Rail transit comprehensive hub-oriented cloud-edge terminal collaborative intelligent operation and maintenance system and method

Abstract

The application discloses a cloud-edge-terminal collaborative intelligent operation and maintenance system and method for rail transit comprehensive hub, comprising a collaborative scheduling module, which is in communication connection with an end-side perception layer, an edge computing layer and a cloud platform layer respectively, is used for collecting multi-dimensional state information, and adopts a D3QN scheduling algorithm to dynamically decide to distribute operation and maintenance tasks to the end side, the edge or the cloud for execution. The application adopts a cloud-edge-terminal collaborative scheduling architecture, dynamically distributes tasks based on a reinforcement learning algorithm, makes real-time tasks preferentially processed by edge nodes, computation-intensive tasks processed by the cloud, and simple rule judgment executed by the end side locally, and realizes efficient utilization of computing resources and optimization of task response time delay.

Description

Technical Field

[0001] This invention belongs to the technical field of the Internet of Things, and in particular, it is a cloud-edge-device collaborative intelligent operation and maintenance system and method for rail transit integrated hubs. Background Technology

[0002] With the acceleration of urbanization and the increasing complexity of transportation networks, integrated transportation hubs, as crucial nodes in urban transportation networks, are constantly expanding in scale, rapidly increasing in the number of equipment, and continuously improving in system complexity. Modern integrated transportation hubs, exemplified by Xiong'an Station, integrate multiple modes of transportation such as high-speed rail, subway, buses, and taxis, handling hundreds of thousands of passengers daily. They involve tens of thousands of sets of equipment, including elevators, air conditioning, lighting, security, and fire protection systems, requiring the coordinated operation of multiple subsystems such as BAS, FAS, and PSCADA. However, the existing operation and maintenance management methods for integrated transportation hubs have the following prominent problems:

[0003] 1. Severe system silos hinder data integration. Each specialized subsystem typically operates independently, using different communication protocols. For example, BAS commonly uses ModBus and BACnet, FAS employs proprietary protocols, and PSCADA adheres to IEC61850. This results in direct data exchange between systems. When FAS detects a fire alarm, it cannot automatically trigger BAS to shut down air conditioning and lighting in the affected area, nor can it notify PSCADA to switch emergency power supplies, requiring manual intervention and delaying emergency response.

[0004] 2. Delayed response mechanism and lack of predictive maintenance. The current operation and maintenance model is mainly based on passive response, that is, the traditional process of "fault occurrence - alarm - manual handling". Equipment failures are often only discovered after a sudden shutdown. There is a lack of long-term data accumulation and trend analysis capabilities for equipment health status, which makes it impossible to realize the transformation from "passive maintenance" to "predictive maintenance", increasing operation and maintenance costs and operational safety risks.

[0005] 3. Wasted computing resources and idle edge computing power. Existing operational architectures mostly adopt a "full cloud" model, transmitting all raw data collected by edge devices to the cloud for centralized processing. On the one hand, the continuous uploading of massive amounts of data consumes a large amount of network bandwidth, and the high load on the cloud leads to increased response latency; on the other hand, the local computing power of edge devices is idle and not fully utilized, creating an inefficient situation of "cloud overload and edge idleness."

[0006] To address the aforementioned issues, some patents have proposed improvement solutions. For example, CN118968744A proposes a digital twin-based method for integrated airport traffic management, achieving digital modeling and visual monitoring. However, it still employs a traditional centralized cloud processing architecture and does not address dynamic task offloading and edge intelligent decision-making mechanisms. CN115018148A proposes a digital twin-based method for predicting passenger flow distribution in urban rail transit networks, but it focuses on a single scenario of passenger flow prediction and does not cover the collaborative needs of multiple subsystems in integrated transportation hubs.

[0007] Therefore, there is an urgent need to design a cloud-edge-device collaborative intelligent operation and maintenance system and method for rail transit hubs, so as to realize data fusion of various subsystems, intelligent fault prediction, efficient utilization of edge resources and emergency linkage and collaboration, so as to improve the operation and maintenance efficiency, safety and economy of transportation hubs. Summary of the Invention

[0008] In response to the problems raised above, this invention proposes a cloud-edge-device collaborative intelligent operation and maintenance system and method for integrated rail transit hubs, the specific solution of which is as follows:

[0009] A cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs includes a collaborative scheduling module, which communicates with the edge perception layer, edge computing layer, and cloud platform layer. This module collects multi-dimensional status information and uses the D3QN scheduling algorithm to dynamically allocate operation and maintenance tasks to the edge, device, or cloud for execution.

[0010] The edge sensing layer is used to collect real-time operating data of the device and perform local judgment and alarm based on preset rules;

[0011] The edge computing layer includes edge computing nodes deployed in multiple areas of the hub. Each node is equipped with a parallel computing unit, which is used to run a lightweight AI model to perform real-time analysis on the real-time operating data to detect abnormal events and output device health scores. It is also used to immediately execute the first level of safety handling actions when an abnormal event is detected that meets preset emergency conditions, and simultaneously upload the abnormal event information to the cloud platform layer.

[0012] The cloud platform layer is used to perform global data storage, complex AI model training, construct a BIM+GIS integrated digital twin, and review the received abnormal event information based on the digital twin. Based on the review results, it decides whether to perform second-level handling and inputs the review results and handling records as feedback data into the reward function of the reinforcement learning scheduling model and the AI ​​model training set of the cloud platform layer.

[0013] Preferably, the D3QN scheduling algorithm incorporates latency constraints, edge node load constraints, network bandwidth constraints, and power consumption constraints as penalty terms into the objective function through the Lagrange relaxation method;

[0014] The reward function of the D3QN scheduling algorithm is constructed as R = α × normalized (1 / processing time) + β × task execution success rate - γ × data transmission volume - δ × energy consumption, where α, β, γ, and δ are weight coefficients and satisfy the normalization constraint.

[0015] Preferably, the lightweight AI model includes a multimodal health assessment model. The multimodal health assessment model adopts an LSTM-Transformer hybrid architecture. It extracts local temporal features of time-series data through 1D-CNN, extracts frequency domain features after FFT transformation, performs temporal modeling through LSTM network, captures global dependencies through the self-attention mechanism of Transformer branch, and outputs device health score and failure probability after fusing the three-modal information of time-series data, state data and discrete event data.

[0016] Preferably, the edge computing layer and the cloud platform layer are connected via URLLC slicing communication through a 5G network. When the network is interrupted, the edge computing node independently maintains local security processing functions and caches data. After the network is restored, it is automatically synchronized to the cloud platform layer.

[0017] Preferably, the edge-side perception layer is also used to locally cache the real-time running data collected when the network is interrupted, and automatically upload the cached data to the edge computing layer after the network is restored, and then synchronize it to the cloud platform layer by the edge computing layer; the cloud platform layer is configured with a message queue middleware, which realizes MQTT protocol access based on a distributed message queue cluster, and the cluster aggregate throughput is greater than 100,000 messages / second.

[0018] Another objective of this invention is to provide a cloud-edge-device collaborative intelligent operation and maintenance method for integrated rail transit hubs. The method, applying the aforementioned cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs, includes the following steps:

[0019] S1: Construct a BIM+GIS integrated digital foundation for transportation hubs, binding physical equipment with virtual objects in digital twins to achieve mapping between physical and digital spaces;

[0020] S2: The LSTM-Transformer hybrid prediction model is used to process the equipment operation data and output the equipment health score;

[0021] S3: The collaborative scheduling module collects multi-dimensional status information and uses the D3QN scheduling algorithm to dynamically decide whether to allocate operation and maintenance tasks to the edge, the cloud, or the terminal side for execution. Among them, tasks with real-time requirements higher than the preset threshold are preferentially allocated to edge computing nodes for execution, computationally intensive tasks are allocated to the cloud platform layer for execution, and simple rule-based judgment tasks are executed locally by the terminal side perception layer.

[0022] S4: When the device health score is lower than the preset threshold, the corresponding level of handling is performed according to the graded response mechanism; the edge computing node immediately performs local rapid handling actions, and at the same time uploads the alarm information to the cloud platform layer for review through the high reliability low latency communication channel, which includes URLLC slices of 5G network;

[0023] S5: Feed back the operation and maintenance results to the digital twin to drive the online iterative optimization of the multimodal health assessment model.

[0024] Preferably, in step S2, an LSTM-Transformer hybrid prediction model is used to perform deep inference on the device operation data at a preset period to generate a device health score. The hybrid prediction model extracts the time-domain and frequency-domain features of the time-series data through 1D-CNN, performs time-series modeling through an LSTM network, captures global dependencies through the self-attention mechanism of the Transformer branch, and outputs the device health score and fault probability after fusing the three-modal information of time-series data, state data, and discrete event data. Furthermore, a sliding window mechanism is used to continuously monitor the real-time operation data stream to perform lightweight anomaly detection and achieve millisecond-level alarm triggering, forming a two-layer detection architecture of lightweight real-time monitoring and deep periodic evaluation.

[0025] Preferably, in step S3, the dynamic decision rules of the D3QN scheduling algorithm include: video intrusion detection and fire alarm linkage tasks are preferentially assigned to edge computing nodes for execution; lightweight model training and optimization tasks are assigned to the cloud platform layer for execution; and temperature threshold alarm and switch status detection tasks are executed locally by the edge perception layer.

[0026] Preferably, in step S4, the graded response mechanism is divided into four levels based on the health score range:

[0027] When the health score is ≥0.8, it is judged to be in a healthy state, and data is collected and the evaluation is updated by the edge sensing layer according to the regular cycle;

[0028] When the health score is between 0.5 and 0.8, it is considered a sub-healthy state, and the data collection frequency is increased and early warning information is pushed.

[0029] When the health score is 0.3 or less and less than 0.5, it is judged as a high-risk state, triggering a dual-channel mechanism of rapid local processing at the edge computing node and cloud review.

[0030] When the health score is <0.3, it is considered a critical state. The edge computing node immediately performs a safety shutdown or isolation operation, and simultaneously triggers the highest-level alarm to the cloud platform layer and the emergency command platform. The above-mentioned rapid local handling under high-risk and critical states is the first-level security handling, and the review and confirmation at the cloud platform layer is the second-level handling.

[0031] Preferably, in step S5, the online iterative optimization includes: triggering the retraining of the multimodal health assessment model when the cumulative amount of feedback data reaches a preset batch update threshold; and triggering the parameter update of the D3QN scheduling algorithm when the deviation between the actual value and the predicted value of the reward function of the D3QN scheduling algorithm exceeds a preset deviation threshold.

[0032] The beneficial effects of this invention are:

[0033] 1. Adopting a cloud-edge-device collaborative scheduling architecture, tasks are dynamically allocated based on reinforcement learning algorithms, so that real-time tasks are prioritized for processing by edge nodes, computationally intensive tasks are processed by the cloud, and simple rule judgments are executed locally on the device side, thereby achieving efficient utilization of computing resources and optimization of task response latency.

[0034] 2. Establish a dual-channel mechanism of local handling and cloud verification. Edge nodes perform immediate local security handling of abnormal events, while the cloud performs verification and confirmation based on digital twins, balancing the real-time nature of rapid response with the accuracy of complex decisions.

[0035] 3. Supports multi-source heterogeneous data fusion access, has self-maintaining capability during network outages, and automatically synchronizes data after network recovery, forming a complete intelligent operation and maintenance closed loop of "perception-analysis-decision-execution-feedback". Attached Figure Description

[0036] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that these drawings are designed for illustrative purposes only and are not intended to limit the scope of the present invention. Furthermore, unless specifically indicated, these drawings are intended only to conceptually illustrate the structural construction described herein and are not necessarily drawn to scale.

[0037] Figure 1 This is a system architecture diagram in an embodiment of the present invention;

[0038] Figure 2 This is a flowchart of the cloud-edge collaborative scheduling decision-making process in an embodiment of the present invention;

[0039] Figure 3This is a flowchart illustrating the implementation of predictive maintenance in an application example of the present invention. Detailed Implementation

[0040] First, it should be noted that the specific structure, features, and advantages of the present invention will be described in detail below by way of examples. However, all descriptions are for illustrative purposes only and should not be construed as limiting the present invention in any way. Furthermore, any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the accompanying drawings, can still be arbitrarily combined or deleted among these technical features to obtain more other embodiments of the present invention that may not be directly mentioned herein. Additionally, for the sake of simplifying the drawings, the same or similar technical features may be indicated only in one place in the same drawing.

[0041] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0042] The following is in conjunction with the appendix Figure 1 -Appendix Figure 3 This invention will be described in detail.

[0043] Example 1:

[0044] like Figure 1 As shown, a cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs includes a collaborative scheduling module, which is communicatively connected to the edge perception layer, edge computing layer, and cloud platform layer. This module collects multi-dimensional status information and uses the D3QN scheduling algorithm to dynamically allocate operation and maintenance tasks to the edge, device, or cloud for execution.

[0045] The edge sensing layer is used to collect real-time operating data of the device and perform local judgment and alarm based on preset rules;

[0046] The edge computing layer includes edge computing nodes deployed in multiple areas of the hub. Each node is equipped with a parallel computing unit, which is used to run a lightweight AI model to perform real-time analysis on the real-time operating data to detect abnormal events and output device health scores. It is also used to immediately execute the first level of safety handling actions when an abnormal event is detected that meets preset emergency conditions, and simultaneously upload the abnormal event information to the cloud platform layer.

[0047] The cloud platform layer is used to perform global data storage, complex AI model training, construct a BIM+GIS integrated digital twin, and review the received abnormal event information based on the digital twin. Based on the review results, it decides whether to perform second-level handling and inputs the review results and handling records as feedback data into the reward function of the reinforcement learning scheduling model and the AI ​​model training set of the cloud platform layer.

[0048] Working principle:

[0049] like Figure 1 , Figure 2 As shown, this embodiment provides a cloud-edge-device collaborative intelligent operation and maintenance system for rail transit integrated hubs. This system is applied to large-scale integrated transportation hubs (such as Xiong'an Station), which integrates multiple modes of transportation such as high-speed rail, subway, bus, and taxi, and involves tens of thousands of sets of various equipment such as elevators, air conditioning, lighting, security, and fire protection.

[0050] The core of the system lies in the collaborative scheduling module, which communicates with the edge perception layer, the edge computing layer, and the cloud platform layer to achieve three-layer collaborative scheduling.

[0051] The collaborative scheduling module makes decisions based on multi-dimensional state information, including task characteristics, edge node status, cloud resource status, and network status. The module employs a constraint optimization method based on the D3QN scheduling algorithm. The D3QN algorithm separates the state value function and action advantage function through a Dueling architecture and combines a Double Q-learning mechanism to reduce Q-value overestimation bias. Under the conditions of meeting latency, computational load, bandwidth, and power consumption constraints, it outputs the optimal task allocation decision, which includes assigning tasks to the edge perception layer, edge computing layer, or cloud platform layer for processing.

[0052] Generally, the task characteristics of the multi-dimensional status information include indicators such as task type, urgency, data volume, and computing requirements; edge node status includes indicators such as CPU load, parallel computing unit load, and memory usage; and cloud resource status includes indicators such as the number of available parallel computing units, task queue length, and network status. Finally, based on the above multi-dimensional status information, the D3QN scheduling algorithm is used to dynamically decide whether to allocate maintenance tasks to the edge, the device, or the cloud for execution.

[0053] The edge perception layer is deployed in various areas and on various devices in the transportation hub. It is responsible for collecting various types of data and performing simple rule judgments and local alarms.

[0054] For example, in terms of equipment sensors, vibration sensors, temperature sensors, and current sensors are deployed in elevator cars; air conditioning units are equipped with temperature, humidity, pressure, and flow sensors; lighting systems are equipped with illuminance sensors and infrared occupancy sensors; and fire smoke detectors and temperature detectors cover the entire station. In terms of video sensing, high-definition cameras are deployed, supporting H.265 encoding and featuring built-in AI inference modules, covering key areas such as waiting areas, passageways, elevator lobbies, and platforms.

[0055] The edge computing layer consists of multiple edge computing nodes deployed in transportation hubs, each equipped with a dedicated server. These servers are rack-mounted devices equipped with multi-core high-performance processors, large-capacity memory, high-speed storage media, and parallel computing units for AI model inference. The edge computing nodes are interconnected via a high-speed network and connected to edge devices via a 5G private network or gigabit Ethernet.

[0056] The edge computing layer runs lightweight AI models, including a device health assessment model, which employs a hybrid LSTM-Transformer network, a video analytics model, and an anomaly detection model (Isolation Forest).

[0057] The cloud platform layer is deployed in a regional data center, employing a microservice architecture and configuring parallel computing units for AI model training and inference. Computing resources include multiple high-performance computing nodes equipped with multi-core processors, large-capacity memory, and high-speed storage, and parallel computing units for lightweight model training and optimization based on knowledge distillation. The storage system uses distributed storage, supporting large-capacity data storage and replica redundancy. Platform services include data access services, stream processing services, model training and compression services, data management services, and digital twin services.

[0058] For tasks with high real-time requirements, such as video intrusion detection and fire alarm linkage, the task is scheduled to be executed on edge computing nodes; for computationally intensive tasks, such as model training and long-term data analysis, the task is scheduled to be executed in the cloud to make full use of powerful computing power; for simple rule judgment tasks, such as temperature threshold alarms, the alarm is triggered on the edge device.

[0059] By deploying network slicing technology through 5G private networks or in cooperation with operators, we can provide enhanced mobile broadband (eMBB) slices for video transmission, ultra-reliable low-latency communication (uRLLC) slices for control commands, and massive machine-type communication (mMTC) slices for IoT devices.

[0060] The edge computing layer is also used to immediately execute the first-level security response action when an abnormal event is detected that meets the preset emergency conditions, and at the same time upload the abnormal event information to the cloud platform layer.

[0061] The cloud platform layer is also used to review the received abnormal event information based on the digital twin, and decide whether to perform the second-level handling based on the review result, and input the review result and handling record as feedback data into the reward function of the reinforcement learning scheduling model and the AI ​​model training set of the cloud platform layer.

[0062] Furthermore, in the embodiments, the D3QN scheduling algorithm can be considered to incorporate latency constraints, edge node load constraints, network bandwidth constraints, and power consumption constraints as penalty terms into the objective function through the Lagrange relaxation method.

[0063] The reward function of the D3QN scheduling algorithm is constructed as R = α × normalized (1 / processing time) + β × task execution success rate - γ × data transmission volume - δ × energy consumption, where α, β, γ, and δ are weight coefficients and satisfy the normalization constraint.

[0064] In this embodiment, the D3QN scheduling algorithm, namely Dueling Double Deep Q-Network for TaskScheduling, integrates the Dueling Network architecture to separate state values ​​and action advantage functions, and the Double Q-learning mechanism to solve the Q-value overestimation bias.

[0065] The D3QN scheduling algorithm incorporates latency constraints, edge node load constraints, network bandwidth constraints, and power consumption constraints as penalty terms into the objective function using the Lagrange relaxation method. The reward function is constructed as follows:

[0066] R = α × Normalized(1 / processing time) + β × task execution success rate - γ × data transmission volume - δ × energy consumption;

[0067] Where α, β, γ, and δ are weighting coefficients, satisfying the normalization constraint α+β+γ+δ=1. Each component is normalized to the range [0,1].

[0068] The constraints are incorporated into the objective function using the Lagrange relaxation method:

[0069] L = R - λ1 × max(0, latency - latency limit) - λ2 × max(0, CPU utilization - utilization limit) - λ3 × max(0, bandwidth usage - bandwidth limit) - λ4 × max(0, power consumption - power consumption limit);

[0070] Where λ1~λ4 are penalty coefficients, and each constraint term has been normalized to a dimensionless comparable range before being incorporated into the objective function.

[0071] The following uses real-world application scenarios to further illustrate this:

[0072] In the reward function of the D3QN scheduling algorithm, the weight coefficients α, β, γ, and δ satisfy the normalization constraint α+β+γ+δ=1. The recommended values ​​for each weight coefficient are α=0.4, β=0.3, γ=0.15, and δ=0.15. In actual deployment, these values ​​can be finely adjusted within ±0.05 based on operational data. The penalty coefficients λ1~λ4 all range from [0.5, 2.0].

[0073] The methods and basis for determining the above numerical range are as follows:

[0074] (a) Experimental method: A simulated hub containing 100 heterogeneous computing nodes (covering edge nodes and cloud nodes with different computing power levels) was constructed in the simulation environment. 100,000 historical task scheduling records were generated as training datasets. The grid search method was used to traverse the four-dimensional parameter space of α / β / γ / δ (step size 0.05). The comprehensive weighted score of three indicators, namely average task completion time, success rate and energy consumption, was used as the evaluation function.

[0075] (b) Determining the range of α: When α < 0.3, the strategy is conservative (overly inclined to cloud offloading, leading to network congestion); when α > 0.5, the strategy is too aggressive (overly inclined to edge processing, leading to edge overload). Therefore, [0.3, 0.5] is taken as the convergence and stability interval.

[0076] (c) Determining the range of β—β controls the incentive intensity for success rate. Experiments show that when β∈[0.2,0.4], the task success rate is maintained above 96% and does not conflict with other objectives;

[0077] (d) Determining the range of γ / δ—the two constrain bandwidth and energy consumption respectively. Experiments show that when the range exceeds 0.3, it causes excessive suppression of the main target, and when it is below 0.1, the constraint fails. Therefore, the range is set to [0.1, 0.3].

[0078] (e) The range λ1~λ4—Based on sensitivity analysis, the penalty signal within this range is sufficient to guide the strategy to converge to the feasible region without causing training oscillations. The above range is the preferred range in this embodiment, and can be fine-tuned based on specific operational data during actual deployment.

[0079] The normalization process employs the Min-Max normalization method, mapping each constraint term to the [0,1] interval. The specific formula is as follows:

[0080] x_norm = (x - x_min) / (x_max - x_min),

[0081] Where x represents the original constraint value (such as latency in milliseconds, CPU utilization in percent), and x_min and x_max represent the minimum and maximum values ​​of this constraint in historical running data, respectively. The x_min and x_max values ​​for each constraint are obtained through continuous sliding window statistics collected during system operation. The D3QN scheduling algorithm combines the advantages of Double DQN (which addresses the Q-value overestimation problem) and DuelingNetwork (which separates state values ​​and action advantage functions), and achieves end-to-end task scheduling decisions through experience replay and stable training of the target network.

[0082] The scheduling strategy comprehensively considers four optimization objectives: latency requirements, computational load, energy consumption cost, and network bandwidth. The state space includes task characteristics (task type, urgency, data volume, computational requirements), edge node states (CPU load, parallel computing unit load, memory usage, number of activated head networks), cloud resource states (number of available parallel computing units, task queue length), and network states (bandwidth, latency, packet loss rate). Actions involve selecting the processing level (edge ​​processing, cloud processing, endpoint processing), with fine-grained details including selecting specific edge nodes and model versions.

[0083] Experiments show that the D3QN scheduling algorithm significantly reduces response time and significantly improves success rate and resource utilization compared to traditional heuristic algorithms.

[0084] Furthermore, in the embodiments, the lightweight AI model may include a multimodal health assessment model. The multimodal health assessment model adopts an LSTM-Transformer hybrid architecture, extracts local temporal features of time-series data through 1D-CNN, extracts frequency domain features after FFT transformation, performs temporal modeling through LSTM network, captures global dependencies through the self-attention mechanism of Transformer branch, and outputs device health score and failure probability after fusing the three-modal information of time-series data, state data and discrete event data.

[0085] In this embodiment, the multimodal health assessment model adopts an LSTM-Transformer hybrid architecture.

[0086] By fusing trimodal information from time-series data, state data, and discrete events into a model, and capturing long-term dependencies through a multi-head self-attention mechanism, the model outputs a device health score (0-1) and a failure probability.

[0087] The temporal branch uses 1D-CNN and LSTM to extract temporal and frequency domain features, while the Transformer branch captures long-term dependencies through a multi-head self-attention mechanism. The state branch mainly processes scalar data such as temperature and pressure, using a fully connected network for encoding. The event branch handles discrete events such as start-up, shutdown, and faults, using an attention mechanism to capture temporal correlations. After the features from the four branches are fused, a health score and fault type probability are output through a fully connected layer.

[0088] Furthermore, in the embodiments, the edge computing layer and the cloud platform layer can be connected via URLLC slicing communication through a 5G network. When the network is interrupted, the edge computing node independently maintains local security processing functions and caches data. After the network is restored, it is automatically synchronized to the cloud platform layer.

[0089] In this embodiment, network slicing technology is deployed through a 5G private network or in cooperation with operators to provide enhanced mobile broadband (eMBB) slices for video transmission and ultra-reliable low-latency communication (uRLLC) slices for control commands. The end-device supports local data caching, which can temporarily store data during network interruptions and automatically upload it after network recovery.

[0090] Furthermore, in the embodiments, the edge-side perception layer can also be used to locally cache the real-time running data collected when the network is interrupted, and automatically upload the cached data to the edge computing layer after the network is restored, and then synchronize it to the cloud platform layer by the edge computing layer; the cloud platform layer is configured with a message queue middleware, which realizes MQTT protocol access based on a distributed message queue cluster, and the cluster aggregate throughput is greater than 100,000 messages / second.

[0091] In this embodiment, the edge device supports local data caching, which can temporarily store data when the network is interrupted and automatically upload it after the network is restored. The cloud platform layer is configured with a message queue middleware for data access, and the cluster aggregate throughput is greater than 100,000 messages / second.

[0092] Example 2:

[0093] A cloud-edge-device collaborative intelligent operation and maintenance method for integrated rail transit hubs, applying the aforementioned cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs, includes the following steps:

[0094] S1: Construct a BIM+GIS integrated digital foundation for transportation hubs, binding physical equipment with virtual objects in digital twins to achieve mapping between physical and digital spaces;

[0095] S2: The LSTM-Transformer hybrid prediction model is used to process the equipment operation data and output the equipment health score;

[0096] S3: The collaborative scheduling module collects multi-dimensional status information and uses the D3QN scheduling algorithm to dynamically decide whether to allocate operation and maintenance tasks to the edge, the cloud, or the terminal side for execution. Among them, tasks with real-time requirements higher than the preset threshold are preferentially allocated to edge computing nodes for execution, computationally intensive tasks are allocated to the cloud platform layer for execution, and simple rule-based judgment tasks are executed locally by the terminal side perception layer.

[0097] S4: When the device health score is lower than the preset threshold, the corresponding level of handling is performed according to the graded response mechanism; the edge computing node immediately performs local rapid handling actions, and at the same time uploads the alarm information to the cloud platform layer for review through the high reliability low latency communication channel, which includes URLLC slices of 5G network;

[0098] S5: Feed back the operation and maintenance results to the digital twin to drive the online iterative optimization of the multimodal health assessment model.

[0099] In this embodiment, the method is illustrated using a practical application scenario:

[0100] Step 1: Establish a BIM+GIS integrated digital foundation for the transportation hub. This involves constructing a digital foundation that integrates the BIM 3D model of the transportation hub with the GIS geographic information system. Physical equipment (such as elevators, air conditioning units, and power distribution cabinets) is bound to virtual objects one by one, and unique identification codes are established for each piece of equipment (such as equipment number, location coordinates, and model parameters). The BIM model adopts the IFC open standard format, and the GIS data adopts GeoJSON or Shapefile format. A seven-parameter model transformation is used to convert the BIM local coordinate system to the GIS coordinate system.

[0101] Step two: Define the equipment health index system. Construct an equipment health assessment model as the core of equipment health assessment. The model integrates three modal information: time-series data, state data, and discrete event data. It captures long-term dependencies through a multi-head self-attention mechanism and outputs an equipment health score (0-1) and failure probability.

[0102] Step 3: Perform dynamic task scheduling based on the D3QN scheduling algorithm.

[0103] The collaborative scheduling module collects four dimensions of information—task characteristics, edge node status, cloud resource status, and network status—as input to the state space. The algorithm separates the state value function V(s) and action advantage function A(s,a) using a Dueling architecture, and combines a Double Q-learning mechanism to reduce Q-value overestimation bias, outputting the optimal task allocation decision (selecting edge processing, cloud processing, or end-side processing). Furthermore, it incorporates latency, load, bandwidth, and power consumption constraints into the objective function using the Lagrange relaxation method. The D3QN scheduling algorithm exhibits the following allocation patterns in actual operation:

[0104] For tasks with high real-time requirements (such as video intrusion detection and fire alarm linkage), the task is scheduled to be executed on edge computing nodes; for computationally intensive tasks (such as lightweight model training and optimization based on knowledge distillation and long-term data analysis), the task is scheduled to be executed in the cloud to make full use of powerful computing power; for simple rule judgment tasks (such as temperature threshold alarms), the alarm is triggered on the edge device.

[0105] Step four triggers a dual-channel mechanism of local handling and cloud verification. When the equipment health assessment result is lower than the preset threshold, the corresponding level of handling is executed according to the graded response mechanism. When the edge node's equipment health assessment model detects a high-risk or critical anomaly, the local channel immediately executes rapid handling actions (such as escalator emergency stop, fan start / stop, and turnstile locking); simultaneously, alarm information is uploaded to the cloud platform through a highly reliable, low-latency communication channel. The cloud platform performs 3D visualization verification using a digital twin model, and maintenance personnel remotely confirm and execute subsequent handling.

[0106] Step 5: Maintenance results are fed back to the digital twin. After completing the repair, maintenance personnel update the equipment status via mobile terminal. The system automatically records maintenance process data (fault symptoms, replaced parts, repair time) and feeds it back to the digital twin, driving online iterative optimization of the health model and continuously improving prediction accuracy.

[0107] Furthermore, in the embodiments, step S2 can be considered as follows: an LSTM-Transformer hybrid prediction model is used to perform deep inference on the device operation data at a preset period to generate a device health score. The hybrid prediction model extracts the time domain and frequency domain features of the time series data through 1D-CNN, performs time series modeling through LSTM network, captures global dependencies through the self-attention mechanism of the Transformer branch, and outputs the device health score and fault probability after fusing the three-modal information of time series data, state data and discrete event data. In addition, a sliding window mechanism is used to continuously monitor the real-time operation data stream to perform lightweight anomaly detection and achieve millisecond-level alarm triggering, forming a two-layer detection architecture of lightweight real-time monitoring and deep periodic evaluation.

[0108] In this embodiment, the model is a dedicated prediction model for equipment health assessment, with an execution cycle of once every 6 hours. Here, 'every 6 hours' refers to the reporting / updating cycle of health score results (i.e., pushing the latest assessment results to the operation and maintenance terminal every 6 hours for trend analysis and decision-making), rather than the model only being able to run once every 6 hours. In fact, the model continuously receives sensor data streams on edge nodes in a sliding window manner and performs real-time feature extraction and anomaly detection (based on lightweight detection algorithms such as isolated forest, with single-window processing time <10ms). When data is detected to deviate abnormally from the normal pattern, an alarm can be triggered immediately, without relying on the 6-hour cycle. The complete LSTM-Transformer deep assessment model, due to its large computational load (845K parameters, multimodal fusion), performs complete inference to generate health scores at 6-hour intervals. The two form a two-layer detection architecture of 'lightweight real-time monitoring + deep periodic assessment'.

[0109] Furthermore, in the embodiments, in step S4, the graded response mechanism is divided into four levels based on the health score interval:

[0110] When the health score is ≥0.8, it is judged to be in a healthy state, and data is collected and the evaluation is updated by the edge sensing layer according to the regular cycle;

[0111] When the health score is between 0.5 and 0.8, it is considered a sub-healthy state, and the data collection frequency is increased and early warning information is pushed.

[0112] When the health score is 0.3 or less and less than 0.5, it is judged as a high-risk state, triggering a dual-channel mechanism of rapid local processing at the edge computing node and cloud review.

[0113] When the health score is less than 0.3, it is judged as a critical state. The edge computing node immediately performs a safe shutdown or isolation operation and simultaneously triggers the highest level alarm to the cloud platform layer and the emergency command platform.

[0114] In this embodiment, the graded response mechanism is divided into four levels based on the health score range:

[0115] (1) Health status (health level ≥ 0.8): Normal operation, data is collected and updated according to the regular cycle (every 6 hours);

[0116] (2) Sub-healthy state (0.5≤health level<0.8): Increase the data collection frequency to once per hour, mark it as a yellow warning in the digital twin and push the prompt information to the operation and maintenance terminal;

[0117] (3) High-risk status (0.3≤health level<0.5): Initiate local rapid handling at the edge node (such as reducing equipment load, switching to backup components), and at the same time send alarms to the cloud through a high-reliability, low-latency communication channel to trigger on-site inspection instructions from maintenance personnel;

[0118] (4) Critical state (health level <0.3): Immediately perform safe shutdown or isolation operations (such as emergency stop of escalator, reduced load operation of air conditioning unit), and simultaneously trigger the highest level alarm to the operation and maintenance center and emergency command platform, and activate the emergency plan.

[0119] Furthermore, in the embodiments, step S5 may include the following: when the cumulative amount of feedback data reaches a preset batch update threshold, triggering the retraining of the multimodal health assessment model; when the deviation between the actual value and the predicted value of the reward function of the D3QN scheduling algorithm exceeds a preset deviation threshold, triggering the parameter update of the D3QN scheduling algorithm.

[0120] In this embodiment, after maintenance personnel complete the repair, they update the equipment status via a mobile terminal. The system automatically records maintenance process data (fault symptoms, replaced parts, repair time) and feeds it back to the digital twin, driving online iterative optimization of the health model and continuously improving prediction accuracy. After this fault case was added to the training set, the model's sensitivity to identifying temperature anomaly patterns in similar transformers improved by approximately 8%.

[0121] The technical solution of the present invention will be further explained below in conjunction with specific application scenarios.

[0122] Application Example 1: Elevator Fault Prediction and Closed-Loop Control.

[0123] This application example involves the intelligent maintenance of an escalator in a transportation hub.

[0124] The end-side sensing layer deploys vibration sensors and current sensors on the escalator drive motor to collect motor operation data in real time. The vibration sensors collect time-domain signals of the motor's radial and axial vibration acceleration; the current sensors collect the motor's three-phase current values.

[0125] Table 1 Elevator Sensor Configuration Parameters

[0126] Vibration acceleration sensor High-frequency sampling Covering commonly used frequency bands Industrial-grade precision motor housing ModBus Current transformer Intermediate frequency sampling Meet monitoring needs Industrial-grade precision Power input line ModBus Temperature sensor Low frequency sampling Standard work scope Industrial-grade precision bearing housing ModBus

[0127] An edge computing layer-based device health assessment model employs a multimodal fusion architecture. This model integrates three modalities: time-series data (vibration / current, etc., with local temporal features extracted via 1D-CNN, frequency features extracted via FFT transformation, and then input into an LSTM network for time-series modeling), state data (scalar parameters such as temperature / pressure), and discrete event data (start / stop / fault event records). The model outputs a device health score (0-1) and fault probability via an LSTM-Transformer hybrid network. The model's inference time on edge computing nodes is approximately 85ms (P95, INT8 quantization), and the health assessment results are updated every 6 hours and uploaded to the cloud via the MQTT protocol.

[0128] The edge computing layer runs an energy efficiency optimization model, which adopts the D3QN scheduling algorithm (using the same deep reinforcement learning framework as the collaborative scheduling module, and specifically designed the state space and reward function for energy consumption optimization scenarios). Based on real-time environmental parameters and personnel distribution, it dynamically adjusts the air conditioning temperature setpoint, air volume, and lighting brightness.

[0129] Application Example 2: Video Intrusion Detection and Emergency Response.

[0130] This application example focuses on unauthorized intrusion detection during off-peak hours. The edge-side perception layer deploys high-definition cameras at the entrances, passageways, and sensitive areas of the transportation hub. These cameras support infrared night vision and possess high-resolution imaging capabilities. Each camera has a built-in AI inference chip, running a lightweight target detection model that can identify people, vehicles, and moving objects in real time.

[0131] The simplified version of the lightweight detection model (containing only basic object detection capabilities, excluding scene adaptive head network) is deployed on the AI ​​chip (computing power ≤ 5 TOPS) built into the edge camera. The quantized model file size is about 8-12MB (including model weights, configuration information and runtime metadata), and the single-frame inference latency is 15-30ms (input resolution 640×640).

[0132] The edge computing layer runs an intrusion detection algorithm that combines target detection and trajectory tracking. The algorithm flow is as follows: the camera captures image frames, each frame uses a lightweight target detection model to extract the person's location, and consecutive frames use a Kalman filter to track the target's trajectory, determining whether the target has entered a prohibited area (such as an equipment room, control room, etc.). Intrusion determination rules include: the target stays in the prohibited area for more than a preset threshold, or the target attempts to enter the prohibited area multiple times in a short period of time.

[0133] Collaboration mode between edge perception layer and edge computing layer:

[0134] The edge only performs lightweight target detection and preliminary filtering (reporting targets with confidence > 0.5), while the complex trajectory tracking, region determination, and intrusion rule engine runs on the edge computing nodes. This two-level architecture of "edge-side preliminary screening + edge-side fine-tuning" reduces the computing pressure on the edge and also reduces the amount of redundant data uploaded to the edge nodes. After the cloud completes a new round of distillation training periodically (e.g., monthly), the updated model weights are distributed to each edge node through the collaborative scheduling module, and then pushed to the edge camera by the edge node, supporting OTA online updates.

[0135] Table 2 Definition of Video Intrusion Detection Area

[0136] Equipment room A01 high 24 hours high urgent power distribution room A02 high 24 hours high urgent Control Center A03 middle Working hours higher important Security check area A04 Low 24 hours middle generally Platform area A05 Low Operating hours middle generally

[0137] Upon detecting an intrusion, the edge computing node immediately executes the first level of security measures: it controls the flashing of lights in the relevant area as a warning; it plays a smart voice broadcast ("You have entered a restricted area, please leave immediately"); it triggers the storage system to record high-definition video footage; and it notifies nearby security personnel to receive the alarm on their handheld terminals. Simultaneously, the alarm information is uploaded to the cloud platform for verification. Security personnel can view the 3D scene of the intrusion area through a digital twin view and retrieve video footage from multiple cameras to confirm the authenticity of the alarm.

[0138] Table 3 Intrusion Detection Emergency Response Actions

[0139] Lights flashing LED lighting system <1 second Lasting 2 minutes Photosensitive detection Voice broadcast Broadcasting system <2 seconds Play 3 times in a loop Audio detection Video recording NVR storage system <1 second 60-second clip Document retrieval Personnel Notification Security handheld terminal <5 seconds Push notifications Reading receipt Large screen display Monitoring Center Large Screen <3 seconds Continuous display Visual confirmation

[0140] Note: Response latency in the table refers to the time interval from the issuance of the alarm trigger command to the start of the device's action; execution duration refers to the time during which the device continuously executes the action.

[0141] In a laboratory testing environment, the system achieved an intrusion detection accuracy of 97.2% (the test dataset contained approximately 500,000 video frames, covering 6 types of intrusion scenarios, and the test hardware was an NVIDIA Jetson AGX Orin edge server), and the false alarm rate was reduced to less than 8 times per month, effectively ensuring the safe operation of the transportation hub.

[0142] Application Example 3: Hierarchical Response Closed-Loop Implementation.

[0143] This application example demonstrates a closed-loop hierarchical response mechanism using transformer #1 as an example.

[0144] At time T0, the edge computing node completes its routine health assessment, outputting a health score of 0.48 for transformer #1, falling within the range of [0.3, 0.5), thus classifying it as a high-risk state. The system immediately executes the following actions:

[0145] (a) Local rapid processing at edge nodes - automatically reducing the load rate of transformer #1 from 85% to 60% (by scheduling the load adjustment task to the edge computing node through the collaborative scheduling module, and the edge node coordinating with the adjacent transformer #2 to share the load).

[0146] (b) Send high-risk alarms to the cloud through a high-reliability, low-latency communication channel. The alarm content includes the device ID, health value, risk level, and recommended measures.

[0147] (c) The cloud platform marks transformer #1 as red and highlights it in the digital twin, and pushes the inspection work order to the handheld terminal of the maintenance personnel, requiring on-site inspection within 24 hours.

[0148] At time T1 (approximately 30 hours later), due to the continuous high load operation of transformer #1 and the abnormal increase in winding temperature (from 65℃ to 82℃), the LSTM-Transformer model of the edge node was re-evaluated after incorporating the newly added time series data. The health score further decreased to 0.26, falling into the <0.3 range, and was upgraded to a critical state.

[0149] The system immediately triggers the highest level of response:

[0150] (a) Safe shutdown operation - the edge node directly controls the circuit breaker to take transformer #1 out of operation (switch to the backup power circuit without affecting the continuity of power supply to the hub).

[0151] (b) Simultaneously trigger the highest level alarm to the operation and maintenance center screen and emergency command platform, and the alarm sound and light prompts continue until manual confirmation;

[0152] (c) Activate the emergency plan - automatically notify the power repair team to rush to the scene with spare parts, and mark the affected area in the BIM digital twin to generate an emergency power supply plan suggestion.

[0153] At time T2, the maintenance personnel completed the on-site inspection (replacing the aging winding insulation pads) and updated the equipment status to "repaired" via mobile terminal.

[0154] The system records complete maintenance process data (fault symptoms: abnormal increase in winding temperature; replaced component: insulation gasket; repair time: 20 hours), feeds it back to the digital twin, and drives online iterative optimization of the health model. Subsequent tracking shows that after the repair of transformer #1, the health level rebounded to 0.89 (healthy status), and the regular 6-hour assessment cycle was restored.

[0155] Application Example 4: Implementation of Predictive Maintenance.

[0156] This application demonstrates the shift from "routine maintenance" to "predictive maintenance." The system integrates multi-source data: equipment sensor data (vibration, temperature, current, etc.), maintenance record data (failure date, repair details, replaced parts), environmental data (temperature, humidity, dust concentration), and historical failure case data. Data is stored in a cloud-based time-series database, partitioned by device ID, supporting fast querying and batch export.

[0157] Table 4 Predictive Maintenance Data Elements

[0158] Sensor time series data Various sensors 1 second - 5 minutes 1200 Time series database 5 years Device status data PLC / SCADA real time 50 relational database 3 years Maintenance record data Maintenance work order system Event-driven 5 relational database 5 years Environmental data Environmental monitoring system 15 minutes 80 Time series database 3 years Fault Case Data Fault Case Library Manual entry 2 Document Database 10 years

[0159] The cloud platform runs a predictive maintenance intelligent analysis engine. The engine employs a multi-period forecasting framework: short-term forecasting (7-30 days ahead) assesses equipment failure probability, medium-term forecasting (1-6 months ahead) assesses equipment maintenance needs, and long-term forecasting (6-24 months ahead) assesses equipment upgrade plans. The forecasting models include a failure probability forecasting model (survival analysis Cox model), a maintenance time window forecasting model (time series forecasting LSTM model), and a spare parts demand forecasting model (inventory optimization model).

[0160] The system generates maintenance work orders based on the prediction results, such as Figure 3 The system displays equipment information, failure risks, recommended maintenance time windows, required spare parts, and personnel skill requirements. Work orders are pushed to maintenance personnel via a mobile app, supporting closed-loop management of order acceptance, execution, and completion feedback. The system automatically tracks maintenance effectiveness, recording changes in equipment status before and after maintenance for continuous model optimization.

[0161] Application Example 5: Energy consumption optimization and energy efficiency management.

[0162] This application example focuses on energy consumption optimization of air conditioning and lighting systems in transportation hubs, demonstrating the application of the D3QN scheduling algorithm in energy consumption optimization scenarios. It reflects the versatility and scalability of the D3QN algorithm in different business scenarios.

[0163] The edge sensing layer deploys temperature and humidity sensors (accuracy ±0.3℃, ±2%RH), CO2 concentration sensors (0-2000ppm, accuracy ±50ppm), illuminance sensors (0-20000lux, accuracy ±5%), and infrared occupancy sensors (detection distance 0-8 meters) in the waiting area and office area. Sensor data is collected every 5 minutes and transmitted to the edge computing nodes via LonWorks or BACnet protocols.

[0164] Table 5 Energy Consumption Sensor Configuration Parameters

[0165] Temperature and humidity sensor 5 minutes Temperature: -20-60℃, Humidity: 0-100%RH Temperature ±0.3℃, Humidity ±2%RH per 500㎡ BACnet CO2 sensor 5 minutes 0-2000ppm ±50ppm per 200㎡ ModBus Illuminance sensor 5 minutes 0-20000 lux ±5% per 300㎡ ModBus Infrared occupancy sensor real time 0-8 meters Detection accuracy ≥95% per 100㎡ ZigBee

[0166] The edge computing layer runs an energy efficiency optimization model, employing the D3QN scheduling algorithm (using the same deep reinforcement learning framework as the collaborative scheduling module, with a dedicated design for the state space and reward function for energy consumption optimization scenarios: the state space includes environmental parameters (indoor temperature, CO2 concentration, illuminance, personnel density), equipment status (air conditioning operation mode, lighting level, fan speed), and energy consumption data (real-time power, cumulative electricity consumption); the reward function is designed as R = α_energy efficiency × normalized (1 / energy consumption per unit area) + β_comfort × comfort score - γ_equipment loss × number of equipment start-stop cycles, where α_energy efficiency, β_comfort, and γ_equipment loss are weight coefficients and satisfy normalization constraints, with default values ​​of 0.5, 0.3, and 0.2 respectively; the constraints are incorporated into the objective function using the Lagrange relaxation method). Based on real-time environmental parameters and personnel distribution, the model dynamically adjusts the air conditioning temperature setpoint, air volume, and lighting brightness. The D3QN scheduling algorithm shares the same core architecture (Dueling Network + Double Q-learning + experience replay mechanism) as the task scheduling D3QN algorithm described in step three, but it has a customized state space and reward function for energy consumption optimization scenarios, demonstrating the versatility and scalability of the D3QN algorithm in different business scenarios. The model training uses historical running data, covering different seasons, time periods, and passenger flow scenarios, totaling 10 million samples. It converged after 2000 training rounds, achieving an energy efficiency improvement of approximately 20%.

[0167] The cloud platform regularly aggregates energy consumption data from various regions and generates energy consumption reports (statistics such as total energy consumption, energy consumption per unit area, and energy saving rate are compiled by day, week, and month).

[0168] The platform supports energy efficiency simulation, allowing maintenance personnel to adjust air conditioning operation strategy parameters (such as temperature setting and start / stop time). The system uses a digital twin model to predict its energy consumption impact and assist in decision-making.

[0169] Table 6. Comparison of daily energy consumption and energy efficiency of the waiting area air conditioning + lighting subsystem under different temperature settings.

[0170] 18℃ 15800 32500 0 7.2 13430 27625 20℃ 14200 29200 10 8.1 12070 24820 22℃ 12600 25800 20 8.8 10710 21930 24℃ 11000 22500 30 9.3 9350 19125 26℃ 9800 20000 38 8.9 8330 17000

[0171] Note: (1) This table contains the measured / simulated daily energy consumption data of the air conditioning + lighting subsystem in the waiting area. It only counts the power consumption of HVAC and lighting equipment, and does not include the energy consumption of power equipment such as elevators, escalators, PSCADA, and fire protection.

[0172] (1) Test conditions: outdoor ambient temperature 28-32℃ (typical summer workday), passenger flow in the waiting area is at a medium-high density level (about 0.8-1.2 people / ㎡), lighting system is running at full power, and HVAC system is in continuous cooling mode. Under these conditions, the HVAC + lighting load intensity in the waiting area is significantly higher than the weighted average of the whole building. This is because the comprehensive energy consumption estimate of the whole building includes a large number of low-load areas (the platform is semi-open / naturally ventilated for about 40% of the time, the passageway uses intermittent lighting during some periods, and the equipment area only maintains standby power consumption when no one is there, etc.). Therefore, it is normal for the energy consumption per unit area of ​​the waiting area to be higher than the comprehensive average of the whole building.

[0173] (2) The operating costs for holidays and weekdays are calculated at RMB 0.85 / kWh for industrial electricity, which corresponds to the electricity cost for total energy consumption on holidays and total energy consumption on weekdays, respectively;

[0174] (3) The total building area is 475,200 square meters;

[0175] (4) The energy saving rate (38% corresponding to the 26℃ set value) in this table is the local energy saving rate of the DQN dynamic optimization strategy of this invention relative to the baseline fixed strategy (18℃ constant temperature full power) in this sub-region. The calculation of the overall energy saving rate of the entire hub needs to be weighted and summarized by combining the area weight of each sub-region and the proportion of equipment type.

[0176] Through the above methods, this system effectively reduces energy consumption per unit area while meeting comfort constraints, and achieves intelligent management of system energy consumption.

[0177] Based on the above laboratory tests and simulation analysis, this system achieves the following technical results:

[0178] The listed technical effect data are divided into two categories:

[0179] (A) Laboratory measured / simulated values ​​– Performance indicators obtained directly by measurement or simulation under a controlled testing environment (the testing environment is a simulation platform of Xiong'an Station or an integrated transportation hub of the same scale, the hardware configuration includes edge computing nodes NVIDIA Jetson AGX Orin / Ascend Atlas 300 series, cloud training cluster equipped with multi-card parallel computing units, and the software environment is a mainstream deep learning framework), such as response latency, detection accuracy, model inference speed, etc.

[0180] (B) Estimated / Estimated Values ​​– Benefit indicators derived from measured data, accelerated degradation test models, industry benchmark comparisons, or engineering experience formulas, marked with the word "estimated" (such as percentage cost reduction, MTBF improvement, safety accident reduction rate, etc.). These values ​​have a certain degree of uncertainty (generally ±10-20% deviation range). The actual effect varies depending on the specific operating conditions at the deployment site (customer traffic, equipment aging, operation and maintenance management level, etc.) and is only for reference on the trend of technical effect.

[0181] All data in the following paragraphs containing the word "expected" are classified as Category B estimates.

[0182] In terms of improving operational efficiency:

[0183] This system employs differentiated response mechanisms for two types of operational tasks with different characteristics:

[0184] Equipment health trend prediction and preventive maintenance tasks: End-side sensors continuously collect equipment operating data (vibration, temperature, current, etc.), and edge nodes continuously detect anomalies in the real-time data stream using a sliding window mechanism (such as a 5-minute time window). Based on lightweight anomaly detection models (such as the isolated forest algorithm) or simple threshold judgment rules, data characteristics can be analyzed in real time within each sliding window, and abnormal patterns that significantly deviate from the normal range can be detected. The average response time from the first detection of an anomaly to the issuance of a local rapid response command is approximately 5 minutes (including the entire process of data acquisition, feature extraction, model inference, judgment confirmation, and issuance of response commands). This type of task belongs to long-cycle trend monitoring tasks, focusing on early identification of equipment degradation trends rather than real-time event response.

[0185] Real-time security event detection and emergency response tasks: For security-sensitive events such as fire alarms, unauthorized intrusions, and elevator entrapments, the system employs an independent, lightweight real-time detection model (such as a lightweight target detection model running at the edge, with a single-frame inference latency of 15-30ms). Combined with a rule engine, this achieves rapid judgment and alarm push at the millisecond to second level, with an end-to-end event response latency of <3 seconds (approximately 15-30ms for edge model inference, approximately 85-185ms for edge model inference, approximately 0.5-2.5 seconds for judgment and confirmation, and approximately 200-500ms for alarm push to the terminal). The response mechanisms for these two types of tasks are independent and do not conflict with each other, together forming the complete operational capability system of this system. Compared to the traditional manual inspection model, the average fault detection cycle of 2 hours is significantly shortened; the cloud-based comprehensive health assessment report is generated and pushed to the operation and maintenance terminal every 6 hours for long-term trend analysis and preventive maintenance decisions; the fault response time is reduced from an average of 30 minutes to 10 minutes; the operation and maintenance labor cost is expected to decrease by 35%, and unplanned maintenance work is expected to decrease by 65%; based on the accelerated degradation test model, the mean time between failures of the equipment is expected to increase by about 120%.

[0186] Regarding safety performance improvements: laboratory tests show that fire detection and emergency response time is less than 30 seconds, and the success rate of personnel evacuation guidance is high; the accuracy rate of illegal intrusion detection is greater than 97%, and the incidence of safety accidents is expected to decrease by 70%; the reliability of elevator safe operation is improved, and the downtime rate is expected to decrease by 65%.

[0187] Regarding energy efficiency improvement: To verify the energy efficiency optimization effect of this system, a comparative simulation test was conducted under the same external conditions (outdoor temperature 28℃, medium passenger flow).

[0188] (1) Local subsystem test: As shown in Table 6, taking the waiting area air conditioning + lighting subsystem as the test object, under the working condition of 26℃ on a weekday, the DQN dynamic optimization strategy of the present invention achieves a local energy saving rate of 38% (daily energy consumption is reduced from 32,500 kWh to 20,000 kWh) compared with the traditional 18℃ constant temperature fixed strategy, and the annualized energy consumption per unit area is significantly reduced (only the HVAC and lighting power consumption of this subsystem are considered). (2) Overall Energy Saving Estimation of the Hub: Based on the energy saving rate curves of each temperature setting obtained from local tests, combined with the area weight and equipment energy consumption ratio of each functional area of ​​the hub, the overall energy efficiency index of the hub is obtained by weighted summation as follows: Under baseline A (traditional fixed strategy: all areas are kept at a constant temperature of 26℃ and lighting is running at full power), the overall energy consumption per unit area of ​​the building is about 165-185 kWh / m²·year; under the strategy of this invention (DQN global dynamic optimization), the overall energy consumption per unit area of ​​the building is reduced to 120-140 kWh / m²·year, and the overall energy saving rate is about 22-28% (of which the air conditioning system contributes about 60-70%, the lighting system contributes about 10-15%, and the power equipment indirectly contributes about 10-20%, and each area is weighted summated according to the area weight and equipment energy consumption ratio). Based on the median estimate: Annual electricity savings = (175-130) kWh / m²·year × 475,200 m² = 21,384,000 kWh; with the electricity price calculated at an average industrial electricity price of 0.85 yuan / kWh (referring to the price published by the local power grid), the annual electricity cost savings are approximately 18 million yuan.

[0189] Explanation of data sources and calculation logic:

[0190] (1) Table 6 shows the measured / simulated daily energy consumption data of the waiting area air conditioning + lighting subsystem under different temperature settings. The energy saving rate (38%) of this local subsystem is the direct verification result of the DQN algorithm of this invention in this subsystem.

[0191] (2) The comprehensive energy saving estimation of the entire hub is not simply multiplying the values ​​in Table 6 by the total area ratio, but adopts a layered weighted method: First, the energy saving rate function curves of air conditioning + lighting loads at various temperatures are determined according to Table 6. Then, the three types of loads, namely air conditioning, lighting and power equipment, are weighted and summed according to their energy consumption ratio and coverage area weight. Although power equipment (elevators, escalators, etc.) does not directly participate in temperature regulation, its start and stop frequency is affected by passenger flow density and thus indirectly benefits from DQN scheduling.

[0192] (3) The above estimated values ​​for the entire building (165-185→120-140 kWh / m²·year) are reasonable ranges after considering the differences in usage intensity in each area. The actual values ​​may have a deviation of ±15% due to factors such as operating season and fluctuations in passenger flow. (4) The energy-saving contribution of the air conditioning system accounts for about 60-70% (because it is the largest energy-consuming item and DQN controls it most directly), the lighting system contributes about 10-15%, and the power equipment indirectly contributes about 10-20%. All data are estimated values ​​based on laboratory tests and simulation analysis, not long-term on-site operation statistics.

[0193] Regarding system performance indicators: Laboratory tests show that the end-to-end event response latency for real-time tasks (such as video intrusion detection and fire alarm linkage) is <3 seconds (of which the edge-side model inference is about 185ms, the judgment and confirmation is about 0.5-2.5 seconds, and the alarm push to the terminal is about 200-500ms), and the inference latency for offline batch processing tasks (such as equipment health assessment) is ≤85ms (P95); the cloud platform supports high concurrency access (message queue aggregation throughput >100,000 messages / second), and the system availability is 99.95% (based on the following high availability architecture design guarantee: (1) The cloud platform adopts a dual-active cluster deployment, with real-time data synchronization between the primary and backup nodes, and failure Switching time < 30 seconds; (2) The edge computing layer adopts an N+1 redundancy design, and the adjacent nodes automatically take over the task load when a single node fails; (3) The message queue middleware adopts master-slave replication + sentinel mode to ensure that messages are not lost and services are not interrupted; (4) The end-side device supports an offline caching mechanism, which can temporarily store data and execute key security policies locally when the network is interrupted, and automatically synchronize after the network is restored); The total data storage supports large-scale data storage (time series database supports >5 billion data points, and can be expanded to support >100,000 sensor access); The digital twin scene rendering frame rate is ≥30fps (1080p resolution), and supports ≥20 users to access simultaneously.

[0194] In summary, the system successfully achieved collaborative operation of multiple subsystems such as BAS, FAS, and PSCADA, and constructed an integrated cloud-edge-device intelligent operation and maintenance system. It provides an innovative solution for the operation and management of smart transportation hubs and has significant demonstration value and application potential.

[0195] The core innovation of this invention lies in the construction of a complete cloud-edge-device collaborative closed-loop system:

[0196] (1) Data flow closed loop: The raw data collected by the edge perception layer is processed in real time by the lightweight model of the edge computing layer. Key features and abnormal alarms are uploaded to the cloud platform layer for long-term storage and analysis. The model trained and optimized in the cloud is then sent to the edge nodes to achieve continuous iteration.

[0197] (2) Command flow closed loop - The global scheduling decision of the cloud platform layer is sent to the edge computing nodes and end-side execution devices through the 5G private network. The local rapid processing results of the edge nodes are fed back to the cloud to form a review and confirmation mechanism to ensure the reliable execution of control commands;

[0198] (3) Feedback flow closed loop—The on-site handling results completed by maintenance personnel through mobile terminals are entered into the digital twin, driving the online update of the health assessment model, forming a complete closed loop of "perception-analysis-decision-execution-feedback-optimization". The above three-flow interaction makes this system not a simple patchwork of functional modules, but an organic whole with self-learning and self-adaptive capabilities. It can dynamically adjust strategies according to actual operating conditions and realize the transformation of the maintenance mode from passive response to proactive prevention.

[0199] The above embodiments and application examples have provided a detailed description of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs, characterized in that: This includes a collaborative scheduling module, which communicates with the edge perception layer, edge computing layer, and cloud platform layer. It collects multi-dimensional status information and uses the D3QN scheduling algorithm to dynamically allocate maintenance tasks to the edge, edge, or cloud for execution. The edge sensing layer is used to collect real-time operating data of the device and perform local judgment and alarm based on preset rules; The edge computing layer includes edge computing nodes deployed in multiple areas of the hub. Each node is equipped with a parallel computing unit, which is used to run a lightweight AI model to perform real-time analysis on the real-time operating data to detect abnormal events and output device health scores. It is also used to immediately execute the first level of safety handling actions when an abnormal event is detected that meets preset emergency conditions, and simultaneously upload the abnormal event information to the cloud platform layer. The cloud platform layer is used to perform global data storage, complex AI model training, construct a BIM+GIS integrated digital twin, and review the received abnormal event information based on the digital twin. Based on the review results, it decides whether to perform second-level handling and inputs the review results and handling records as feedback data into the reward function of the reinforcement learning scheduling model and the AI ​​model training set of the cloud platform layer.

2. The cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs according to claim 1, characterized in that: The D3QN scheduling algorithm incorporates latency constraints, edge node load constraints, network bandwidth constraints, and power consumption constraints as penalty terms into the objective function through the Lagrange relaxation method. The reward function of the D3QN scheduling algorithm is constructed as R = α × normalized (1 / processing time) + β × task execution success rate - γ × data transmission volume - δ × energy consumption, where α, β, γ, and δ are weight coefficients and satisfy the normalization constraint.

3. The cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs according to claim 1, characterized in that: The lightweight AI model includes a multimodal health assessment model, which adopts an LSTM-Transformer hybrid architecture. It extracts local temporal features from time-series data through 1D-CNN, extracts frequency domain features after FFT transformation, performs temporal modeling through an LSTM network, captures global dependencies through the self-attention mechanism of the Transformer branch, and outputs device health score and failure probability after fusing information from time-series data, state data, and discrete event data.

4. The cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs according to claim 1, characterized in that: The edge computing layer and the cloud platform layer are connected via URLLC slicing communication through a 5G network. When the network is interrupted, the edge computing node independently maintains local security processing functions and caches data. After the network is restored, it is automatically synchronized to the cloud platform layer.

5. The cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs according to claim 1, characterized in that: The edge-side perception layer is also used to locally cache the real-time running data collected when the network is interrupted. After the network is restored, the cached data is automatically uploaded to the edge computing layer, which then synchronizes it to the cloud platform layer. The cloud platform layer is configured with a message queue middleware, which implements MQTT protocol access based on a distributed message queue cluster, with a cluster aggregate throughput of more than 100,000 messages / second.

6. A cloud-edge-device collaborative intelligent operation and maintenance method for integrated rail transit hubs, employing the cloud-edge-device collaborative intelligent operation and maintenance system for integrated rail transit hubs as described in any one of claims 1-5, characterized in that, Includes the following steps: S1: Construct a BIM+GIS integrated digital foundation for transportation hubs, binding physical equipment with virtual objects in digital twins to achieve mapping between physical and digital spaces; S2: The LSTM-Transformer hybrid prediction model is used to process the equipment operation data and output the equipment health score; S3: The collaborative scheduling module collects multi-dimensional status information and uses the D3QN scheduling algorithm to dynamically decide whether to allocate operation and maintenance tasks to the edge, the cloud, or the terminal side for execution. Among them, tasks with real-time requirements higher than the preset threshold are preferentially allocated to edge computing nodes for execution, computationally intensive tasks are allocated to the cloud platform layer for execution, and simple rule-based judgment tasks are executed locally by the terminal side perception layer. S4: When the device health score is lower than the preset threshold, the corresponding level of handling is performed according to the graded response mechanism; the edge computing node immediately performs local rapid handling actions, and at the same time uploads the alarm information to the cloud platform layer for review through the high reliability low latency communication channel, which includes URLLC slices of 5G network; S5: Feed back the operation and maintenance results to the digital twin to drive the online iterative optimization of the multimodal health assessment model.

7. The cloud-edge-device collaborative intelligent operation and maintenance method for integrated rail transit hubs according to claim 6, characterized in that: In step S2, an LSTM-Transformer hybrid prediction model is used to perform deep inference on the device operation data at a preset period to generate a device health score. The hybrid prediction model extracts local temporal features of the time-series data through 1D-CNN, extracts frequency-domain features after FFT transformation, performs temporal modeling through an LSTM network, captures global dependencies through the self-attention mechanism of the Transformer branch, and outputs the device health score and fault probability after fusing the three-modal information of time-series data, state data, and discrete event data. Furthermore, a sliding window mechanism is used to continuously monitor the real-time operation data stream to perform lightweight anomaly detection and achieve millisecond-level alarm triggering, forming a two-layer detection architecture of lightweight real-time monitoring and deep periodic evaluation.

8. The cloud-edge-device collaborative intelligent operation and maintenance method for integrated rail transit hubs according to claim 6, characterized in that: In step S4, the graded response mechanism is divided into four levels based on the health score range: When the health score is ≥0.8, it is judged to be in a healthy state, and data is collected and the evaluation is updated by the edge sensing layer according to the regular cycle; When the health score is between 0.5 and 0.8, it is considered a sub-healthy state, and the data collection frequency is increased and early warning information is pushed. When the health score is 0.3 or less and less than 0.5, it is judged as a high-risk state, triggering a dual-channel mechanism of rapid local processing at the edge computing node and cloud review. When the health score is less than 0.3, it is judged as a critical state. The edge computing node immediately performs a safe shutdown or isolation operation and simultaneously triggers the highest level alarm to the cloud platform layer and the emergency command platform.

9. The cloud-edge-device collaborative intelligent operation and maintenance method for integrated rail transit hubs according to claim 6, characterized in that: In step S5, the online iterative optimization includes: triggering the retraining of the multimodal health assessment model when the cumulative amount of feedback data reaches a preset batch update threshold; and triggering the parameter update of the D3QN scheduling algorithm when the deviation between the actual value and the predicted value of the reward function of the D3QN scheduling algorithm exceeds a preset deviation threshold.

Images