Railway dedicated line intelligent scheduling method, device and equipment based on vehicle identity automatic identification, and medium

Abstract

The invention relates to a railway special line intelligent scheduling method and device based on vehicle identity automatic identification, equipment and a medium. The method comprises the following steps: acquiring a multi-modal evidence of a vehicle with a to-be-identified identity, and preprocessing the multi-modal evidence to obtain an evidence set; according to the evidence set, constructing an identity confidence map of the to-be-identified vehicle, and based on the identity confidence map, performing evidence fusion and confidence calculation on the to-be-identified vehicle to obtain a vehicle identity label and confidence thereof; if the confidence exceeds a preset threshold, determining a corresponding digital twinborn model according to the vehicle identity, and binding the real-time position information associated with the UWB positioning evidence into the digital twinborn model to obtain digital twinborn update information; and based on the state constraint and the safety constraint of the current track section, scheduling solution is performed according to the digital twin update information, and a driving scheduling instruction is generated. By adopting the method, the vehicle identity recognition is more reliable through evidence flow driving, and the scheduling decision is more accurate and predictable.

Description

Technical Field

[0001] This invention belongs to the field of railway vehicle dispatching technology, and in particular relates to an intelligent dispatching method, device, equipment and medium for dedicated railway lines based on automatic vehicle identification. Background Technology

[0002] With the development of rail transit and intelligent management technology, vehicle scheduling technology has emerged. By using real-time vehicle location, operating status and historical records, it is possible to achieve automatic identification, real-time tracking and intelligent arrangement of vehicles within a dedicated line.

[0003] In traditional technologies, vehicle identification mainly uses single RFID identification, single visual OCR identification, or a simple combination of both, while scheduling relies on manual planning or independent optimization systems that are disconnected from vehicle identification.

[0004] However, the aforementioned methods suffer from several drawbacks. RFID read rates are affected by metal obstructions, multipath interference, and tag damage, and are easily cloned or forged. Visual OCR is susceptible to contamination, lighting conditions, and motion blur, leading to recognition failures or misidentifications. Existing multi-sensor parallel solutions often rely on simple overlay or voting, lacking a systematic evidence quality assessment and a dynamic fusion mechanism based on evidence sources and historical priors. This makes it difficult to reconstruct credible identities when evidence conflicts arise or some sensors fail. Furthermore, existing scheduling systems typically do not tightly bind real-time confirmed vehicle identities with vehicle files for simulation and rolling optimization. This results in scheduling decisions failing to fully utilize historical operational characteristics and dynamically adjust scheduling schemes based on vehicle identities, thus limiting scheduling efficiency and accuracy. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for intelligent scheduling of dedicated railway lines based on vehicle identity automatic identification, which can accurately identify vehicle identity through evidence fusion and schedule according to vehicle identity.

[0006] Firstly, this application provides an intelligent scheduling method for dedicated railway lines based on automatic vehicle identification, including:

[0007] Multimodal evidence of the vehicle to be identified is obtained, and the multimodal evidence is preprocessed to obtain an evidence set; the evidence set includes each modal evidence and its corresponding evidence quality score; the multimodal evidence includes trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence;

[0008] An identity confidence map of the vehicle to be identified is constructed based on the evidence set, and evidence fusion and confidence calculation are performed on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level.

[0009] If the confidence level exceeds a preset threshold, the corresponding digital twin model is determined based on the vehicle identification, and the real-time location information associated with the UWB positioning evidence is bound to the digital twin model to obtain the digital twin update information; the digital twin model is constructed from the historical vehicle file corresponding to the vehicle identification; the digital twin update information includes the current operating status and expected stay information;

[0010] Based on the current track section status constraints and safety constraints, scheduling solutions are obtained by updating information from the digital twin, and train dispatching instructions are generated. The train dispatching instructions are used to instruct track line equipment to be adjusted according to the target schedule. The track section status is updated by track event evidence.

[0011] In one embodiment, the multimodal evidence is preprocessed to obtain an evidence set, including:

[0012] Multimodal evidence is timestamped using a unified clock and converted into a unified data event format to obtain a standardized raw event stream. The standardized raw event stream includes event identifiers, timestamps, sensor identifiers, and payload fields.

[0013] The standardized raw event stream is imported into the corresponding channel according to the sensor modality type for denoising and feature extraction. Based on the feature extraction results, the evidence quality score of each modality evidence is calculated to obtain the evidence set.

[0014] In one embodiment, the standardized raw event stream is imported into the corresponding channel according to the sensor modality type for denoising and feature extraction. Based on the feature extraction results, the evidence quality score for each modality is calculated to obtain an evidence set, including:

[0015] Active challenge-response verification, antenna multipath read merging, RSSI noise reduction and anti-collision decoding are performed on the trackside RFID evidence to obtain RFID identification results and RFID evidence quality scores.

[0016] The image frames of visual OCR evidence are preprocessed by denoising, motion deblurring and low light enhancement, and character recognition is performed on the license plate area of ​​the preprocessed image frames to obtain OCR recognition results and OCR evidence quality scores.

[0017] UWB positioning evidence was subjected to outlier removal using Kalman filtering, and the UWB positioning data after outlier removal was fused and calculated to obtain position-velocity estimation and UWB positioning evidence quality score.

[0018] After performing a short-time Fourier transform on the waveform of the acoustic-vibration fingerprint evidence, the Mel frequency cepstral coefficients and power spectral density are extracted. Based on the Mel frequency cepstral coefficients and power spectral density, the acoustic-vibration fingerprint is matched with the fingerprint database to obtain the fingerprint matching results and the quality score of the acoustic-vibration fingerprint evidence.

[0019] In one embodiment, an identity confidence map of the vehicle to be identified is constructed based on the evidence set, and evidence fusion and confidence calculation are performed on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level, including:

[0020] The vehicle to be identified is designated as the vehicle entity node, and the trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence are respectively set as evidence nodes.

[0021] Construct directed relation edges from each evidence node to the corresponding vehicle entity node, and determine the evidence quality score and sensor credibility as the edge weights of the directed relation edges to obtain the identity confidence graph;

[0022] Based on the vehicle archive database, at least one vehicle candidate identity identifier and its evidence-identity matching degree are determined according to each evidence node and edge weight in the identity confidence graph.

[0023] Based on Dempste-Shafer evidence fusion, the cumulative identity confidence is calculated when the vehicle entity node is the candidate identity identifier of each vehicle according to the evidence-identity matching degree; when different evidence leads to mutually exclusive identity inference, Dempste-Shafer evidence fusion requantifies conflicting identities based on evidence consistency, historical prior and evidence quality score and triggers candidate retention or manual verification process.

[0024] When the cumulative identity confidence score exceeds the pre-configured threshold, the corresponding vehicle candidate identity identifier will be used as the obtained vehicle identity identifier and its confidence score.

[0025] In one embodiment, a corresponding digital twin model is determined based on the vehicle identification identifier, and real-time location information associated with UWB positioning evidence is bound to the digital twin model to obtain digital twin update information, including:

[0026] The current operating status and route requirements of vehicles are determined based on real-time location information and historical vehicle records; historical vehicle records include vehicle registration information, owner information, and historical operating characteristics.

[0027] The expected dwell time of vehicles is predicted by a pre-trained prediction model based on historical operating characteristics and current operating status, and the expected dwell time information is obtained by combining the route demand.

[0028] The current operating status and expected dwell time information are structured and mapped to obtain digital twin update information.

[0029] In one embodiment, based on the current track segment state constraints and safety constraints, scheduling solutions are performed according to the updated digital twin information to generate train scheduling instructions, including:

[0030] The current operating status of all vehicles and known vehicle scheduling operations are determined from track event evidence, thus obtaining the current track segment state constraints;

[0031] Based on the current track section state constraints and safety constraints, a scheduling scheme that satisfies the constraints is obtained by using mixed-integer linear programming and rolling optimization based on the updated information from the digital twin. The scheduling scheme includes the movement path, timing arrangement, and track equipment resources.

[0032] The scheduling scheme is converted into a structured train dispatching instruction; the train dispatching instruction includes the dispatching start time, the track section number involved, and the route turnout.

[0033] In one embodiment, the method further includes:

[0034] According to the train dispatching instructions, a safety token is requested from the track equipment execution unit; the track equipment includes the turnout controller; before issuing the token, the turnout controller verifies whether the associated track section is currently occupied, and if there is no occupancy, it issues the safety token.

[0035] In response to receiving the safety tokens issued by the execution units of each line equipment, the track equipment begins to execute the train dispatching instructions.

[0036] Secondly, this application also provides an intelligent dispatching device for dedicated railway lines based on automatic vehicle identification, comprising:

[0037] The evidence collection module is used to acquire multimodal evidence of the vehicle to be identified and to preprocess the multimodal evidence to obtain an evidence set. The evidence set includes each modal evidence and its corresponding evidence quality score. The multimodal evidence includes trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence.

[0038] The identity recognition module is used to construct an identity confidence map of the vehicle to be identified based on the evidence set, and to perform evidence fusion and confidence calculation on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level.

[0039] The driving demand module is used to determine the corresponding digital twin model based on the vehicle identification if the confidence level exceeds a preset threshold, and bind the real-time location information associated with UWB positioning evidence to the digital twin model to obtain digital twin update information; the digital twin model is constructed from the historical vehicle file corresponding to the vehicle identification; the digital twin update information includes the current operating status and expected stay information;

[0040] The scheduling module is used to solve scheduling problems based on the current track section status constraints and safety constraints, and generate train dispatching instructions according to the updated information of the digital twin. The train dispatching instructions are used to instruct the track line equipment to adjust according to the target scheduling. The track section status is updated by track event evidence.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described intelligent scheduling methods for dedicated railway lines based on automatic vehicle identification.

[0042] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described intelligent scheduling methods for dedicated railway lines based on automatic vehicle identification.

[0043] The aforementioned intelligent scheduling method, device, equipment, and media for dedicated railway lines based on automatic vehicle identification simultaneously collect multiple types of evidence, including trackside RFID, visual OCR, UWB positioning, acoustic-vibration fingerprints, and track events, avoiding data silos caused by single sensor failures or environmental factors. Evidence quality scores and evidence sets provide quantifiable reliability metrics for the fusion stage, enabling the differentiation between strong and weak evidence and weighted decision-making during fusion, thereby reducing the risk of misjudgment due to low-quality or noisy evidence and improving the overall reliability and robustness of identification. The identity confidence graph enables a time-series, traceable identity inference process, accumulating temporal evidence to improve judgment stability and address the problem of blind decision-making in conflict situations. The digital twin model utilizes historical archives and current dynamic information to more accurately predict dwell time and route demand, providing more binding and predictable input for subsequent scheduling. Using digital twin update information as initial conditions, and combining the current track section status and safety constraints, scheduling solutions are obtained to generate train dispatching instructions and drive track equipment to execute them. This can reflect the confirmed vehicle identity and real-time location information in real time, taking into account safety constraints and the current track status, reducing manual scheduling errors, and improving the timeliness and rationality of scheduling response. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating the intelligent scheduling method for dedicated railway lines based on automatic vehicle identification according to the present invention.

[0046] Figure 2 This is a step-by-step flowchart of step S101;

[0047] Figure 3 This is a flowchart illustrating the steps of step S102.

[0048] Figure 4 This is a structural diagram of the intelligent dispatching device for dedicated railway lines based on automatic vehicle identification according to the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] In one embodiment, such as Figure 1 As shown, a method for intelligent scheduling of dedicated railway lines based on automatic vehicle identification is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S101. Obtain multimodal evidence of the vehicle to be identified, and preprocess the multimodal evidence to obtain an evidence set; the evidence set includes each modal evidence and its corresponding evidence quality score; the multimodal evidence includes trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence.

[0052] As an illustration, active RFID (Radio Frequency Identification) reading gates are deployed at key nodes such as dedicated line entrances and switch areas. Trackside RFID evidence refers to the read-write interaction results between the trackside active RFID reading gates and onboard RFID tags, including tag identification, read / write time, received signal strength (RSSI), and challenge-response verification information, used to quickly obtain tag-level identification information. Cameras are deployed on the dedicated line, with pre-calibrated camera angles to cover the carriage nameplate or car number plate location. Short-focal-length infrared supplementary lighting components are used to cope with low-light scenarios such as nighttime and tunnels, continuously capturing images of the carriage nameplate at a frame rate of no less than 60fps to obtain visual OCR (Optical Character Recognition) evidence, used to supplement identity verification with visual evidence when tags cannot be read or are suspicious. Based on the deployment of UWB (Ultra Wide Band) short-range positioning base stations within the dedicated line plant area, these base stations measure the time of arrival (TOA) or time of flight (TOF) of signals to output the vehicle's three-dimensional coordinates (x, y, z) and position confidence, obtaining UWB positioning evidence. This evidence provides continuous spatiotemporal information for position verification and track occupancy determination. Acoustic-vibration fingerprint evidence is collected by contact accelerometers or non-contact microphone arrays near the track. When a vehicle's wheelset crosses the track, or when the vehicle vibrates or moves, the sensors capture short-time window signals with a sampling rate of 8-48kHz, such as vibration segments of wheels passing through rail gaps or acoustic waveforms of the vehicle moving. Due to the uniqueness of wheelset wear and vehicle structure differences, this data can serve as the vehicle's physical fingerprint, providing physical identity compensation when tags and visual data are missing. Track event evidence includes discrete track-side events such as axle meter pulses and turnout occupancy feedback, primarily used to accurately determine the train's passing time and confirm track resource occupancy status.

[0053] Furthermore, the collected multimodal evidence undergoes preprocessing to ensure data fusionability. For example, preprocessing objectives include time synchronization, formatting, noise reduction, and basic recognition. Specifically, raw data from different sensors are timestamped using a unified clock, enabling all evidence to be correlated within the same time reference frame. Further, various types of raw data are converted into a unified data event format, forming a standardized raw event stream. This event stream includes at least event identifiers, timestamps, sensor identifiers, and payload fields. Optionally, raw data is channelized according to the characteristics of different modalities, such as image enhancement and character detection to output OCR candidates, anti-collision merging of RFID readings while retaining information usable for security verification, smoothing of positioning data to obtain spatiotemporal trajectories, and extracting spectral or cepstral features from acoustic waveforms. A corresponding evidence quality score is calculated for each piece of evidence. The evidence quality score is a quantitative representation of the credibility and usability of a single piece of evidence, comprehensively reflecting factors such as signal-to-noise ratio, recognition confidence, sensor operating status, and security verification results.

[0054] S102. Construct an identity confidence map of the vehicle to be identified based on the evidence set, and perform evidence fusion and confidence calculation on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level.

[0055] This illustration demonstrates how multimodal evidence can be used to construct an abstract structure representing the relationship between evidence and candidate entities: the Identity Confidence Graph (ICG). The purpose of the ICG is to organize originally scattered, heterogeneous, and temporally distributed evidence in a structured and traceable manner, thereby supporting dynamically updated identity inference. Specifically, the ICG is a temporally ordered relational network that connects candidate identifiers of vehicle entities with various evidence nodes through weighted edges. The weights are determined by the evidence quality score and the prior confidence of the sensors. The ICG graph structure allows for the parallel representation of short-term, sudden evidence and long-term historical evidence, facilitating the accumulation of confidence over time. The weighted edges allow for differentiated processing of evidence from different sources, ensuring that the fusion process respects the inherent reliability of the evidence.

[0056] Based on ICG, evidence fusion is not a simple vote, but rather uses formalized fusion rules to cumulatively calculate the support for candidate identities. For example, multi-source evidence linked to the same candidate identity is synthesized and evaluated. During the synthesis process, evidence quality scores are weighted, and the posterior confidence of candidate identities is updated for newly arriving evidence. When mutually exclusive evidence appears, ICG can provide a basis for conflict identification. This is achieved by comparing the quality scores, historical priors, and consistency with existing long-term behavioral patterns of conflicting evidence to determine whether to retain multiple candidates and trigger stronger evidence collection methods, or to elevate a candidate to a more subjectively reliable identity candidate. Through a structured graph-based and weighted fusion mechanism, ICG can output the candidate vehicle identity identifier corresponding to each vehicle entity node and its real-time updated identity confidence. This yields the vehicle identity identifier, which uniquely represents the vehicle, and the identity confidence, a quantitative measure of the identifier's consistency with the current observation set.

[0057] S103. If the confidence level exceeds the preset threshold, the corresponding digital twin model is determined based on the vehicle identification, and the real-time location information associated with the UWB positioning evidence is bound to the digital twin model to obtain the digital twin update information. The digital twin model is constructed from the historical vehicle file corresponding to the vehicle identification. The digital twin update information includes the current operating status and expected stay information.

[0058] When the confidence level of a candidate vehicle identity output by the ICG reaches a preset confirmation threshold, it is determined that the candidate identity is supported by sufficient evidence, thus entering the binding stage with the digital twin model. Illustratively, a digital twin model refers to a structured entity built based on historical vehicle archives, used to represent the mapping between a single vehicle in the physical world and a virtual model. This model includes at least static attributes such as vehicle registration information, owner information, historical operating characteristics, and known dimensions / wheelbase, as well as dynamic fields for real-time scheduling, such as current location, track segment, current status, and resource occupancy. Specifically, by indexing and matching the identity identifier output by the ICG with digital twin model entries in the historical vehicle archive, associated static information is obtained. Subsequently, the real-time location information from UWB positioning evidence is associated with the current location segment field of the digital twin model to reflect the vehicle's current location and serve as the initial condition for scheduling. The digital twin model supports inferences about expected dwell time, route requirements, and formation constraints.

[0059] S104. Based on the current track section status constraints and safety constraints, the scheduling solution is performed according to the digital twin update information to generate train dispatching instructions; the train dispatching instructions are used to instruct the track line equipment to be adjusted according to the target scheduling; the track section status is updated by track event evidence.

[0060] After the digital twin model reflects the current operating status, real-time location information, and expected dwell time of the confirmed vehicles, real-time constraint scheduling is solved by combining the track section state constraints and safety constraints on site. The track section state constraints include the current occupancy status of each track section, the availability of switches, and known vehicle scheduling operation information. Safety constraints cover the spatiotemporal constraints necessary to ensure operational safety, such as train separation time, switch switching time window, maximum permissible speed, and locomotive resource limitations. Illustratively, the objectives and expected dwell time of each vehicle in the digital twin model are converted into scheduling tasks, and the track section state and safety constraints are encoded into a set of solvable constraints. Within a short-time rolling window, heuristic or mixed-integer linear programming methods are used to solve the scheduling problem. The optimization objective can be to minimize the overall operation completion time, reduce energy consumption, or reduce delay penalties, while ensuring that the constraints are satisfied. During the solution process, the location information, identity confidence level, and historical behavior patterns provided by the digital twin model are used to predict execution risks and time window stability, thereby influencing the setting of constraint weights or priorities. After the solution is completed, a scheduling scheme that satisfies the constraints is obtained, including the movement path, time arrangement, and required track equipment resources for each vehicle, including the vehicle to be identified. Furthermore, the scheduling scheme is converted into structured train dispatching instructions and issued to the execution layer. The instructions include, but are not limited to, the scheduling start time, the track section number involved, and route turnout information, which are used to instruct the track equipment and execution units to implement adjustments according to the target schedule.

[0061] The aforementioned intelligent scheduling method for dedicated railway lines based on automatic vehicle identification effectively suppresses errors caused by misidentification or omission of single modalities by integrating multi-source evidence from RFID, vision, positioning, acoustic-vibration, and track events, thereby significantly improving the overall accuracy and stability of vehicle identification. Quality scores are used to quantify the credibility of each piece of evidence, and weighted accumulation and conflict resolution are performed in the identity confidence graph. When faced with scenarios such as RFID occlusion / cloning, visual occlusion, or positioning errors, the system can reduce the risk of erroneous actions due to single-point failures by retaining candidate evidence, waiting for subsequent evidence, or prioritizing high-quality evidence. The identity confidence graph structures the relationship between evidence and candidate entities and records the source and weight of the evidence. This allows the system to re-quantify conflicts based on quality scores and historical priors when mutually exclusive evidence appears, deciding whether to retain evidence or trigger manual verification. By binding verified identities and real-time location information to a digital twin model, and maintaining the current operating status and expected dwell time information within the digital twin model, the scheduling module can perform rolling optimization solutions based on a more complete and structured set of vehicle information. This results in more timely scheduling schemes that conform to real-world constraints within a short time window, which helps shorten response time and reduce operational conflicts. Because the scheduling solution considers the expected dwell time and route requirements provided by the digital twin model, the generated scheduling scheme, while meeting safety constraints, facilitates the rational allocation of resources such as locomotives, track sections, and switches, reducing blind occupation and empty runs. This improves the utilization rate of on-site resources and overall operational efficiency, accurately avoids dangerous scenarios such as track conflicts, switch misalignment, or overtime operations, and thus enhances the compliance and operational safety of scheduling execution.

[0062] In one embodiment, the multimodal evidence is preprocessed to obtain an evidence set, including:

[0063] S11. Based on a unified clock, timestamp-align the multimodal evidence and convert it into a unified data event format to obtain a standardized raw event stream. The standardized raw event stream includes event identifier, timestamp, sensor identifier, and payload field.

[0064] As an illustration, a unified clock is constructed using PTP (Precision Time Protocol) / NTP (Network Time Protocol) protocols combined with GPS time. This timestamps all raw data collected by sensors with the same time reference, avoiding time discrepancies caused by sensor response delays. Simultaneously, various types of evidence are converted into a unified JSON data event format, including an event identifier (event_id), timestamp, sensor identifier (source_id), sequence number (seq), and payload fields, such as unique identifiers from RFID (tag_id) and image frames from visual recognition, ultimately forming a standardized raw event stream. Optionally, this standardized raw event stream is transmitted via a Kafka (distributed stream processing platform) / MQTT (Message Queuing Telemetry Transport) message bus to ensure data consistency.

[0065] S12. The standardized raw event stream is imported into the corresponding channel according to the sensor modality type for denoising and feature extraction. The evidence quality score of each modality evidence is calculated based on the feature extraction results to obtain the evidence set.

[0066] Standardized raw event streams are imported into dedicated processing channels based on sensor modality type, such as RFID and vision channels. Adaptive algorithms are used to remove environmental interference, such as RSSI moving average denoising for RFID and frame denoising and motion deblurring for vision. Core features for identification are extracted, such as valid RFID tag information, vehicle license plate character regions for vision, and Mel-frequency cepstral coefficients for acoustic-vibrational data. Furthermore, an evidence quality score is calculated for each modality, based on factors including signal-to-noise ratio (SNR), feature sharpness, and data integrity. For example, if RFID evidence passes the challenge-response check and has a high RSSI signal strength (e.g., -55dBm, above the -70dBm threshold), the quality score can be set to 0.9, while visual evidence affected by lighting interference may only have a quality score of 0.6.

[0067] In one embodiment, such as Figure 2 As shown, the standardized raw event stream is imported into the corresponding channel according to the sensor modality type for denoising and feature extraction. Based on the feature extraction results, the evidence quality score for each modality is calculated, resulting in an evidence set, including:

[0068] S201. Active challenge-response verification, antenna multipath read merging, RSSI noise reduction and anti-collision decoding are performed on the trackside RFID evidence to obtain the RFID identification result and RFID evidence quality score.

[0069] As an illustration, the active RFID tag supports a challenge-response mechanism. During the deployment phase of the RFID reader gate, a unique key is preset for each vehicle's tag. When a vehicle passes through the reader gate, the reader sends a random challenge code (nonce) to the tag. The tag encrypts the challenge code based on the preset key, generates a challenge response (nonce_response), and sends it back to the reader gate. Furthermore, after receiving the nonce_response, the reader gate reverses the calculation using the same key and encryption algorithm to verify whether the calculation result is consistent with the local result. If they are consistent, it means that the tag is a genuine and valid tag and has not been cloned, and the RFID data is retained. If they are inconsistent, it is determined to be a malicious tag or a cloned tag, the data is marked as invalid, and a score of 0 is directly assigned to it in the subsequent quality score calculation to avoid interference from false identity information.

[0070] Furthermore, the trackside RFID reading gate is equipped with 3-4 antennas. When the same vehicle passes by, multiple antennas may read the information of the same tag one after another, forming redundant data. In order to avoid redundant calculations caused by data redundancy, the reading data of multiple antennas is merged. For example, a fixed time window is set. The window duration can be determined according to the time when the vehicle passes through the reading gate to ensure that the reading cycle of all antennas is covered. All the same tag_id data read within the window is integrated into a single arrival event. At the same time, the maximum RSSI value of each antenna is retained to reflect the state when the tag signal is strongest, the earliest reading timestamp, and duplicate tag_id and other fields are removed.

[0071] RSSI is a key indicator reflecting the distance between RFID tags and readers and signal stability. However, it is easily affected by metal carriage obstruction and electromagnetic interference from surrounding electrical equipment, resulting in instantaneous fluctuations. For example, a normal RSSI of -58dBm may jump to -75dBm when interfered with. For instance, a moving average filtering algorithm is used to denoise the RSSI. The RSSI data of 5 consecutive reading cycles are selected as a sliding window, and the average value within the window is calculated as the valid RSSI at the current moment. If the deviation between the original RSSI at a certain moment and the window average value exceeds a preset threshold of 10dBm, it is determined to be an outlier based on the electromagnetic environment calibration of the dedicated line and replaced with the window average value. Finally, a smooth RSSI curve is output to ensure the stability of the signal strength data.

[0072] Optionally, when multiple vehicles pass through the reader gate consecutively, the RFID tag signals from different vehicles may overlap at the reader's receiver, causing tag collisions and data decoding errors, such as missing tag_id fields or CRC check failures. For example, a dynamic frame slotted ALOHA algorithm is used. The reader divides the reading period into multiple slots, and each tag randomly selects a slot to send data. The reader monitors the signal status of each slot in real time. If a slot receives only a single tag signal, decoding proceeds normally; if signal overlap is detected, the slot is marked as invalid, and the tag is notified to reselect a slot to send data, until all tag data is successfully decoded. Data that fails to decode is directly marked as invalid and not included in subsequent feature extraction.

[0073] Furthermore, core identification features, including the vehicle's unique tag ID (tag_id), reader ID (reader_id), and the denoised RSSI value (valid_rssi), are extracted from the valid RFID data to form the RFID identification result. The RFID evidence quality score is calculated based on three dimensions: authenticity verification result, signal stability, and data integrity. That is, the quality score is obtained by weighting the verification result, the normalized RSSI value, and the data integrity according to a set weight.

[0074] S202. Perform noise reduction, motion deblurring, and low-light enhancement preprocessing on the image frames of the visual OCR evidence, and perform character recognition on the license plate area of ​​the preprocessed image frames to obtain the OCR recognition result and the OCR evidence quality score.

[0075] To illustrate, frame denoising employs a Gaussian filtering algorithm to remove random noise from the image, while motion deblurring utilizes the deep learning-based DeblurGAN network. This network, trained on numerous blurred-to-sharp image pairs, learns vehicle motion trajectory features and reverses this process to restore sharp images of blurred areas. Furthermore, for low-light scenes such as nighttime and tunnels, the Retinex algorithm enhances image brightness and contrast. By decomposing the image into illuminance and reflectance components, adjusting the illuminance component improves overall brightness while simultaneously enhancing the contrast of the reflectance component. This allows the vehicle's nameplate to stand out from the dark background in low-light environments, ensuring clear and legible character details and avoiding the local overexposure problem caused by traditional histogram equalization. Furthermore, after preprocessing, the license plate area needs to be located first, and then YOLOv5 or SSD object detection models are used to extract character features for recognition. A combination of CRNN (Convolutional Recurrent Neural Network) and CTC (Connectionist Temporal Classification) algorithms is used to adapt to the sequence features of the license plate characters. Specifically, CRNN first extracts spatial features such as strokes and contours of the characters in the license plate ROI (Region of Interest) through convolutional layers, then learns the temporal dependencies of the characters through a long short-term memory network recurrent layer, and finally decodes the temporal features through a CTC layer, outputting the character sequence and the probability value of each character. For example, a recognition window is set to ensure that 10-20 frames of valid license plate ROIs are obtained, and the majority consistent sequence of all recognition results within the window is used as the final result. The OCR evidence quality score is obtained by weighting recognition consistency, character probability, and image quality.

[0076] S203. Outlier removal is performed on the UWB positioning evidence using Kalman filtering, and the UWB positioning data after outlier removal is fused and calculated to obtain the position-velocity estimate and the UWB positioning evidence quality score.

[0077] To illustrate, the UWB positioning data is updated at a frequency of 10-50Hz. Each frame contains the vehicle's x, y, and z coordinates and the signal strength (TOA / TOF_rssi) based on the time of arrival (TOA) / time of flight. The Kalman filter predicts the theoretical position at the current moment based on the vehicle's position and velocity from the previous moment, combined with the vehicle's motion model. The raw UWB positioning data at the current moment is compared with the predicted position to calculate the residual. If the residual exceeds a preset value, it is considered an outlier, and the raw data is not used; instead, the predicted position is taken as the current valid position. If the residual is less than or equal to a threshold, the raw data and the predicted position are weighted and fused. The weights of the raw data are adjusted according to TOF_rssi, with stronger signals receiving higher weights. The resulting smoothed position at the current moment is then output.

[0078] Optionally, more than three UWB base stations are typically deployed within the dedicated line factory area. Time-of-Flight (TOF) data from a single base station is susceptible to signal delay, necessitating the fusion of data from multiple base stations to improve positioning accuracy. Extended Kalman Filtering (EKF) is used for fusion. Specifically, EKF uses multiple sets of distance data as observations, combined with the known coordinates of the base stations, to solve for the vehicle's optimal position (x, y) using a nonlinear state equation. During the fusion process, weights are assigned based on the TOF_rssi (signal strength) of each base station. Base station data with high TOF_rssi (e.g., -50dBm) have higher weights, while data from base stations with low TOF_rssi have lower weights, ensuring that high-quality data dominates the positioning results and further reducing the impact of single-base station errors. Based on the smoothed position after fusion, the vehicle's real-time speed is calculated using position differences between adjacent frames. The final output UWB positioning result includes the vehicle's three-dimensional coordinates, real-time speed v, positioning time (timestamp), and the base station number used for positioning. The UWB positioning evidence quality score (pos_confidence) is calculated based on filter residuals, multi-base station consistency, and signal strength.

[0079] S204. After performing a short-time Fourier transform on the waveform of the acoustic-vibration fingerprint evidence, extract the Mel frequency cepstral coefficients and power spectral density. Based on the Mel frequency cepstral coefficients and power spectral density, perform similarity matching between the acoustic-vibration fingerprint and the fingerprint database to obtain the fingerprint matching result and the acoustic-vibration fingerprint evidence quality score.

[0080] In illustrative terms, the raw signals acquired by acoustic-vibration sensors are continuous time-domain waveforms. Direct analysis makes it difficult to distinguish vehicle characteristics from environmental noise. Therefore, a Short-Time Fourier Transform (STFT) is needed to convert the time-domain signal into a two-dimensional time-frequency spectrum, enabling feature visualization and noise separation. For example, a sliding time window of 20-50ms with a step size of 10ms is set, and Fourier transforms are performed on the raw waveform segment by segment to obtain the frequency spectrum corresponding to each time window. Two types of features are extracted from the STFT spectrum: Mel Frequency Cepstral Coefficients (MFCC) and Power Spectral Density (PSD), to best reflect the physical uniqueness of the vehicle. MFCC simulates the sensitivity of the human ear to different frequencies. The linear frequency axis of the STFT is converted to the Mel frequency axis, and the Mel spectrum is extracted using a triangular filter bank. Then, a Discrete Cosine Transform (DCT) is performed on the Mel spectrum, and the first 13 DCT coefficients are taken as the MFCC features. Wheelset abrasion causes a unique peak in the vibration signal in a specific Mel frequency band, and its MFCC coefficients differ significantly from those of other vehicles. DCT reflects the power distribution of a signal at different frequencies, calculated by averaging the squares of the STFT spectrum. PSD highlights the frequency at which vehicle vibration energy is concentrated. For example, an 8-axle coal truck has more wheelsets, and its vibration signal power in the 2kHz band is significantly higher than that of a 6-axle container truck. Furthermore, the peak position and amplitude of the PSD curve are unique, serving as an important basis for distinguishing different vehicles. Further, the extracted MFCC and PSD features need to be normalized to eliminate differences in feature scale caused by different sensor sensitivities and acquisition distances, ensuring fairness in subsequent matching.

[0081] The backend database pre-stores the baseline acoustic-vibration fingerprints of all vehicles, forming a fingerprint database. For example, a combination algorithm of Dynamic Time Warping (DTW) and cosine similarity is used to calculate fingerprint similarity. Specifically, for MFCC features, the DTW algorithm stretches or compresses the time axes of the real-time MFCC sequence and the baseline MFCC sequence to find the optimal matching path, calculating the sum of the distances along the path; the smaller the distance, the higher the similarity. For PSD features, the cosine similarity algorithm treats the real-time PSD and the baseline PSD as vectors, calculating the cosine of the angle between the two vectors; the closer the cosine value is to 1, the higher the similarity. The similarities of the two features are weighted and fused into the final similarity. If the similarity is ≥ a preset threshold, a successful match is determined, and the acoustic-vibration fingerprint recognition result is output based on the similarity matching result. The acoustic-vibration fingerprint evidence quality score (acoustic_score) is based on similarity, signal-to-noise ratio (SNR), and feature integrity. The signal SNR is obtained by calculating the ratio of the power of the vehicle feature frequency band to the power of the environmental noise frequency band in the acquired signal.

[0082] In one embodiment, such as Figure 3 As shown, an identity confidence map of the vehicle to be identified is constructed based on the evidence set. Evidence fusion and confidence calculation are then performed on the vehicle based on the identity confidence map to obtain the vehicle's identity identifier and its confidence level, including:

[0083] S301. Identify the vehicle to be identified as the vehicle entity node, and set the trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence as evidence nodes respectively.

[0084] As an illustration, the vehicle entity node corresponds to the vehicle to be identified. In the initial stage, a temporary internal identifier can be assigned, which will be bound to the formal vehicle identity identifier after the identity is confirmed. The evidence node corresponds to each type of modal evidence, and each evidence node stores the features of that modality.

[0085] S302. Construct directed relation edges from each evidence node to the corresponding vehicle entity node, and determine the evidence quality score and sensor credibility as the edge weights of the directed relation edges to obtain the identity confidence graph.

[0086] For example, the edge weights are calculated by weighted fusion of evidence quality score and sensor credibility, with sensor credibility being a preset parameter. Optionally, the identity confidence graph needs to be stored in a graph database such as Neo4j, retaining node and edge data from the past N days to form a vehicle behavior signature. This historical association can serve as an auxiliary basis for subsequent identity inference, further improving recognition stability.

[0087] S303. Based on the vehicle archive database, determine at least one vehicle candidate identity identifier and its evidence-identity matching degree according to each evidence node and edge weight in the identity confidence graph.

[0088] As an illustration, possible vehicle identities, i.e., candidate identities, are matched in the backend vehicle archive. For example, the RFID TAG-K1001 corresponds to K1001 (8-axle coal truck) in the archive. The visual OCR recognition of K1001 further verifies the candidate. If there are fuzzy features, such as acoustic-vibration fingerprints that match both K1001 and K1002, multiple candidate identities K1001 and K1002 are generated, and the evidence-identity matching degree of each candidate identity with each evidence node is calculated. For example, the matching degree of K1001 with R1 is 1.0, and the matching degree with V1 is 0.88.

[0089] S304. Based on Dempste-Shafer evidence fusion, calculate the cumulative identity confidence when the vehicle entity node is the candidate identity identifier of each vehicle according to the evidence-identity matching degree; when Dempste-Shafer evidence fusion is used to deal with the mutual exclusion of identity inference caused by different evidence, the conflicting identities are requantified based on the consistency between evidence, historical prior and evidence quality score and the candidate retention or manual verification process is triggered.

[0090] Furthermore, the confidence score is calculated using a Bayesian update and the Dempster-Shafer (DS) evidence fusion algorithm. The Bayesian update is used to integrate independent supporting evidence, as shown by the formula... The calculation is as follows: P(evidence|entity) is the likelihood of the evidence to the candidate identity, and P_prior(entity) is the historical prior probability of the candidate identity. For example, if K1001 has appeared 20 times on this dedicated line in the past 30 days, the prior probability is 0.8. When there is evidence conflict, such as one piece of evidence pointing to K1001 and another piece of evidence pointing to K1002, the DS evidence fusion framework is activated. The basic probability allocation (BPA) is used to quantify the degree of support of each piece of evidence for the candidate identity. Combined with the consistency between evidence, such as the majority of evidence supporting K1001, and the historical prior such as K1002 not being dispatched to this area recently, the evidence quality score is also used to requantify the conflict. For example, if the quality score of the evidence supporting K1002 is only low quality (0.4), its weight is reduced in the fusion to avoid misleading the results.

[0091] S305. When the cumulative identity confidence score exceeds the pre-configured threshold, the corresponding vehicle candidate identity identifier is used as the obtained vehicle identity identifier and its confidence score.

[0092] Finally, the cumulative identity confidence score of each candidate identity corresponding to the vehicle entity node is calculated. The candidate identity with the highest confidence score and exceeding the preset threshold is the final vehicle identity identifier. Optionally, if the confidence scores of all candidate identities are in the low confidence interval, the candidate identities are retained and await subsequent supplementary evidence, such as new evidence of the vehicle passing through the next RFID gate. If the confidence score is lower than 0.4, a manual verification process is triggered to ensure the rigor of identity inference.

[0093] In one embodiment, a corresponding digital twin model is determined based on the vehicle identification identifier, and real-time location information associated with UWB positioning evidence is bound to the digital twin model to obtain digital twin update information, including:

[0094] S21. Determine the current operating status and route requirements of the vehicle based on real-time location information and historical vehicle records; historical vehicle records include vehicle registration information, owner information, and historical operating characteristics.

[0095] The construction of a digital twin model requires a foundation of historical vehicle records, stored in a backend database. These records encompass both static and dynamic information throughout the vehicle's entire lifecycle. Static information includes vehicle registration details such as the vehicle number (wagon_no), unique tag (tag_id), owner information, vehicle technical parameters, vehicle type (e.g., coal car / container car), number of axles, maximum load capacity, and car dimensions. Dynamic historical information includes historical operational characteristics such as average dwell time over the past three months, frequently stopped track sections, and typical driving speed. It also includes maintenance records, such as the date of the most recent maintenance, fault history, and load history, i.e., the type of goods frequently transported (e.g., coal / ore) and their weight.

[0096] Furthermore, the vehicle's three-dimensional coordinates (x, y, z) provided by UWB positioning evidence need to be mapped to the electronic track map of the dedicated line to determine the current track segment where the vehicle is located, such as track 3 at x=120.4m. The real-time position (x, y) and track segment are then written into the current_position and track_segment fields of the digital twin model. At the same time, combined with track event evidence, such as the vehicle's movement status recorded by the axle counter and the switch occupancy status, the current operating status of the vehicle is determined, i.e., moving, stopped, shunting, or blocked.

[0097] S22. Using a pre-trained prediction model, predict the expected dwell time of vehicles based on historical operating characteristics and current operating status, and combine this with route requirements to obtain expected dwell information.

[0098] The expected dwell time is the predicted dwell time of a vehicle in the current track section. This prediction is based on historical operating characteristics and the current operating status, and is achieved through a pre-trained XGBoost lightweight machine learning model. The predicted expected dwell time is obtained by inputting features such as vehicle load type, current time, and track section operation efficiency.

[0099] S23. The current running status and expected dwell time information are structured and mapped to obtain the digital twin update information.

[0100] In one embodiment, based on the current track segment state constraints and safety constraints, scheduling solutions are performed according to the updated digital twin information to generate train scheduling instructions, including:

[0101] S31. Determine the current operating status of all vehicles and known vehicle scheduling operations based on track event evidence to obtain the current track section state constraints.

[0102] Indicatively, the current track section status constraints can include track occupancy constraints, i.e., only one vehicle is allowed to stop on each track section. If a track is occupied by a vehicle, other vehicles are prohibited from entering that track during its dwell time. Equipment status constraints, i.e., the loading and unloading equipment and turnout equipment on the dedicated line have fixed operating capabilities. For example, one unloader can only serve one vehicle at a time, and turnout switching takes 2 seconds. The constraints need to clearly define the current busy / idle status and operating time of the equipment.

[0103] S32. Based on the current track segment state constraints and safety constraints, a scheduling scheme that satisfies the constraints is obtained by using mixed integer linear programming and rolling optimization based on the digital twin update information; the scheduling scheme includes the movement path, timing arrangement and track equipment resources.

[0104] Safety constraints include conflict prevention constraints, which prohibit two cars from occupying the same track section or crossing routes at the same time; turnout time window constraints, which prevent vehicles from being introduced into occupied routes during emergencies, thus causing safety risks; locomotive resource constraints, which require the scheduling plan to match the number of locomotives currently available; and emergency priority constraints, which require emergency vehicles to have the highest scheduling priority and require the suspension of non-emergency operations of ordinary vehicles to make way for emergency vehicles. To illustrate, the algorithm collects updated digital twin information of all confirmed vehicles, current track segment status constraints, safety constraint rules, and manual priority rules to form a complete scheduling data matrix, ensuring that the algorithm can obtain all decision information. Furthermore, it employs a combination of Mixed Integer Linear Programming (MILP) and heuristic algorithms for rolling optimization. Addressing the real-time requirements of dedicated railway line scheduling, the solution window is divided into short-term and long-term windows. Within the short-term window, if the number of vehicles to be scheduled is small, the MILP algorithm is directly used, with the objective functions of minimizing total operation time, minimizing energy consumption, and minimizing train delay penalties, to find the optimal scheduling scheme that satisfies all constraints. If the number of vehicles to be scheduled is large, a hierarchical optimization strategy is adopted. First, track allocation is completed using MILP, and then locomotive allocation and operation sequence optimization are completed using a greedy algorithm or local search algorithm to determine which locomotive to pull and which vehicle to perform operations first, balancing solution efficiency and optimality.

[0105] S33. Convert the scheduling scheme into a structured train dispatching instruction; the train dispatching instruction includes the scheduling start time, the track section number involved, and the route turnout.

[0106] Furthermore, the optimized scheduling scheme is converted into structured train dispatching instructions. The scheduling scheme includes the vehicle's movement path, timing arrangement, and track equipment resource requirements. This information needs to be converted into an instruction format that the equipment can recognize. The instructions must include the dispatching instruction ID, generation timestamp, involved vehicle ID, involved track section number, start time, required locomotive, turnout adjustment requirements, and safety token identifier.

[0107] In one embodiment, the method further includes:

[0108] S41. Apply for a safety token from the track equipment execution unit according to the train dispatching instructions; the track equipment includes the turnout controller; before issuing the token, the turnout controller verifies whether the associated track section is currently occupied, and if there is no occupancy, it issues the safety token.

[0109] As an illustration, the track section occupancy verification process by the turnout controller requires verification based on track event evidence. It checks whether the track section involved in the scheduling instruction, particularly the target track connected to the turnout, is idle. For example, if an application is made to switch turnout 1 to position and connect track 5, it verifies whether track 5 is occupied by other vehicles. If track 5 is currently occupied, but the digital twin model shows its expected dwell time is earlier than the scheduling start time, it is determined to be available in the future, meeting the application requirements. If track 5 is occupied and there is no expected idle time, an occupancy conflict is determined. Furthermore, if all verifications pass, the turnout controller generates a unique security token and sends it to the scheduler via an encrypted channel. If any verification fails, the token is rejected, and the reason for failure is returned. Upon receiving the rejection information, the scheduler triggers a rescheduling of the scheduling plan.

[0110] S42. In response to receiving the safety tokens issued by the execution units of each line equipment, the track equipment begins to execute the train dispatching instructions.

[0111] As an illustration, an action execution command is issued to the turnout controller. This command must carry a corresponding safety token. The turnout controller only accepts action commands with valid tokens to prevent unauthorized operations without a token. A travel execution command is also issued to the locomotive and rolling stock's onboard terminals, including departure time, target track, and speed limits, ensuring synchronization between the locomotive and turnout actions. After receiving the command, each execution unit coordinates its operations according to a preset sequence, ensuring that track equipment switches and locks before the vehicle enters, thus guaranteeing safety.

[0112] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0113] Based on the same inventive concept, this application also provides an intelligent scheduling device for dedicated railway lines based on automatic vehicle identification, used to implement the aforementioned intelligent scheduling method for dedicated railway lines based on automatic vehicle identification. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent scheduling device for dedicated railway lines based on automatic vehicle identification provided below can be found in the limitations of the intelligent scheduling method for dedicated railway lines based on automatic vehicle identification described above, and will not be repeated here.

[0114] In one exemplary embodiment, such as Figure 4 As shown, a smart dispatching device for dedicated railway lines based on automatic vehicle identification is provided, comprising:

[0115] The evidence collection module 401 is used to acquire multimodal evidence of the vehicle to be identified and to preprocess the multimodal evidence to obtain an evidence set. The evidence set includes each modal evidence and its corresponding evidence quality score. The multimodal evidence includes trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence.

[0116] The identity recognition module 402 is used to construct an identity confidence map of the vehicle to be identified based on the evidence set, and to perform evidence fusion and confidence calculation on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level.

[0117] The driving demand module 403 is used to determine the corresponding digital twin model based on the vehicle identification if the confidence level exceeds a preset threshold, and bind the real-time location information associated with UWB positioning evidence to the digital twin model to obtain digital twin update information; the digital twin model is constructed from the historical vehicle file corresponding to the vehicle identification; the digital twin update information includes the current operating status and expected stay information;

[0118] The scheduling module 404 is used to solve the scheduling problem based on the current track section status constraints and safety constraints, and according to the updated information of the digital twin, and generate train dispatching instructions. The train dispatching instructions are used to instruct the track line equipment to be adjusted according to the target scheduling. The track section status is updated by track event evidence.

[0119] In one embodiment, a data preprocessing module is also included, for:

[0120] Multimodal evidence is timestamped using a unified clock and converted into a unified data event format to obtain a standardized raw event stream. The standardized raw event stream includes event identifiers, timestamps, sensor identifiers, and payload fields.

[0121] The standardized raw event stream is imported into the corresponding channel according to the sensor modality type for denoising and feature extraction. Based on the feature extraction results, the evidence quality score of each modality evidence is calculated to obtain the evidence set.

[0122] In one embodiment, a feature extraction module is also included, for:

[0123] Active challenge-response verification, antenna multipath read merging, RSSI noise reduction and anti-collision decoding are performed on the trackside RFID evidence to obtain RFID identification results and RFID evidence quality scores.

[0124] The image frames of visual OCR evidence are preprocessed by denoising, motion deblurring and low light enhancement, and character recognition is performed on the license plate area of ​​the preprocessed image frames to obtain OCR recognition results and OCR evidence quality scores.

[0125] UWB positioning evidence was subjected to outlier removal using Kalman filtering, and the UWB positioning data after outlier removal was fused and calculated to obtain position-velocity estimation and UWB positioning evidence quality score.

[0126] After performing a short-time Fourier transform on the waveform of the acoustic-vibration fingerprint evidence, the Mel frequency cepstral coefficients and power spectral density are extracted. Based on the Mel frequency cepstral coefficients and power spectral density, the acoustic-vibration fingerprint is matched with the fingerprint database to obtain the fingerprint matching results and the quality score of the acoustic-vibration fingerprint evidence.

[0127] In one embodiment, the identity recognition module 402 is further configured to:

[0128] The vehicle to be identified is designated as the vehicle entity node, and the trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence are respectively set as evidence nodes.

[0129] Construct directed relation edges from each evidence node to the corresponding vehicle entity node, and determine the evidence quality score and sensor credibility as the edge weights of the directed relation edges to obtain the identity confidence graph;

[0130] Based on the vehicle archive database, at least one vehicle candidate identity identifier and its evidence-identity matching degree are determined according to each evidence node and edge weight in the identity confidence graph.

[0131] Based on Dempste-Shafer evidence fusion, the cumulative identity confidence is calculated when the vehicle entity node is the candidate identity identifier of each vehicle according to the evidence-identity matching degree; when different evidence leads to mutually exclusive identity inference, Dempste-Shafer evidence fusion requantifies conflicting identities based on evidence consistency, historical prior and evidence quality score and triggers candidate retention or manual verification process.

[0132] When the cumulative identity confidence score exceeds the pre-configured threshold, the corresponding vehicle candidate identity identifier will be used as the obtained vehicle identity identifier and its confidence score.

[0133] In one embodiment, the driving demand module 403 is further configured to:

[0134] The current operating status and route requirements of vehicles are determined based on real-time location information and historical vehicle records; historical vehicle records include vehicle registration information, owner information, and historical operating characteristics.

[0135] The expected dwell time of vehicles is predicted by a pre-trained prediction model based on historical operating characteristics and current operating status, and the expected dwell time information is obtained by combining the route demand.

[0136] The current operating status and expected dwell time information are structured and mapped to obtain digital twin update information.

[0137] In one embodiment, the scheduling module 404 is further configured to:

[0138] The current operating status of all vehicles and known vehicle scheduling operations are determined from track event evidence, thus obtaining the current track segment state constraints;

[0139] Based on the current track section state constraints and safety constraints, a scheduling scheme that satisfies the constraints is obtained by using mixed-integer linear programming and rolling optimization based on the updated information from the digital twin. The scheduling scheme includes the movement path, timing arrangement, and track equipment resources.

[0140] The scheduling scheme is converted into a structured train dispatching instruction; the train dispatching instruction includes the dispatching start time, the track section number involved, and the route turnout.

[0141] In one embodiment, a security protection module is also included, for:

[0142] According to the train dispatching instructions, a safety token is requested from the track equipment execution unit; the track equipment includes the turnout controller; before issuing the token, the turnout controller verifies whether the associated track section is currently occupied, and if there is no occupancy, it issues the safety token.

[0143] In response to receiving the safety tokens issued by the execution units of each line equipment, the track equipment begins to execute the train dispatching instructions.

[0144] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0145] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0146] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0147] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent scheduling of dedicated railway lines based on automatic vehicle identification, characterized in that, The method includes: Multimodal evidence of the vehicle to be identified is acquired, and the multimodal evidence is preprocessed to obtain an evidence set; the evidence set includes each modal evidence and its corresponding evidence quality score; the multimodal evidence includes trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence; Based on the evidence set, an identity confidence map of the vehicle to be identified is constructed, and based on the identity confidence map, evidence fusion and confidence calculation are performed on the vehicle to be identified to obtain the vehicle identity identifier and its confidence level. If the confidence level exceeds a preset threshold, a corresponding digital twin model is determined based on the vehicle identification, and the real-time location information associated with the UWB positioning evidence is bound to the digital twin model to obtain digital twin update information; the digital twin model is constructed from the historical vehicle file corresponding to the vehicle identification; the digital twin update information includes the current operating status and expected stay information; Based on the current track section status constraints and safety constraints, scheduling solutions are performed according to the updated digital twin information to generate train scheduling instructions; the train scheduling instructions are used to instruct track line equipment to adjust according to the target scheduling; the track section status is updated by the track event evidence.

2. The method according to claim 1, characterized in that, The preprocessing of the multimodal evidence to obtain the evidence set includes: The multimodal evidence is timestamped based on a unified clock and converted into a unified data event format to obtain a standardized raw event stream; the standardized raw event stream includes event identifier, timestamp, sensor identifier and payload field; The standardized raw event stream is imported into the corresponding channel according to the sensor modality type for denoising and feature extraction. Based on the feature extraction results, the evidence quality score of each modality evidence is calculated to obtain the evidence set.

3. The method according to claim 2, characterized in that, The process involves importing the standardized raw event stream into the corresponding channel according to the sensor modality type for denoising and feature extraction, and calculating the evidence quality score for each modality based on the feature extraction results to obtain the evidence set, which includes: The trackside RFID evidence is subjected to active challenge-response verification, antenna multipath read merging, RSSI noise reduction and anti-collision decoding to obtain RFID identification results and RFID evidence quality scores. The image frames of the visual OCR evidence are preprocessed with denoising, motion deblurring and low-light enhancement, and character recognition is performed on the license plate area of ​​the preprocessed image frames to obtain the OCR recognition result and the OCR evidence quality score. UWB positioning evidence was subjected to outlier removal using Kalman filtering, and the UWB positioning data after outlier removal was fused and calculated to obtain position-velocity estimation and UWB positioning evidence quality score. After performing a short-time Fourier transform on the waveform of the acoustic-vibration fingerprint evidence, the Mel frequency cepstral coefficients and power spectral density are extracted. Based on the Mel frequency cepstral coefficients and power spectral density, the acoustic-vibration fingerprint is matched with the fingerprint database to obtain the fingerprint matching result and the acoustic-vibration fingerprint evidence quality score.

4. The method according to claim 1, characterized in that, The step of constructing an identity confidence map of the vehicle to be identified based on the evidence set, and performing evidence fusion and confidence calculation on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level, includes: The vehicle to be identified is determined as the vehicle entity node, and the trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence are respectively set as evidence nodes. Construct directed relation edges between each of the evidence nodes and the corresponding vehicle entity nodes, and determine the evidence quality score and sensor credibility as the edge weights of the directed relation edges to obtain the identity confidence graph; Based on the vehicle archive database, at least one vehicle candidate identity identifier and its evidence-identity matching degree are determined according to the evidence nodes and edge weights in the identity confidence graph. Based on Dempste-Shafer evidence fusion, the cumulative identity confidence is calculated when the vehicle entity node is the candidate identity identifier of each vehicle according to the evidence-identity matching degree; when different evidence leads to mutually exclusive identity inference, Dempste-Shafer evidence fusion requantifies conflicting identities based on evidence consistency, historical prior and evidence quality score and triggers candidate retention or manual verification process. When the cumulative identity confidence score exceeds a pre-configured threshold, the corresponding vehicle candidate identity identifier is used as the obtained vehicle identity identifier and its confidence score.

5. The method according to claim 1, characterized in that, The step of determining the corresponding digital twin model based on the vehicle identification identifier and binding the real-time location information associated with the UWB positioning evidence to the digital twin model to obtain digital twin update information includes: The current operating status and route requirements of the vehicle are determined based on the real-time location information and historical vehicle records; the historical vehicle records include vehicle registration information, owner information, and historical operating characteristics. The expected dwell time of a vehicle is predicted by a pre-trained prediction model based on the historical operating characteristics and the current operating status, and the expected dwell time information is obtained by combining the route requirements. The current operating state and the expected dwell time information are structured and mapped to obtain the digital twin update information.

6. The method according to claim 1, characterized in that, The process of generating train dispatch instructions based on the current track segment state constraints and safety constraints, and according to the updated digital twin information, includes: The current operating status of all vehicles and known vehicle scheduling operations are determined from the track event evidence, thus obtaining the current track segment state constraints; Based on the current track segment state constraints and safety constraints, a scheduling scheme that satisfies the constraints is obtained by using mixed integer linear programming and rolling optimization based on the updated digital twin information; the scheduling scheme includes movement paths, timing arrangements, and track equipment resources; The scheduling scheme is converted into a structured train scheduling instruction; the train scheduling instruction includes the scheduling start time, the track section number involved, and the route turnout.

7. The method according to claim 6, characterized in that, The method further includes: According to the train dispatching instruction, a safety token is requested from the track equipment execution unit; the track equipment includes a turnout controller; before issuing the token, the turnout controller verifies whether the associated track section is currently occupied, and if there is no occupancy, it issues the safety token. In response to receiving the security tokens issued by each of the track equipment execution units, the track equipment begins to execute the train dispatching instructions.

8. A smart dispatching device for dedicated railway lines based on automatic vehicle identification, characterized in that, The device includes: The evidence collection module is used to acquire multimodal evidence of the vehicle to be identified and to preprocess the multimodal evidence to obtain an evidence set. The evidence set includes each modal evidence and its corresponding evidence quality score. The multimodal evidence includes trackside RFID evidence, visual OCR evidence, UWB positioning evidence, acoustic-vibration fingerprint evidence, and track event evidence. The identity recognition module is used to construct an identity confidence map of the vehicle to be identified based on the evidence set, and to perform evidence fusion and confidence calculation on the vehicle to be identified based on the identity confidence map to obtain the vehicle identity identifier and its confidence level. The driving demand module is used to determine the corresponding digital twin model based on the vehicle identification if the confidence level exceeds a preset threshold, and bind the real-time location information associated with the UWB positioning evidence to the digital twin model to obtain digital twin update information; the digital twin model is constructed from the historical vehicle file corresponding to the vehicle identification; the digital twin update information includes the current operating status and expected stay information; The scheduling module is used to solve the scheduling problem based on the current track section status constraints and safety constraints, and according to the updated information of the digital twin, to generate train scheduling instructions; the train scheduling instructions are used to instruct the track line equipment to be adjusted according to the target scheduling; the track section status is updated by the track event evidence.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images