A metro intelligent maintenance system with brake energy consumption collaborative optimization

Abstract

The present application belongs to the field of artificial intelligence and rail transit intelligent operation and maintenance technology, and specifically relates to a subway intelligent maintenance system for collaborative optimization of braking energy consumption, aiming to solve the problems of low braking energy recovery efficiency, equipment perception lag and passive maintenance strategy. The system realizes the determination of regenerative braking energy absorbability and the quantitative evaluation of component health state by constructing a whole-network train dynamic topology and energy-health coupling analysis engine. Combined with a four-quadrant decision matrix, the system generates a hierarchical maintenance instruction and uses a rolling time domain optimization algorithm to coordinate train group operation time to improve energy matching degree. The system integrates a self-adaptive learning hub to continuously calibrate model parameters. The system realizes the collaborative optimization of braking energy efficiency improvement and key component life extension, and reduces the measured resistance energy consumption by more than 28% and the sudden failure rate by 41%.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent operation and maintenance technology of rail transit, and specifically relates to a subway intelligent maintenance system for coordinated optimization of braking energy consumption. Background Technology

[0002] As the backbone of modern urban public transportation, urban rail transit has always been a core concern in the industry regarding operational efficiency and energy consumption. With the continuous expansion of subway networks and the increasing density of train operations, the overall energy consumption of the system has increased significantly, with traction and braking processes constituting the most significant energy consumption stages. During the braking phase, traditional subway vehicles mainly use resistance braking or mechanical braking, converting a large amount of kinetic energy into heat dissipation, which not only wastes energy but also places an additional burden on tunnel ventilation and equipment cooling. In recent years, regenerative braking technology has provided a new path for energy conservation by converting braking kinetic energy into electrical energy fed back to the overhead contact line; however, its actual benefits are limited by energy utilization efficiency and system coordination levels.

[0003] Among them, the intelligent maintenance system for subway vehicles, as a key support for ensuring the safe and stable operation of trains, is gradually evolving from reactive maintenance to predictive maintenance. This system relies on onboard sensors, condition monitoring units, and an operation and maintenance management platform to achieve real-time perception and fault warning of the operating status of various train subsystems. However, under the current technological architecture, maintenance strategies are mostly focused on equipment reliability and safety indicators, with less consideration given to their impact on the overall energy consumption characteristics of the train, especially neglecting the dynamic correlation between changes in the braking system's state and energy consumption fluctuations. The timing of maintenance tasks, component replacement standards, and functional debugging parameters are often based on fixed cycles or threshold settings, lacking deep coupling with the energy management system.

[0004] Existing technologies have significant shortcomings in achieving coordinated optimization of braking energy consumption and maintenance decisions: the recovery efficiency of regenerative braking energy is constrained by fluctuations in overhead contact line voltage and the absorption capacity of nearby trains, making it difficult to maintain a stable high-efficiency range; the maintenance system lacks quantitative modeling of the changing trends of key state parameters such as brake disc wear, brake pad clearance, and inverter performance degradation, making it impossible to predict their impact on subsequent energy consumption; more importantly, maintenance scheduling and energy consumption management belong to different business systems, and data silos prevent them from forming a closed-loop feedback, causing potential energy-saving opportunities to be overlooked. Against this backdrop, how to break through the limitations of traditional maintenance models and construct an intelligent collaborative mechanism that can dynamically sense the braking system status, predict energy consumption evolution trends, and adaptively adjust maintenance strategies has become a pressing technical challenge for improving the overall energy efficiency of subway systems. Summary of the Invention

[0005] The purpose of this invention is to provide a smart subway maintenance system with coordinated optimization of braking energy consumption, addressing the current problems of low braking energy recovery efficiency, lagging equipment status perception, passive maintenance strategy response, and uneven energy utilization caused by differences in the operating conditions of multiple trains in subway operations. With the continuous expansion of urban rail transit networks, the enormous energy consumption resulting from frequent train starts and stops has become a key bottleneck restricting green and low-carbon development. Existing braking energy recovery technologies mostly focus on instantaneous energy feedback control for single trains, lacking accurate prediction and coordinated adjustment capabilities for dynamic fluctuations in traction network voltage. This results in recovered energy not being effectively absorbed by other trains and being largely dissipated through resistance. Simultaneously, traditional maintenance models rely on periodic maintenance or emergency repairs after failures, failing to integrate the health status of the power system with energy consumption behavior characteristics for joint diagnosis, leading to both misallocation of maintenance resources and potential safety hazards. Furthermore, the highly uneven operating density and load status of train groups in different sections and time periods further exacerbate the spatial and temporal mismatch of energy supply and demand. These multiple contradictions intertwine, causing the overall system energy efficiency improvement to reach a bottleneck.

[0006] The technical solution of this invention is to construct a closed-loop intelligent subway maintenance system integrating state perception, energy consumption modeling, collaborative scheduling, and intelligent decision-making, achieving the dual goals of efficient reuse of braking energy and extended lifespan of key components. The system is deployed in a regional control center and achieves real-time data interaction with all trains on the entire line through onboard units and trackside communication facilities. After system startup, a dynamic topology model of the entire train network is first established. Based on global positioning information and timetable data, the system tracks the position, speed, acceleration, and traction / braking command status of each train within the section in real time. Furthermore, the system connects to the voltage and current sensor arrays at the output terminals of the regenerative braking units of each train, collecting the energy amplitude, duration, and power curve of each braking event, and simultaneously acquiring voltage monitoring point data from the overhead contact line feeder to identify the local overvoltage risk level of the traction network. As one embodiment of the present invention, the mechanism for determining the absorbability of regenerative braking energy is as follows: the system calculates the probability of effective transfer of braking energy based on the current grid voltage level, the distance to the nearest preceding train, and the traction power demand window; when the grid voltage exceeds a preset safety threshold and there is no receiving train in traction condition ahead, it is determined to be a high dissipation risk event, triggering a priority warning signal. Further, the system integrates multi-source vibration, temperature, and acoustic emission sensing subsystems for the traction motor, gearbox, and brake disc, continuously collecting their operating noise spectrum, root mean square value of axial vibration acceleration, and surface thermal distribution gradient field. As one embodiment of the present invention, the component health index quantification model adopts a hierarchical weighted fusion algorithm, assigning the kurtosis index of the vibration signal a first weight coefficient, the temperature gradient change rate a second weight coefficient, and the cumulative energy of the acoustic emission event a third weight coefficient. The three are weighted and summed to generate a real-time health score; when the score is lower than a set degradation threshold, it enters the degradation tracking queue.

[0007] Furthermore, the system establishes an energy consumption-health coupling analysis engine to analyze the nonlinear correlation between braking intensity and wear rate of mechanical components. As one embodiment of the invention, the coupling relationship modeling process is as follows: Data samples of the number of emergency braking operations per 100 kilometers and the corresponding traction motor bearing replacement cycle for the same vehicle model under similar load conditions are extracted from the historical database; the original data are projected to a high-dimensional feature space through kernel function mapping, and an empirical decay curve is fitted, showing that for every one increase in braking frequency, the expected remaining bearing life is shortened by 0.7%; this curve serves as one of the basic parameters for subsequent predictive maintenance cycle adjustments. Furthermore, the system is configured with a dynamic maintenance task planner, receiving life prediction results from the coupling analysis engine and high dissipation risk warnings from the energy consumption assessment module, and comprehensively generating a graded intervention strategy. In one embodiment of the present invention, the planner executes a four-quadrant decision matrix: the horizontal axis represents the component health score, and the vertical axis represents the average daily braking energy dissipation rate of the power supply section. Target trains falling into the upper left quadrant (low health + high dissipation) are marked as special-grade maintenance objects, immediately arranged for inspection in the depot, and their braking distribution ratio in the timetable is adjusted simultaneously. Trains falling into the lower right quadrant (high health + low dissipation) have their maintenance cycle extended to release maintenance resources. Furthermore, the system introduces an energy scheduling coordination mechanism between train groups, which, under the premise of meeting the safe operating interval, actively fine-tunes the interval running time of subsequent trains, so that their traction acceleration phase and the braking period of the preceding train precisely overlap in time and space. As one embodiment of the present invention, the coordination mechanism is implemented through a rolling time-domain optimization algorithm: with the next 15 minutes as the prediction window, a state vector set containing N trains is constructed, and the objective function is set as the weighted sum of minimizing the total energy consumption of the entire network resistance and maximizing the operating condition matching degree; the constraints include the minimum headway, the maximum allowable delay time and the vehicle's own energy consumption limit; the optimal speed trajectory obtained by the solution is sent to the corresponding train automatic driving system for execution via a wireless communication link.

[0008] Furthermore, the system is equipped with an adaptive learning hub for continuously calibrating the parameter drift of the energy consumption prediction model and the health degradation model. In one embodiment of the invention, the adaptive learning process is as follows: after each actual maintenance and disassembly inspection report, residual analysis is performed between the measured wear degree and the system's previous prediction values; if the absolute error exceeds 10% for three consecutive times, the model parameter backpropagation correction program is initiated, adjusting the slope factor and intercept term in the aforementioned empirical decay curve; simultaneously, combined with humidity and precipitation records from the meteorological database, the influence weight of environmental factors on the brake disc friction coefficient is analyzed, and the friction model parameter library in the energy consumption simulation module is dynamically updated. Furthermore, the system sets up multi-level human-computer interaction interfaces, providing differentiated views for different functional roles. In one embodiment of the invention, the dispatching and command personnel terminal displays a heat map of energy flow across the entire network and voltage trend curves of key nodes, supporting manual injection of virtual trains to test the effectiveness of dispatching strategies; the maintenance management personnel terminal presents a list of vehicles to be inspected aggregated by station and their corresponding energy consumption contribution sorting, assisting in the formulation of weekly schedules; vehicle technicians receive personalized work instructions via handheld terminals, containing historical vibration spectrum comparison diagrams of target components and recommended torque parameter tables. Furthermore, the system employs a dual-channel redundant communication architecture to ensure data transmission reliability. The primary channel utilizes the 5G-U communication protocol to achieve high-bandwidth, low-latency transmission, while the backup channel uses the enhanced LTE-M protocol to maintain basic status reporting functionality. In one embodiment of the invention, the dual-channel switching logic is triggered based on a comprehensive link quality score: when the average packet loss rate of the primary channel exceeds 3% for 5 consecutive seconds or the end-to-end latency exceeds 200 milliseconds, it automatically switches to the backup channel and issues a link alarm; after recovery, bidirectional data retransmission and consistency verification are performed to ensure the integrity of historical records.

[0009] Furthermore, the system incorporates a digital twin simulation platform for energy efficiency pre-assessment and emergency response simulation before new lines are commissioned. In one embodiment of the invention, the simulation platform loads a three-dimensional geographic information model of the line, a substation distribution map, and a typical passenger flow OD matrix to simulate the annual cumulative power consumption and peak grid voltage frequency under different combinations of train density. For identified weak power supply sections, mobile energy storage devices are deployed in advance, or it is suggested to add inverter feedback stations. Furthermore, the system supports a cross-line data federated learning framework, sharing component degradation mode feature vectors in non-sensitive layers while protecting the data privacy of each operator. In one embodiment of the invention, the federated learning process is as follows: each participant locally trains and generates gradient update packages, which are then homomorphically encrypted and uploaded to a trusted third-party aggregation server; the server performs gradient averaging in encrypted form, generates global model update parameters, and broadcasts them back to all parties; each party decrypts the data locally and integrates it into the original model, achieving iterative upgrades of collective intelligence without exposing the original data.

[0010] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This solution, for the first time, integrates regenerative braking efficiency and the health management of core transmission components into a unified decision-making framework, breaking the traditional separation of energy efficiency optimization and equipment maintenance as two separate technical approaches. By establishing a quantitative mapping relationship between braking intensity and mechanical wear rate, each high-energy-consumption braking action is not only considered an energy event but also a key indicator reflecting the intensity of equipment use, thus achieving a paradigm shift from "post-event repair" to "pre-event prevention." The proposed inter-train group spatiotemporal collaborative scheduling mechanism can significantly improve the online utilization rate of regenerative braking energy without increasing hardware investment. Actual measurement data shows that it can reduce resistance energy consumption by more than 28% during peak hours. This mechanism achieves precise matching of energy supply and demand by fine-tuning runtime, avoiding negative impacts on passenger service levels. The four-quadrant decision matrix designed in this solution conducts a two-dimensional joint assessment of equipment health status and regional energy consumption levels, enabling limited maintenance resources to be allocated to the most needed aspects. Field verification shows that the sudden failure rate of key components has decreased by 41%, while the average maintenance cycle has been extended by 19%, significantly improving the economic efficiency of operation and maintenance. This solution integrates an adaptive learning hub and a digital twin platform, endowing the system with continuously evolving engineering intelligence, which can automatically respond to long-term disturbances such as vehicle aging and environmental changes, ensuring the long-term stability of model prediction accuracy and extending the effective service life of the system. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of energy consumption-health coupling analysis and four-quadrant decision-making in this invention. Detailed Implementation

[0012] Please refer to Figure 1 and Figure 2This invention relates to a smart subway maintenance system for collaborative optimization of braking energy consumption. The system aims to construct a closed-loop technical architecture that deeply integrates train operation status perception, dynamic assessment of regenerative braking energy, health monitoring of key transmission components, and a multi-level collaborative decision-making mechanism. Deployed at the regional control center, the system establishes bidirectional real-time connections with the onboard computing units and trackside data acquisition nodes of all subway trains along the entire line via a highly reliable communication network. Upon system startup, a global initialization process is first executed, loading the daily timetable plan, line topology parameters, substation location distribution, contact network resistivity model, and power system design specifications database for each train model. Based on this, the system continuously receives operational data from each online train, including global positioning information, real-time speed vector, acceleration rate of change, traction output commands, electric braking engagement ratio, and air brake trigger status. Combined with timetable calculations, the system dynamically reconstructs the spatial relative positions of all trains in the network at any given time, forming a dynamic topology model of the train group that is updated every minute.

[0013] This dynamic topology model not only records the current section of each train, its distance to the next station, and its estimated arrival and departure times, but also analyzes its expected operating condition transition sequence in the next operating cycle, including the time window for entering the braking phase, the estimated deceleration curve, the accuracy requirements for the target stopping position, and possible temporary speed limit interventions. Simultaneously, the system connects to the voltage sensor and current transformer array at the DC bus end of each train's regenerative braking unit via a dedicated high-speed data channel. It collects the energy amplitude, instantaneous power peak, total energy integral, and its temporal distribution characteristics fed back to the traction network for each braking event with a sampling period of 100 milliseconds. All collected electrical parameters are precisely timestamped and associated with the corresponding train's identification, direction of travel, and passenger capacity level (estimated based on door opening / closing records and passenger counting system), constructing a structured braking energy event database.

[0014] To further evaluate the actual usable value of the braking energy, the system simultaneously acquires real-time data from voltage monitoring points set at the feeder ends of each power supply section along the line to determine whether the local traction network is in an overvoltage critical state. As one of the core functions of this invention, the system has a built-in regenerative braking energy absorbability determination mechanism, the determination logic of which relies on the fusion analysis of information in three spatiotemporal dimensions: the first dimension is the current traction network voltage level. When the voltage of a power supply section is detected to rise above 1850 volts, it is considered to be close to the safety limit. At this time, any additional feedback energy will greatly increase the risk of resistance dissipation; the second dimension is the position and operating status of the adjacent trains. Based on the aforementioned dynamic topology model, the system identifies all trains in the same power supply section in front of the target braking train and queries whether they are currently in the traction acceleration or constant speed climbing stage, i.e., they have the ability to absorb external electrical energy; the third dimension is the energy demand duration window of the potential receiving train. Based on its current speed, target speed, and remaining acceleration distance, the system predicts the length of its traction power maintenance period.

[0015] Based on the above three inputs, the system calculates the probability value of the effective transfer of braking energy, denoted as . The specific calculation process is as follows: If the current network voltage is below 1750 volts, then... A direct value of 0.95 indicates that the power grid has sufficient absorption capacity; if the grid voltage is between 1750 volts and 1850 volts, the value is determined based on the number of trains ahead that are in traction operation. The linear decay process is performed, and the expression is: in, This indicates the probability that the braking energy is effectively transferred in this instance. This indicates the number of trains in traction mode within the same power supply zone ahead. This formula shows that under medium-high voltage conditions, even with receiving trains, the absorption efficiency is limited. However, as the number of receiving trains increases, the overall transfer probability gradually increases, with an upper limit set at 0.9 to reflect the physical limit. If the grid voltage exceeds 1850 volts and there are no traction trains ahead, then... Forced zeroing is classified as a high-dissipation-risk event.

[0016] When the system determines that a braking event falls into the high-dissipation-risk category, it immediately generates a priority warning signal and incorporates it into the input queue for subsequent scheduling, coordination, and maintenance decisions. Simultaneously, the system runs a component health monitoring subsystem in parallel. This subsystem integrates a multi-source sensor network installed on the traction motor base, gearbox housing, and brake disc bracket. Sensor types include triaxial microelectromechanical system vibration accelerometers, non-contact infrared thermal imager arrays, and broadband acoustic emission probes. All sensors are connected to the local edge computing module via a heterogeneous protocol. After preliminary filtering and feature extraction, key indicators are uploaded to the regional control center.

[0017] The vibration accelerometer captures the transient vibration response of the equipment casing in the X, Y, and Z directions at a sampling rate of 10 kHz. The system performs joint time-domain and frequency-domain analysis on the signal, extracting five core statistical features: root mean square value of axial vibration acceleration, peak factor, impulse index, margin index, and kurtosis. Among these, kurtosis, as a key indicator for measuring the intensity of the impact component in the signal, is assigned the first weighting coefficient. Because it is extremely sensitive to early pitting corrosion faults, the infrared thermal imager scans the surface temperature field distribution of the traction motor stator winding ends, gearbox lubricating oil outlet pipe, and brake disc friction surface at a rate of one frame every 5 seconds. The system calculates the maximum temperature rise gradient between two adjacent frames. It is used as an indicator of the rate of change of temperature gradient and assigned a second weighting coefficient. The acoustic emission probe monitors the elastic wave signals generated by the propagation of micro-cracks or abnormal slippage of friction pairs within materials. The system sets an energy threshold and only records events with amplitudes exceeding 45 dB, accumulating the total energy released per unit time. Assign a third weighting coefficient .

[0018] After normalization, the above three weighted indicators are substituted into a hierarchical weighted fusion algorithm to generate a real-time health score. The normalization method uses the range method to map the original index to the [0,1] interval, as shown in the following formula: in, Provides real-time health scores; This is the current kurtosis value; This represents the current rate of change of the temperature gradient; This represents the current cumulative acoustic emission energy; the subscripts "min" and "max" represent the historical minimum and maximum observed values ​​for this vehicle model during its normal service life, respectively, derived from statistical results from the maintenance database of the same model over the past two years. The scoring results are updated every 30 seconds. If the device falls below the set degradation threshold of 0.65 for five consecutive times, the system will automatically mark the device as a degradation tracking object, start the long-term trend analysis program, and push an initial alarm to the maintenance task planner.

[0019] To further reveal the intrinsic relationship between braking operation modes and the lifespan loss of mechanical components, an energy consumption-health coupling analysis engine was established in the system. This engine extracted data samples from historical databases for the same vehicle model under similar ambient temperature and average load conditions (defined as a full load rate of 70% ± 5%), collecting a data set of 128 valid data points, including the number of emergency braking events (N_brake) per 100 kilometers and the corresponding actual replacement cycle (L_bearing) of the traction motor bearing. To eliminate noise interference, the system performed local weighted regression smoothing on the raw data, and then used a Gaussian kernel function to map the input variable N_brake to a high-dimensional feature space, fitting a nonlinear decay curve in this space. The fitting results show that the expected reduction rate of the bearing's remaining life (ΔL_per_brake) is approximately linearly related to the increase in braking frequency, with a slope of 0.7% reduction in the expected bearing life for every one additional emergency braking event per 100 kilometers. This empirical decay curve was solidified as one of the system's fundamental parameters for dynamic adjustment of subsequent predictive maintenance cycles.

[0020] The dynamic maintenance task planner, acting as the system's decision-making center, receives life prediction results from the coupled analysis engine, high dissipation risk warnings from the energy consumption assessment module, degradation alarms from the health monitoring subsystem, and runtime constraints from the scheduling system, and comprehensively generates a tiered intervention strategy. The planner executes a four-quadrant decision matrix model, where the horizontal axis represents the component health score. The system divides the power supply interval into two ranges: high health (≥0.65) and low health (<0.65). The vertical axis represents the daily average braking energy dissipation rate R_dissipation of the power supply interval, which is divided into two ranges: high dissipation (>35%) and low dissipation (≤35%). This forms four decision quadrants. Trains falling into the upper left quadrant (low health + high dissipation) are marked as special-level maintenance targets. The system immediately generates an emergency repair order, arranges for them to enter the depot for comprehensive inspection after completing their current route, and simultaneously adjusts the braking distribution ratio in their timetable, forcibly reducing the intensity of electric braking during peak hours and moderately increasing the proportion of air braking to alleviate grid pressure. Trains falling into the upper right quadrant (high health + high dissipation) are assigned to participate in the energy dispatch coordination mechanism, and are given priority to assume the role of traction load during specific periods to improve energy matching. Trains falling into the lower left quadrant (low health + low dissipation) are included in the regular maintenance queue and perform maintenance according to the original cycle. Trains falling into the lower right quadrant (high health + low dissipation) have their current maintenance cycle automatically extended by 15%, freeing up maintenance resources for higher priority tasks.

[0021] To achieve efficient reuse of regenerative braking energy, the system introduces an energy scheduling and coordination mechanism among train groups. This mechanism, under strict constraints such as a minimum headway of 90 seconds, a single train's allowed delay of no more than 2 minutes, and vehicle energy consumption not exceeding 10% of the rated baseline, proactively fine-tunes the interval running times of subsequent trains, ensuring precise temporal and spatial overlap between their traction acceleration phase and the braking period of the preceding train. The coordination process is implemented using a rolling time-domain optimization algorithm. Using a 15-minute prediction window, a state vector set S(t) containing N trains is constructed. Each state vector includes four elements: position, speed, acceleration, operating mode, remaining braking energy potential, and traction power demand. The objective function is set as the weighted sum of minimizing the total network resistive energy consumption E_resist and maximizing the operating mode matching degree M_match, with weight coefficients α and β set to 0.7 and 0.3 respectively, reflecting the energy-saving priority principle. The constraint set C includes: minimum tracking interval constraint, maximum allowable speed deviation, traction network voltage fluctuation range limit, train dynamic performance boundary, and passenger comfort acceleration threshold.

[0022] An improved particle swarm optimization algorithm was used to optimize the solution, with a population size of 50. The iteration termination condition was that the optimal fitness value changed by less than 0.1% over 10 consecutive generations. The obtained optimal speed trajectory V_opt(t) was transmitted to the corresponding train's automatic driving system via a 5G-U wireless communication link, which then executed speed curve tracking control. To ensure consistent command execution, a feedback verification mechanism was implemented, receiving the actual train trajectory data every 30 seconds. If the deviation exceeded the allowable range of ±1.5%, the optimization calculation was restarted and the command was updated. Experimental results showed that this mechanism could reduce resistive energy consumption by more than 28% during peak hours without causing a significant cumulative delay effect.

[0023] The adaptive learning center is responsible for continuously calibrating the parameter drift of each prediction model within the system. After each actual maintenance operation is completed and a disassembly and inspection report is obtained, the system performs residual analysis between the measured wear degree (e.g., bearing raceway spalling area, gear tooth surface pitting depth, brake disc thickness reduction) and the remaining life estimate predicted earlier by the system based on health scores and coupled models. If the absolute error exceeds 10% for three consecutive times, the model parameter backpropagation correction procedure is initiated, adjusting the slope factor k and intercept term b in the empirical decay curve, and updating the formula as follows: ,in, This is the updated slope factor. For the current (old) slope factor, With a learning rate of 0.05, The average error ratio is used. Simultaneously, the system combines hourly humidity, precipitation intensity, and temperature records from the meteorological database to analyze the weighting of environmental factors on the brake disc friction coefficient μ. Through multiple linear regression modeling, it is found that for every 10 percentage point increase in relative humidity, μ decreases by an average of 0.03; under rainfall conditions, μ decreases by an additional 0.05; and under low temperature conditions (<5 degrees Celsius), μ increases by 0.02. These correction coefficients are dynamically injected into the friction model parameter library in the energy consumption simulation module to ensure the long-term stability of braking energy estimation accuracy.

[0024] The human-computer interaction interface adopts a multi-level role-adaptive architecture, providing differentiated views for users with different functions. The dispatching and command personnel terminal displays a heat map of energy flow across the entire network, using color intensity to represent the net energy inflow and outflow intensity of each power supply zone per unit time, and overlays voltage trend curves of key nodes. It supports manually injecting virtual trains to test voltage fluctuation responses under different dispatching strategies. The maintenance management personnel terminal displays a list of vehicles to be inspected aggregated by station. The list items include train number, health score, time of the most recent high dissipation event, recommended maintenance level, and energy consumption contribution ranking index, assisting in the formulation of weekly schedules. Vehicle technicians receive personalized work instructions on handheld terminals. The documents embed historical vibration spectrum comparison charts of target components (current spectrum and factory reference spectrum), recommended disassembly and assembly torque parameter tables, lubricant type descriptions, and typical fault case collections, improving the standardization of on-site operations.

[0025] The communication system employs a dual-channel redundancy architecture to ensure data transmission reliability. The primary channel, relying on the 5G-U communication protocol, achieves high-performance transmission with a downlink bandwidth of 300 Mbps, an uplink bandwidth of 150 Mbps, and an end-to-end latency of less than 50 milliseconds within a dedicated frequency band. This is used to carry high-definition video streams, large-capacity sensor data packets, and real-time control commands. The backup channel uses the enhanced LTE-M protocol, providing basic communication capabilities of 5 Mbps downlink and 2 Mbps uplink, maintaining the periodic reporting functions of train identification, critical alarm signals, and basic operating status. The dual-channel switching logic is triggered based on a comprehensive link quality score, which is calculated by weighting four indicators: average packet loss rate, end-to-end latency, signal strength, and signal-to-noise ratio over a continuous 5-second period. When the primary channel score falls below a preset threshold (corresponding to a packet loss rate higher than 3% or a latency exceeding 200 milliseconds), the system automatically switches to the backup channel and issues a link alarm. After the primary channel recovers, bidirectional data retransmission and consistency verification are performed, using a timestamp alignment mechanism to fill in missing data segments and hash verification to ensure the integrity of historical records.

[0026] The digital twin simulation platform, serving as a forward-looking capability support module, is used for energy efficiency pre-assessment and emergency response simulation before the opening of new lines. The platform loads a 3D geographic information model of the line, including spatial attributes such as gradient, curvature, and tunnel ventilation conditions; a substation distribution map annotates rectifier unit capacity, inverter feedback device configuration, and energy storage unit deployment; and a typical passenger flow OD matrix describes the passenger flow distribution between origin and destination stations at different times. The system simulates the annual cumulative power consumption and peak grid voltage frequency under different train density combinations (from 4-minute intervals during off-peak to 1.5-minute intervals during peak), identifying weak power supply sections prone to frequent starts and stops due to large terrain undulations and short station spacing. For these sections, the platform recommends the early deployment of mobile energy storage devices or the addition of inverter feedback stations, and quantifies the investment payback period and carbon emission reduction benefits of the renovation plan.

[0027] The cross-line data federated learning framework supports the sharing of component degradation mode feature vectors in non-sensitive layers while protecting the data privacy of each operator. The federated learning process is coordinated and executed by a trusted third-party aggregation server. Each participant trains and generates gradient update packages based on its accumulated equipment monitoring data in its local training environment. The gradient vectors are then encrypted using the Paillier homomorphic encryption algorithm to ensure that the original data remains within the domain. The encrypted gradient packages are uploaded to the aggregation server, which performs gradient averaging in the encrypted state to generate global model update parameters. These parameters are broadcast back to all parties, and each local node decrypts them and merges them with the existing model, completing a round of collective intelligence iterative upgrade. The entire process does not expose any original observation data, complying with the stringent data security standards of the rail transit industry.

[0028] The overall system operation constitutes a multi-layered closed-loop feedback system. The bottom layer is the real-time data perception loop, continuously collecting data on train operating status, electrical parameters, and mechanical health indicators. The middle layer is the analysis and decision-making loop, covering energy consumption assessment, health scoring, coupled modeling, and four-quadrant decision-making. The upper layer is the scheduling and execution loop, involving energy coordination, maintenance task assignment, and timetable fine-tuning. The top layer is the evolutionary learning loop, continuously optimizing model parameters through adaptive calibration and federated learning. These four loops progress step-by-step and feedback to each other, forming an autonomous system with engineering intelligence. For example, when construction in a section causes a temporary speed reduction leading to an increase in braking frequency, the system not only detects a downward trend in the relevant train health score but also identifies an increase in the dissipation rate of that section. This triggers a special maintenance warning and initiates a scheduling coordination mechanism to optimize the subsequent train operation rhythm. Finally, through the adaptive learning hub, the characteristics of this event are incorporated into the model training set, improving the system's ability to respond to similar future scenarios.

[0029] Existing technologies generally build energy management systems and equipment maintenance systems independently. The former focuses on instantaneous power balance, while the latter emphasizes periodic component replacement. These two systems lack data sharing and collaborative decision-making. This invention achieves joint modeling of energy consumption behavior and health status by establishing a unified data platform and a shared feature space. Each braking action is no longer merely an energy event, but also a historical imprint reflecting the intensity of equipment use. The system can identify drivers who habitually use high-deceleration braking, which, while achieving rapid stopping, significantly exacerbates gearbox impact loads. This generates a driving behavior evaluation report in the background, driving optimization of passenger service management. Similarly, some routes naturally create a favorable "front vehicle braking, rear vehicle traction" working condition combination due to their gradient design. The system can then formulate differentiated energy consumption assessment indicators based on this, incentivizing operating units to leverage the inherent advantages of the routes.

[0030] At the information security level, the system employs a domain isolation strategy, physically separating control command channels from data acquisition channels. Critical command execution requires dual authentication. All remotely issued dispatch commands are accompanied by digital signatures, and the receiving end verifies the source's legitimacy through public key infrastructure. The database implements hierarchical access control; maintenance personnel can only view data for vehicles assigned to their specific segment, and dispatchers cannot access component-level vibration spectrum raw files. Audit logs fully record all operational activities and are retained for no less than three years, meeting industry regulatory requirements.

[0031] The system boasts excellent scalability, supporting the integration of new energy storage devices, compatibility with hydrogen hybrid trains, and adaptation to fully automated driverless scenarios through the addition of plug-in modules. Hardware interfaces are reserved for multiple industry standard protocols such as Modbus TCP, IEC 61850, and CANopen. The software architecture is based on microservice design, with each functional module deployed independently and communicating loosely, facilitating future feature iterations and version upgrades. All core algorithms are encapsulated as containerized services, enabling flexible migration to private or hybrid cloud environments and adapting to the existing IT infrastructure of different urban rail transit companies.

[0032] This embodiment, through the systematic integration of the aforementioned technical means, constructs a subway intelligent maintenance system with deep perception, precise assessment, intelligent decision-making, and autonomous evolution capabilities. Its innovation lies in the deep integration of the traditionally separate fields of energy efficiency optimization and equipment health management, using the common behavior of braking operation as a bridge to establish a quantitative connection between energy flow and material flow. The system not only significantly reduces resistive energy consumption and improves the utilization rate of regenerative energy, but also, through source analysis of high-energy-consumption braking behavior, can proactively identify potential mechanical failure risks, achieving a fundamental shift from passive maintenance to proactive prevention. Actual measurement data shows that after applying this system, the sudden failure rate of key components decreased by 41%, the average maintenance cycle was extended by 19%, and the economic efficiency of operation and maintenance was substantially improved. More importantly, through spatiotemporal coordinated scheduling among train groups, the system unlocks enormous energy-saving potential without additional hardware investment, providing a replicable technical path for the green and low-carbon development of urban rail transit.

[0033] The specific implementation of the energy consumption-health coupling analysis engine is as follows: The system extracts data samples from historical databases of the number of emergency braking operations per 100 kilometers for the same vehicle model under similar load conditions, along with the corresponding traction motor bearing replacement cycle. The raw data is projected onto a high-dimensional feature space using a kernel function mapping, and an empirical decay curve is fitted to show that for every one increase in braking frequency, the expected remaining bearing life is shortened by 0.7%. This curve serves as one of the basic parameters for subsequent predictive maintenance cycle adjustments. The core objective of this process is to establish a quantifiable relationship between braking intensity and mechanical wear, enabling the system to predict the remaining service life of key components based on daily operating data. The data sample collection period covers the past 24 months, eliminating abnormal data points caused by abnormal operating conditions (such as emergency braking in severe weather or emergency stops due to signal failures). The radial basis function is selected as the kernel function used for projection. ,in, The output value of the kernel function. For the input sample vector, The square of the Euclidean distance. The kernel width parameter was determined to be optimally 0.8 using cross-validation. Support vector regression was used to fit the decay curve in high-dimensional space, employing ε-insensitive loss as the loss function and setting the regularization parameter C to 1.2. After fitting, the curve parameters were fixed as entries in the system's knowledge base, and a quarterly automatic review mechanism was implemented to ensure the model's timeliness.

[0034] The execution logic of the four-quadrant decision matrix is ​​as follows: the horizontal axis represents the component health score, and the vertical axis represents the average daily braking energy dissipation rate of the power supply section. Target trains falling into the upper left quadrant (low health + high dissipation) are marked as special-grade maintenance objects, immediately arranged for inspection in the depot, and their braking distribution ratio in the timetable is adjusted simultaneously. Trains falling into the lower right quadrant (high health + low dissipation) have their maintenance cycle extended to release maintenance resources. The essence of this mechanism is to divide the two-dimensional evaluation space into four typical operating condition combinations, each corresponding to a different resource allocation strategy. A low health state reflects that the equipment has entered a period of performance degradation and requires key monitoring; a high dissipation rate indicates low regional energy utilization efficiency and room for optimization. The combination of the two constitutes the highest priority scenario, which must be intervened immediately. Conversely, a combination of high health and low dissipation represents that the system is operating in an ideal state, and the maintenance frequency can be appropriately relaxed to save costs. The decision boundary setting is not fixed but dynamically adjusted according to the season, traffic volume, and equipment batches. For example, during the high-temperature period in summer, the health score threshold is temporarily lowered to 0.6 to cope with the additional stress caused by the decline in lubrication performance; during the Spring Festival travel rush, the dissipation rate threshold is raised to 40% to avoid excessive intervention affecting transportation capacity.

[0035] The specific implementation of the rolling time-domain optimization algorithm is as follows: A state vector set containing N trains is constructed using a 15-minute prediction window. The objective function is set as the weighted sum of minimizing the total resistive energy consumption of the entire network and maximizing the operating condition matching degree. Constraints include minimum headway, maximum allowable delay time, and vehicle energy consumption limits. The optimal speed trajectory obtained is transmitted to the corresponding train automatic driving system via a wireless communication link for execution. The algorithm performs a rolling update every 2 minutes to ensure that decisions closely follow actual operational changes. Each train's state vector set contains 7 dimensions: current position, current speed, current acceleration, target station name, estimated arrival time, operating mode (traction / coasting / braking), and remaining braking energy potential. The total resistive energy consumption in the objective function is calculated by integrating the energy dissipation of each power supply zone exceeding the safe voltage threshold. The operating condition matching degree is defined as the weighted average of the time overlap rate between the braking period of the preceding train and the traction period of the following train. The weight coefficients are dynamically adjusted according to the train's passenger load, with heavier-load trains receiving higher matching weights. The constraints include a minimum headway of 90 seconds, a maximum allowable delay of 120 seconds, and vehicle energy consumption not exceeding 110% of the baseline value. The solver employs an interior-point method for nonlinear programming, with the initial solution obtained by shifting the optimal solution from the previous time window, accelerating the convergence process. The issued commands consist of a target speed-time series matrix, spaced 10 seconds apart, for a total of 90 sets, covering the entire next 15 minutes.

[0036] The adaptive learning process is implemented as follows: After each actual maintenance and disassembly inspection report, residual analysis is performed between the measured wear level and the system's previous predictions. If the absolute error exceeds 10% for three consecutive times, the model parameter backpropagation correction procedure is initiated, adjusting the slope factor and intercept term in the aforementioned empirical decay curve. Simultaneously, combined with humidity and precipitation records from the meteorological database, the influence weight of environmental factors on the brake disc friction coefficient is analyzed, and the friction model parameter library in the energy consumption simulation module is dynamically updated. The residual analysis uses root mean square error (RMSE) and mean absolute percentage error (MAPE) as dual indicators for evaluation. When MAPE exceeds 10% for three consecutive times, the correction mechanism is triggered. The slope factor k is updated using the gradient descent method, with a learning rate of 0.02 and the direction being the negative gradient of the error. The intercept term b is adjusted synchronously to maintain the physical rationality of the curve at the zero point. The influence weight of environmental factors is obtained through generalized linear regression (GLM), with explanatory variables including relative humidity, precipitation intensity, temperature, and wind speed, and the response variable being the observed value of the brake disc friction coefficient. The regression results are updated monthly, generating a new parameter mapping table which is then injected into the simulation module. All model update operations must be recorded by the version control system, supporting rollback and audit trail.

[0037] The specific implementation of the dual-channel switching logic is as follows: When the average packet loss rate of the primary channel exceeds 3% for 5 consecutive seconds or the end-to-end latency exceeds 200 milliseconds, the system automatically switches to the backup channel and issues a link alarm. After recovery, bidirectional data retransmission and consistency verification are performed to ensure the integrity of historical records. The switching decision is executed by the communication management agent module, which collects link quality indicators every second, including IP layer packet loss rate, RTT latency, RSRP signal strength, and SINR signal-to-noise ratio. If any indicator exceeds the standard within 5 consecutive seconds, the switching process is initiated. The switching process includes three stages: the first stage is command freezing, suspending the issuance of all non-emergency control commands; the second stage is channel switching, closing the primary channel session and establishing a backup channel connection; the third stage is state synchronization, requesting retransmission of data that has not been acknowledged in the past 60 seconds through incremental data packets. Consistency verification uses the SHA-256 hash algorithm to generate a digest for the retransmitted data block and compare it with the source end to ensure that the content has not been tampered with. After the primary channel recovers, the system will not immediately switch back, but will continuously monitor its stability for 5 minutes before performing a reverse switch to prevent frequent fluctuations.

[0038] The digital twin simulation platform is implemented as follows: It loads a 3D geographic information model of the railway line, a substation distribution map, and a typical passenger flow OD matrix to simulate the annual cumulative power consumption and the frequency of peak grid voltage occurrence under different combinations of train density. For identified weak power supply sections, it pre-deploys mobile energy storage devices or suggests adding inverter feedback stations. The platform uses the Unity3D engine to build a visual scene, and the underlying physical simulation is implemented based on the MATLAB / Simulink power system toolbox. The railway line model contains more than 2000 spatial nodes, each defining elevation, radius of curvature, track resistance coefficient, and contact network resistance per unit length. The substation model includes rectifier transformer capacity, inverter feedback efficiency curves, and a battery SOC-power mapping table. The passenger flow OD matrix is ​​divided into 24 time periods by hourly granularity, with each time period defining passenger flow between 100 major stations. The simulation step size is set to 1 second, and the annual simulation takes approximately 6 hours. Output indicators include: total traction energy consumption, resistive energy dissipation, regenerative energy utilization rate, and the frequency and location of maximum grid voltage occurrence. The criteria for identifying weak sections are: the average daily peak grid voltage exceeds 1850 volts more than 30 times per year, and the resistance dissipation ratio is higher than 40%. For such sections, the platform recommends solutions and estimates the investment return cycle to assist decision-makers in formulating infrastructure renovation plans.

[0039] The specific implementation of the federated learning process is as follows: Each participant generates a gradient update package locally, which is then homomorphically encrypted and uploaded to a trusted third-party aggregation server. The server performs gradient averaging in the encrypted state, generates global model update parameters, and broadcasts them back to each participant. Each participant decrypts the data locally and integrates it into the original model, achieving iterative upgrades of collective intelligence without exposing the original data. Local training uses a lightweight convolutional neural network (CNN) to extract fault feature vectors from vibration signals. The network structure includes 3 convolutional layers, 2 pooling layers, and 1 fully connected layer, with a total of less than 50,000 parameters, making it suitable for edge device deployment. Homomorphic encryption uses the open-source library SEAL to implement the Paillier encryption scheme, supporting additive homomorphic and scalar multiplication operations. The gradient vector is encrypted locally and uploaded. The server performs the ∑Enc(g_i) / N operation to obtain the average gradient, then encrypts and distributes it. Each participant decrypts the data and executes the corresponding steps. Complete the model update. Among them, For the updated model parameters, For the old model parameters, This refers to the learning rate. To achieve a global average gradient, the entire process is executed monthly to ensure that the frequency of knowledge sharing matches the speed of data accumulation. The server does not store any raw gradient data, but only performs real-time aggregation operations, maximizing the protection of the data sovereignty of the participants.

[0040] The specific implementation of the human-computer interaction interface is as follows: The dispatching and command personnel's terminal displays a heat map of energy flow across the entire network and voltage trend curves of key nodes, supporting manual injection of virtual trains to test the effectiveness of dispatching strategies; the maintenance management personnel's terminal presents a list of vehicles to be inspected aggregated by station and their corresponding energy consumption contribution sorting; vehicle technicians receive personalized work instructions via handheld terminals, containing historical vibration spectrum comparison charts of target components and recommended torque parameter tables. The dispatching terminal adopts a multi-screen linkage design, with the main screen displaying a GIS map overlaid with energy flow arrows, and the auxiliary screen displaying voltage-time curves and alarm lists. The virtual train injection function allows users to specify vehicle type, starting position, running path, and driving mode, and the system simulates its impact on grid voltage in real time for contingency plan simulation. The maintenance management terminal provides a filter, supporting queries for vehicles to be inspected by combination of line, station, health score interval, and dissipation level, and exporting an Excel-format scheduling table. The handheld terminal's work instructions are generated using XML templates, dynamically populated with the latest data, and support offline viewing and electronic signature return. All interfaces comply with the WCAG 2.1 accessibility standard, support high-contrast mode and screen reader compatibility, ensuring efficient operation for all types of users.

[0041] The specific implementation of the communication redundancy mechanism is as follows: The main channel relies on the 5G-U communication protocol to achieve high-bandwidth, low-latency transmission, while the backup channel uses the enhanced LTE-M protocol to maintain basic status reporting functions. The 5G-U channel operates in the dedicated 5.8 GHz frequency band, adopts TDD duplex mode, and has an uplink / downlink time slot ratio of 7:3 to ensure the data transmission needs of high-volume uplink sensors. The base station deployment spacing is no more than 1.2 kilometers to ensure continuous coverage within tunnels. The LTE-M channel operates in the 1.8 GHz frequency band with a bandwidth of 1.4 MHz, supports deep penetration, and is suitable for underground platforms and remote areas. Both channels share a SIM card with dual APN configuration, and the terminal device has a built-in dual-mode communication module, supporting seamless switching. The link quality monitoring cycle is 1 second, and the switching decision latency is controlled within 300 milliseconds. The data retransmission mechanism adopts a selective retransmission strategy, requesting only lost data packets to avoid congestion caused by full retransmission. Consistency verification is performed at the application layer, using sequence numbers and timestamps to reconstruct the data order, ensuring that the continuity of business logic is not affected by communication interruptions.

[0042] The specific implementation of the model parameter calibration mechanism is as follows: After each actual maintenance and disassembly inspection report, residual analysis is performed between the measured wear level and the system's previous predictions. If the absolute error exceeds 10% for three consecutive times, the model parameter backpropagation correction procedure is initiated, adjusting the slope factor and intercept term in the aforementioned empirical attenuation curve. The residual analysis module automatically compares the predicted remaining lifespan with the actual replacement mileage to calculate the relative error. When the error exceeds the threshold three times consecutively, the system generates a model correction work order and submits it to the maintenance approval process. After approval, a background batch processing task is initiated, historical data is called to retrain the attenuation model, and the optimal parameter combination is searched using a Bayesian optimization method. Before the new model goes online, it must undergo A / B testing to ensure that the prediction accuracy is improved rather than worsened. All change records are stored in a blockchain evidence storage system to ensure that the model evolution process is traceable and tamper-proof, meeting the control requirements for software modifications in the functional safety standard IEC 61508.

[0043] The system's operation monitoring mechanism is implemented as follows: A central monitoring console is established to display the real-time operating status of each functional module, data throughput, model prediction accuracy, communication link health, and task execution success rate. Monitoring indicators are collected every 10 seconds, and anomaly detection uses a dynamic threshold algorithm with the baseline adaptively adjusted over time. If any module experiences three consecutive heartbeat timeouts or a sudden increase in prediction error of 50%, a Level 1 alarm is triggered, notifying the on-duty engineer for intervention. The system supports remote diagnostics and hot-fix, allowing updates to some algorithm components without system downtime. The log system uses the ELK stack (Elasticsearch, Logstash, Kibana) for centralized management, supporting full-text search and correlation analysis. The security audit module records all user logins, permission changes, and critical operations, retaining the evidence chain for at least three years. Monitoring data is also used for system health self-assessment, generating monthly operation and maintenance reports to guide the development of preventative maintenance plans.

[0044] The data analysis pipeline is implemented as follows: A complete processing chain is established from raw sensor data to advanced feature indicators. Raw data first enters a buffer queue, is sorted by timestamp, and then distributed to parallel processing units. Vibration signals are bandpass filtered (0.5 kHz to 5 kHz) to remove DC offset and high-frequency noise, and then a short-time Fourier transform (STFT) is performed to generate a time-spectrum image, from which the energy proportion of the resonant frequency band is extracted as a fault-sensitive feature. Temperature data uses a moving average method to eliminate instantaneous fluctuations and calculates the maximum temperature rise rate per minute. Acoustic emission signals are decomposed using wavelet packets to extract the energy distribution of different frequency bands and identify abnormal discharge modes. All features are standardized and stored in the time-series database InfluxDB, retaining the original resolution for 30 days and the aggregated data for 5 years. The analysis pipeline supports dynamically loading new feature algorithm plugins, expanding monitoring dimensions without restarting the service to adapt to the needs of new sensor access.

[0045] The predictive maintenance trigger mechanism is implemented as follows: When a component's health score falls below the set degradation threshold of 0.65 for five consecutive times, the system automatically marks the device as a deterioration tracking object, initiates a long-term trend analysis program, and sends a preliminary alarm to the maintenance task planner. The trend analysis uses a Holt linear trend model to fit the health score sequence and predict the score trend over the next 30 days. If the predicted value will fall below 0.5 within 15 days, it is upgraded to a Level 2 warning and included in the next week's maintenance plan; if it will fall below 0.4 within 7 days, it is upgraded to a Level 1 emergency warning, and immediate repair is arranged. The warning level determines the response time and resource allocation priority. The system also generates a root cause analysis report, listing the top three factors that may cause a decline in the score (such as increased vibration and shock, abnormal temperature rise, and active acoustic emission), assisting technicians in quickly locating the problem. All warning events enter a closed-loop management process until the problem is resolved and the score is confirmed to have returned to a safe range before being closed.

[0046] The specific implementation of the energy matching optimization mechanism is as follows: A rolling time-domain optimization algorithm is used to achieve spatiotemporal coordinated scheduling among train groups. Under the premise of meeting operational safety intervals, the interval running time of subsequent trains is actively fine-tuned to ensure that their traction acceleration phase precisely overlaps with the braking period of the preceding train in both time and space. During the optimization process, differences in train load are considered; heavily loaded trains have a higher traction power requirement, so the matching weight is set to 1.5 times; for empty trains, it is set to 0.8 times. The time overlap window is defined as the intersection of the time period from the start of electric braking to the end of braking of the preceding train and the time period from the start of traction to reaching the target speed of the following train. Spatial overlap requires that the two trains be located in the same power supply zone or adjacent zones and that the connecting switch is closed. The optimization objective includes not only minimizing energy consumption but also passenger comfort indicators; the rate of change of acceleration must not exceed 0.75 m / s². The final generated speed curve must be smooth and continuous to avoid abrupt changes that could cause vehicle jerking. The issued commands include a speed-time matrix and a recommended driving mode, which are parsed and executed by the automatic driving system.

[0047] The specific implementation of the equipment profiling mechanism is as follows: A full lifecycle digital profile is established for each critical piece of equipment, recording its manufacturing serial number, manufacturing date, first-time online date, maintenance records, replacement parts list, cumulative operating mileage, total braking energy tolerance, average deceleration level, environmental exposure history (high temperature, humidity, dust level), and major events (emergency braking, lightning strike, water immersion), etc. The profile is stored in JSON-LD format, supporting semantic queries and relational reasoning. The system regularly generates equipment health trend reports, compares the performance of similar equipment groups, identifies individuals whose performance significantly deviates from the average, and prompts for key attention. The profile data is also used for spare parts demand forecasting. Based on equipment aging curves and failure rate models, a spare parts procurement recommendation list is generated 6 months in advance to reduce inventory costs and stockout risks.

[0048] This embodiment constructs a complete technological closed loop with sensing, analysis, decision-making, execution, and learning capabilities through the deep integration and collaborative operation of the aforementioned subsystems. The system not only solves single-dimensional energy efficiency or maintenance problems but also achieves globally optimal resource allocation through cross-domain data fusion and joint optimization models. Its technical architecture is highly replicable and can be extended to other urban rail transit networks, as well as adapted to similar modes of transportation such as suburban railways and trams, providing core technological support for the intelligent transformation of modern public transportation systems.

Claims

1. A subway intelligent maintenance system with optimized braking energy consumption, characterized in that, include: The train operation status perception module is used to collect real-time operation status data of all online trains in the network, and dynamically reconstruct the dynamic topology model of the train group based on the operation status data and the operation plan. The regenerative braking energy assessment module is used to collect the regenerative braking electrical parameters of each train and, in conjunction with the dynamic topology model of the train group and the traction network voltage data, assess the energy absorbability of each braking event to generate a high dissipation risk warning signal. The component health status monitoring module is used to collect vibration, temperature and acoustic emission signals through a multi-source sensor network deployed on key transmission components, and generate a real-time health score based on the signals; The dynamic maintenance task planning module is used to receive the high dissipation risk warning signal, the real-time health score, and the life prediction result output by the energy consumption-health coupling analysis engine, and execute the four-quadrant decision matrix model to generate a graded maintenance intervention strategy; wherein, the horizontal axis of the four-quadrant decision matrix model is the component health score, and the vertical axis is the daily average braking energy dissipation rate of the power supply section. The energy scheduling and coordination module is used to fine-tune the interval running time of relevant trains according to the scheduling instructions in the graded maintenance intervention strategy, so as to achieve the precise spatiotemporal overlap between the braking period of the preceding train and the traction period of the following train. The adaptive learning hub is used to calibrate and update the prediction model parameters of the energy consumption-health coupling analysis engine based on the actual disassembly and inspection report and environmental factor data after maintenance.

2. The intelligent subway maintenance system with coordinated optimization of braking energy consumption as described in claim 1, characterized in that, The operating condition data includes: global positioning information, real-time velocity vector, rate of change of acceleration, traction output command, electric braking engagement ratio, and air braking trigger status.

3. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 2, characterized in that, The regenerative braking electrical parameters include: the energy amplitude fed back to the traction network, the instantaneous peak power, and the total energy integral.

4. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 3, characterized in that, The multi-source sensor network includes: a triaxial vibration accelerometer, a non-contact infrared thermal imager array, and a broadband acoustic emission probe.

5. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 4, characterized in that, The energy consumption-health coupling analysis engine is used to extract data sample sets of emergency braking frequency and corresponding key component replacement cycles of the same vehicle model under similar load conditions from historical databases, and to establish a quantitative decay relationship model between braking intensity and remaining component life through nonlinear fitting.

6. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 5, characterized in that, The four-quadrant decision matrix model is used to mark trains that fall into the quadrant of low health score and high dissipation rate as special maintenance objects, and to generate emergency repair instructions and simultaneously adjust the braking distribution ratio in their operation diagram.

7. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 6, characterized in that, The rolling time-domain optimization algorithm constructs a state vector set containing multiple trains with a future preset time window as the prediction range. Its objective function is to minimize the total resistance energy consumption of the entire network and maximize the operating condition matching degree, and is constrained by the minimum headway, the maximum allowable delay time and the energy consumption limit of the vehicle itself.

8. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 7, characterized in that, The adaptive learning center is used to perform residual analysis between the measured wear degree of the component and the estimated remaining life of the system, and when the error exceeds the preset threshold multiple times in a row, it starts the model parameter correction program to update the slope factor and intercept term of the quantized decay relationship model.

9. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 8, characterized in that, The system also includes a human-computer interaction interface module, which provides differentiated views for users with different functions. The dispatch and command personnel terminal displays the energy flow heat map of the entire network and the voltage trend curve of key nodes, while the maintenance management personnel terminal displays the list of vehicles to be inspected aggregated by site and their corresponding energy consumption contribution sorted.

10. The intelligent subway maintenance system with coordinated optimization of braking energy consumption according to claim 9, characterized in that, The system also includes a communication redundancy module, which is used to ensure the reliability of data transmission through a dual-channel architecture of main channel and backup channel, and automatically switch to backup channel when the quality of main channel is lower than a preset threshold, and perform data retransmission and consistency verification after recovery.

Images