Edge-cloud collaborative approach to decision-making for the entire lifecycle of hydrogen fuel cell vehicles: operation and maintenance.
By adopting an edge-cloud collaborative approach to the full lifecycle operation and maintenance decision-making of hydrogen fuel cell vehicles, the problem of data silos in the operation and maintenance management of hydrogen fuel cell vehicles is solved by utilizing edge computing and cloud decision-making, and real-time, efficient operation and maintenance decision-making and unified scheduling are achieved.
Patent Information
- Application Number
- CN202611486596.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-17
AI Technical Summary
In the operation and maintenance management of hydrogen fuel cell vehicles, there is a serious problem of data silos, which makes it difficult to make overall optimal decisions based on comprehensive and multi-dimensional information, resulting in low efficiency of operation and maintenance decisions.
The edge-cloud collaborative operation and maintenance decision-making method for hydrogen fuel cell vehicles adopts a full life cycle approach. It aggregates multi-source heterogeneous data through an edge computing gateway, performs data preprocessing and feature extraction, combines a priority preemptive task scheduler to perform millisecond-level real-time inference, and makes decisions in the cloud, generating decision result data and sending it to the mobile terminal.
It improves the real-time performance and collaborative efficiency of operation and maintenance decisions for hydrogen-powered vehicles, reduces reliance on manual judgment and operation, increases the speed of safety alarm response and the efficiency of unified dispatching of the entire fleet, and enhances overall effectiveness.
Smart Images

Figure CN123123449A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of hydrogen fuel cell vehicle technology, and in particular to a method, system, computer equipment, and storage medium for full lifecycle operation and maintenance decision-making of hydrogen fuel cell vehicles based on edge-cloud collaboration. Background Technology
[0002] In the field of hydrogen fuel cell vehicle technology, hydrogen fuel cell vehicles, with their advantages of zero emissions, long range, and fast refueling, are becoming an important technological route to replace fossil fuels in the medium and heavy-duty commercial vehicle sector. However, in stark contrast to the rapid deployment of vehicles, the operation and maintenance management technology for hydrogen fuel cell vehicles is severely lagging behind.
[0003] Currently, the industry offers several representative existing technical solutions for the operation and maintenance management of hydrogen fuel cell vehicles: single-function fault monitoring solutions, single-function hydrogen refueling scheduling solutions, and discrete fleet management solutions. In the current state of hydrogen fuel cell vehicle operation and maintenance management, various aspects are stored in isolated systems, creating data silos. This lack of comprehensive, multi-dimensional information for optimal overall decision-making, and the absence of end-to-end automated closed-loop processes, with many steps relying on manual judgment and operation, results in low overall efficiency and poor effectiveness of hydrogen fuel cell vehicle operation and maintenance decisions. Summary of the Invention
[0004] This application provides a method, system, computer equipment, and storage medium for edge-cloud collaborative operation and maintenance decision-making for the entire lifecycle of hydrogen fuel cell vehicles. More specifically, this application provides a method, system, computer equipment, computer storage medium, and computer program product for edge-cloud collaborative operation and maintenance decision-making for the entire lifecycle of hydrogen fuel cell vehicles, thereby improving the overall efficiency of operation and maintenance decision-making for hydrogen fuel cell vehicles and enhancing the effectiveness of overall operation and maintenance decision-making.
[0005] In a first aspect, embodiments of this application provide a method for decision-making on the entire lifecycle of hydrogen-powered vehicles based on edge-cloud collaboration, including: Based on the data perception layer device deployed on hydrogen fuel cell vehicles, multi-source heterogeneous vehicle perception data is acquired, and the vehicle perception data is aggregated to the edge computing gateway at the edge. The vehicle perception data of the edge computing gateway is preprocessed and feature extracted to obtain feature-extracted data; priority information of each operation and maintenance decision task is calculated based on a priority preemptive task scheduler; according to the task execution order represented by the priority information, the feature-extracted data is inferred in millisecond-level real time based on each pre-deployed edge inference model, and the feature data and alarm events of the inference process are summarized as edge inference result data; the edge inference result data is uploaded to the cloud. The edge inference result data uploaded by the edge terminal is received based on the message queue, and a time-series database is constructed based on the edge inference result data of each hydrogen energy vehicle; in response to the triggering event of cloud decision, cloud decision is performed on the fleet operation data of the entire fleet in the time-series database based on the pre-trained cloud decision engines, and cloud decision result data is obtained. The cloud-based decision-making results data are sent to a mobile device, and displayed on the mobile device for vehicle users associated with the hydrogen fuel cell vehicle to make decisions.
[0006] Optionally, in some embodiments of this application, the step of performing millisecond-level real-time inference on the feature-extracted data based on pre-deployed edge inference models, and summarizing the feature data and alarm events of the inference process as edge inference result data, includes: Hydrogen concentration, hydrogen storage cylinder pressure and hydrogen storage cylinder temperature data are extracted from the data after feature extraction as input data for hydrogen storage safety monitoring. Based on the hydrogen storage safety monitoring model, the hydrogen storage safety monitoring input data are inferred to obtain hydrogen storage safety monitoring output data. The voltage data of each section of the fuel cell stack is extracted from the data after feature extraction. The fuel cell stack anomaly detection and lifetime prediction model is used to infer the voltage data of each section of the fuel cell stack to obtain fuel cell stack anomaly detection and lifetime prediction data. Operating condition data is extracted from the data after feature extraction, and energy consumption prediction data is obtained by reasoning based on the energy consumption prediction model. Driving operation data is extracted from the feature-extracted data, and driving behavior recognition data is obtained by reasoning based on the driving behavior recognition model. The edge inference result data is determined by summarizing the hydrogen storage safety monitoring output data, the stack anomaly detection and lifetime prediction data, the energy consumption prediction data, and the driving behavior recognition data.
[0007] Optionally, in some embodiments of this application, the step of reasoning about the hydrogen storage safety monitoring input data based on the hydrogen storage safety monitoring model to obtain hydrogen storage safety monitoring output data includes: The initial hydrogen storage safety monitoring model was constructed based on a fully connected multilayer autoencoder architecture. The training dataset was constructed based on historical operating data of the vehicle under normal operating conditions. Based on the training dataset, with minimizing the reconstruction error as the training objective, the mean squared error is used as the loss function to train the initial hydrogen storage safety monitoring model, resulting in the trained hydrogen storage safety monitoring model. After the model training is completed, the distribution data of the reconstruction error is calculated on the validation set of the training dataset, and a dynamic threshold is set based on the distribution data of the reconstruction error. The real-time reconstruction error is obtained by inferring from the hydrogen storage safety monitoring input data based on the trained hydrogen storage safety monitoring model. Anomaly determination is performed based on the real-time reconstruction error and the dynamic threshold to obtain the anomaly determination result; The hydrogen storage safety monitoring input data, the reconstruction error distribution data, the dynamic threshold, and the anomaly determination result recorded during the reasoning process are used as the feature data. The alarm event is generated based on the anomaly determination result to determine the hydrogen storage safety monitoring output data.
[0008] Optionally, in some embodiments of this application, the step of constructing the training dataset based on historical operating data under normal vehicle conditions includes: Extract historical operating data of the target vehicle model under normal operating conditions covering preset scenarios from the time-series database in the cloud; A three-dimensional input vector, including hydrogen concentration, hydrogen storage cylinder pressure, and hydrogen storage cylinder temperature, is extracted from each time point in the historical operation data. The three-dimensional input vector is standardized in each dimension to obtain standardized data. The standardized data of continuous time series is divided into fixed-length sample sequences using a sliding window method. Each window of the fixed-length sample sequence is used as an independent training sample to obtain the segmented normal sample set. The segmented normal sample set is divided into a training set and a validation set according to a preset ratio to generate the training dataset.
[0009] Optionally, in some embodiments of this application, the step of reasoning about the voltage data of each section of the fuel cell stack based on the fuel cell stack anomaly detection and lifetime prediction model to obtain fuel cell stack anomaly detection and lifetime prediction data includes: An isolated forest model is constructed using the isolated forest algorithm. The voltage data of each section of the fuel cell stack is input into the isolated forest model to obtain the abnormal cluster judgment result. The stack reconstruction error is output through an autoencoder; Based on the abnormal cluster judgment results and the fuel cell reconfiguration error, the overall health deviation of the fuel cell stack is calculated. Update the remaining useful life estimate of the fuel cell stack based on the health deviation; The voltage data of each section of the fuel cell stack, the abnormal cluster judgment result, the fuel cell stack reconstruction error, the health deviation, and the updated estimated remaining lifespan of the fuel cell stack are recorded as the feature data. The alarm event is generated based on the abnormal cluster judgment result, the health deviation, and the updated estimated remaining lifespan of the fuel cell stack to determine the fuel cell stack anomaly detection and lifespan prediction data.
[0010] Optionally, in some embodiments of this application, the step of constructing an isolated forest model using the isolated forest algorithm, inputting the voltage data of each section of the fuel cell stack into the isolated forest model, and obtaining the abnormal cluster judgment result includes: Collect historical segment voltage data of the target type of hydrogen fuel cell stack under normal operating conditions, and construct a training set of feature vectors for normal operating conditions; The initial isolated forest model is trained based on the normal operating condition feature vector training set to obtain the trained isolated forest model. Based on the voltage data of each section of the fuel cell stack, a feature vector with the same dimension as the training set is constructed. The feature vector with the same dimension as the training set is then input into the trained isolation forest model to determine the anomaly score. The anomaly score output by the isolated forest is compared with a preset anomaly determination threshold to determine the anomaly cluster judgment result.
[0011] Optionally, in some embodiments of this application, the step of calculating the overall health deviation of the fuel cell stack based on the abnormal cluster judgment result and the fuel cell stack reconstruction error includes: If the abnormal cluster judgment result indicates that the isolated forest is normal, the health deviation is obtained based on the fuel cell reconfiguration error mapping. If the abnormal cluster judgment result indicates that the isolated forest is abnormal, the abnormal score is used as an amplification factor for the health deviation, and the health deviation is calculated.
[0012] Optionally, in some embodiments of this application, generating the alarm event based on the abnormal cluster judgment result, the health deviation degree, and the updated estimated remaining useful life of the fuel cell stack includes: If the abnormal cluster judgment result indicates that the abnormal score is greater than or equal to the preset abnormal alarm threshold, an abnormal cluster alarm is generated. If the estimated remaining useful life of the fuel cell stack is less than or equal to a preset useful life alarm threshold, a useful life decay alarm will be generated. If the health deviation is greater than or equal to a preset health deviation alarm threshold, a health deviation alarm is generated. The abnormal cluster alarm, the lifetime decay alarm, and the health deviation alarm are defined as the alarm events.
[0013] Optionally, in some embodiments of this application, updating the estimated remaining useful life of the fuel cell stack based on the health deviation includes: A conventional aging model for fuel cell stacks is derived by fitting the factory aging rate curve or the relationship between the cumulative operating time of the fuel cell stack and performance degradation. The basic lifetime is derived based on the conventional aging rate of the conventional aging model of the fuel cell stack. The lifespan correction coefficient is calculated based on the health deviation correction function and the health deviation. The estimated remaining useful life of the fuel cell stack is calculated based on the base lifetime duration and the lifetime correction factor.
[0014] Optionally, in some embodiments of this application, the step of reasoning about the operating condition data based on the energy consumption prediction model to obtain energy consumption prediction data includes: Based on the lightweight converter prediction model, the operating data is predicted to obtain the expected hydrogen consumption rate sequence within a future preset time period. Based on the physical model of the hydrogen fuel cell system, and with the optimal stack operating point and optimal driving strategy as constraints, the theoretical minimum hydrogen consumption curve is obtained by solving. The energy consumption deviation is calculated based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve. The operating condition dimension with the largest deviation from the theoretical minimum hydrogen consumption curve and the corresponding maximum deviation value are determined based on the energy consumption deviation. The operating condition data, the expected hydrogen consumption rate sequence, the theoretical minimum hydrogen consumption curve, the operating condition dimension with the largest deviation, and the maximum deviation value are recorded as the feature data. The alarm event is generated based on the relationship between the energy consumption deviation degree and the preset energy consumption deviation alarm threshold to determine the energy consumption prediction data.
[0015] Optionally, in some embodiments of this application, calculating the energy consumption deviation based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve includes: Based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve, a first integral operation is performed within the future preset time period to obtain the total difference between the predicted hydrogen consumption and the ideal hydrogen consumption. Based on the theoretical minimum hydrogen consumption curve, a second integral operation is performed within the future preset time period to obtain the ideal total hydrogen consumption. The ratio of the total difference to the total ideal hydrogen consumption is taken as the energy consumption deviation.
[0016] Optionally, in some embodiments of this application, the step of reasoning about the driving operation data based on the driving behavior recognition model to obtain driving behavior recognition data includes: A driving operation feature vector is constructed based on the driving operation data; The driving operation feature vector is input into the trained fusion model, and the current driving style classification result and the corresponding driving style confidence score are output. The driving behavior score is obtained by weighting and summing the scores of each item according to their respective weights. Personalized energy-saving suggestions are generated based on the current driving style classification results; Based on the current driving style classification result, the driving style confidence score, the driving behavior score, the personalized energy-saving suggestion, and the driving operation data, the feature data is generated, and the alarm event is generated based on the driving behavior score to determine the driving behavior identification data.
[0017] Optionally, in some embodiments of this application, generating the feature data based on the current driving style classification result, the driving style confidence level, the driving behavior score, the personalized energy-saving suggestion, and the driving operation data includes: Determine the suggestion type and trigger time associated with the personalized energy-saving suggestion; The driving operation data is subjected to cumulative driving behavior statistical analysis to obtain driving behavior statistics. The feature data is generated based on the current driving style classification result, the driving style confidence level, the driving behavior score, the suggestion type, the trigger time, and the driving behavior statistics.
[0018] Optionally, in some embodiments of this application, the cloud-based decision engines, based on pre-training, perform cloud-based decision-making on the fleet operation data in the time-series database to obtain cloud-based decision result data, including: Historical order data, vehicle utilization rate, maintenance cycle, and operating cost structure are extracted from the entire fleet operation data as input data for fleet size optimization; the fleet size optimization engine solves the fleet size optimization input data to obtain the vehicle configuration scheme; The distribution of hydrogen refueling stations, real-time traffic, weather forecasts, and road restrictions are extracted from the entire fleet's operational data as input data for route network planning. The route network planning engine is used to solve the route network planning input data to obtain a route network planning scheme. The current health status of vehicles, hydrogen reserves, hydrogen refueling station queue status, and hydrogen price are extracted from the entire fleet's operational data as input data for hydrogen refueling optimization. The hydrogen refueling optimization engine solves the hydrogen refueling optimization input data to obtain a hydrogen allocation scheme. The fuel cell stack and individual component attenuation trends, group comparison anomaly reports, historical maintenance records, and spare parts inventory status are extracted from the entire fleet's operational data as input data for predictive maintenance decisions. The predictive maintenance decision engine then solves the predictive maintenance decision input data to obtain maintenance work orders. The cloud-based decision-making results data are determined based on the vehicle configuration scheme, the route network planning scheme, the hydrogen refueling allocation scheme, and the maintenance work order.
[0019] Optionally, in some embodiments of this application, the step of solving the fleet size optimization input data based on the fleet size optimization engine to obtain the vehicle configuration scheme includes: With the objective function of minimizing total operating costs and maximizing revenue, and with constraints such as total capacity, vehicle utilization rate, maintenance queue limit, purchase cost, and number of vehicle types, a fleet size optimization model is constructed. The vehicle configuration scheme is obtained by solving the fleet size optimization model using a fleet size optimization genetic algorithm.
[0020] Optionally, in some embodiments of this application, the step of using a fleet size optimization genetic algorithm to solve the fleet size optimization model to obtain the vehicle configuration scheme includes: Using the number of vehicles of each model as the decision variable, and employing real-number vector encoding, an initial population is randomly generated within the constraints. For each individual in the initial population, the objective function is calculated, and the penalty function method is used to multiply the individuals that violate the constraints by a penalty coefficient and include them in the fitness. A predetermined number of individuals are randomly selected from the initial population, and those with the best fitness are selected to enter the mating pool until a new generation of parent parents is generated. Based on the new generation parent set, simulated binary crossover is performed to generate the crossover offspring; Apply a polynomial mutation perturbation to the gene loci of the offspring to generate perturbed offspring; Dynamic parameter tuning is performed at each preset algebra iteration. Repair the crossover offspring and the perturbation offspring that exceed the boundary, and retain the historical best individual; The iteration stops when the maximum number of generations or the convergence condition is reached, and the vehicle configuration scheme is output based on the historical best individual.
[0021] Optionally, in some embodiments of this application, the step of solving the route network planning input data based on the route network planning engine to obtain a route network planning scheme includes: Retrieve externally loaded data; Three-dimensional cost data is calculated for each candidate path based on the externally loaded data; Perform a multi-objective search based on the aforementioned three-dimensional cost data to determine the set of non-dominated paths; Dominant path elimination is performed on the non-dominated path set after the search is completed, and the Pareto optimal path set is output. The target path is determined from the Pareto optimal path set, corresponding to the dispatcher's selection operation or the recommended operation based on operational preferences. Generate the route network planning scheme based on the target path.
[0022] Optionally, in some embodiments of this application, obtaining externally loaded data includes: Road network topology data and hydrogen refueling station distribution data are obtained from static map data that is persisted locally in the cloud. Real-time traffic data is retrieved from a third-party traffic data service interface. Obtain weather forecast data by connecting to the cloud-based meteorological data service interface; Obtain road restriction data from the map service provider's interface; The external loading data is generated based on the road network topology data, the hydrogen refueling station distribution data, the real-time traffic data, the weather forecast data, and the road restriction data.
[0023] Optionally, in some embodiments of this application, the step of solving the hydrogen refueling optimization input data based on the hydrogen refueling optimization engine to obtain a hydrogen allocation scheme includes: Using vehicle station allocation, hydrogen refueling volume, and arrival time as decision variables, an objective function is constructed based on hydrogen refueling cost, detour cost, waiting cost, and unallocated penalty. A mixed integer programming model is constructed with constraints such as hydrogen surplus, station allocation, capacity limits of hydrogen refueling stations for each time period, health status limits, and inventory limits of hydrogen refueling stations. The mixed integer programming model is solved using a mixed integer solution algorithm to obtain the hydrogenation allocation scheme.
[0024] Optionally, in some embodiments of this application, solving the mixed-integer programming model according to the mixed-integer solution algorithm includes: If the number of vehicles to be dispatched or the size of the preset variables exceeds the preset threshold, or if the branch and bound cannot converge within the preset time limit, the solution size is determined as the target size, and heuristic solution steps are executed to solve the problem.
[0025] Optionally, in some embodiments of this application, the extraction of the degradation trend of the fuel cell stack and its components includes: The original sequence of remaining lifespan was extracted from the entire fleet's operational data; The original remaining lifetime sequence is denoised and smoothed using sliding window filtering to obtain a smoothed remaining lifetime trend curve. Based on the smoothed remaining lifetime trend curve, trend inflection point detection is performed to obtain the degradation trend of the fuel cell stack and its components.
[0026] Optionally, in some embodiments of this application, the acquisition of multi-source heterogeneous vehicle perception data based on the data perception layer device deployed on the hydrogen fuel cell vehicle includes: The vehicle's native electronic control data is composed of vehicle control data, battery management data, fuel cell control data, and hydrogen management data obtained from the vehicle's native electronic control system. The sensor data consists of temperature and pressure data of the hydrogen storage tank, hydrogen concentration sensing data, and vehicle camera data acquired by the independently installed sensor groups. Location data is obtained based on the positioning module; The vehicle perception data is obtained based on the vehicle's native electronic control data, the sensor data, and the positioning data.
[0027] Optionally, in some embodiments of this application, the method further includes: The alarm events in the hydrogen storage safety monitoring output data at the edge are displayed on the vehicle display screen. Personalized energy-saving suggestions from driving behavior recognition data at the edge are output in real time via local audio-visual devices.
[0028] Optionally, in some embodiments of this application, the method further includes: A component life prediction model is built in the cloud, and a lightweight deployment is carried out at the edge based on the component life prediction model to obtain a fuel cell stack anomaly detection and life prediction model. An energy consumption baseline model is built in the cloud, and a lightweight deployment is carried out at the edge based on the energy consumption baseline model to obtain an energy consumption prediction model. A driving behavior group profile model is built in the cloud, and a driving behavior recognition model is obtained by lightweight deployment on the edge based on the driving behavior group profile model. A configuration file is obtained in the cloud through a security rule base, and a hydrogen storage safety monitoring model is deployed at the edge based on the configuration file; The stack anomaly detection and lifetime prediction model, the energy consumption prediction model, the driving behavior recognition model, and the hydrogen storage safety monitoring model are used as the edge inference model.
[0029] Optionally, in some embodiments of this application, the method further includes: Obtain the decision execution result obtained by the vehicle user after making a decision based on the cloud decision result data, and send the decision execution result back to the cloud; The vehicle status in the time series database is updated based on the decision execution results, and each of the cloud decision engines is calibrated based on the decision execution results.
[0030] Optionally, in some embodiments of this application, the method further includes: From the edge inference result data, obtain the difficult example inference results where the model confidence is lower than the threshold, and the abnormal working condition identification results where the deviation from the cloud prediction exceeds the threshold; The difficult example reasoning results and the abnormal operating condition identification results are identified as difficult example data in the reasoning process; The difficult example data is uploaded to the cloud to construct the incremental dataset required for the next model training.
[0031] Secondly, embodiments of this application provide a hydrogen fuel cell vehicle lifecycle operation and maintenance decision-making system based on edge-cloud collaboration, which has the function of implementing the hydrogen fuel cell vehicle lifecycle operation and maintenance decision-making method based on edge-cloud collaboration provided in the first aspect above. The function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware.
[0032] In one possible design, the system includes: The vehicle data sensing module is used to acquire multi-source heterogeneous vehicle sensing data based on the data sensing layer device deployed on hydrogen energy vehicles, and to aggregate the vehicle sensing data to the edge computing gateway at the edge. The edge inference module is used to preprocess and extract features from the vehicle perception data of the edge computing gateway to obtain feature-extracted data; calculate the priority information of each operation and maintenance decision task based on the priority preemptive task scheduler; perform millisecond-level real-time inference on the feature-extracted data according to the task execution order represented by the priority information and based on the pre-deployed edge inference models; summarize the feature data and alarm events of the inference process as edge inference result data; and upload the edge inference result data to the cloud. The cloud decision module is used to receive edge inference result data uploaded from the edge terminal based on the message queue, and construct a time series database based on the edge inference result data of each hydrogen energy vehicle; in response to the triggering event of cloud decision, it performs cloud decision on the fleet operation data of the entire fleet in the time series database based on the pre-trained cloud decision engines to obtain cloud decision result data. The mobile collaboration module is used to send the cloud decision result data to the mobile terminal and display the cloud decision result data on the mobile terminal for vehicle users associated with the hydrogen energy vehicle to make decisions.
[0033] In another aspect, this application provides a computer device including at least one connected processor and a memory, wherein the memory is used to store program code, and the processor is used to call the program code in the memory to execute the methods described in the above aspects.
[0034] In another aspect, embodiments of this application provide a computer storage medium including instructions that, when executed on a computer, cause the computer to perform the methods described in the above aspects.
[0035] In another aspect, this application provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods described in the above aspects.
[0036] Compared with traditional technologies, the technical solution of this application improves the real-time performance and collaborative efficiency of hydrogen fuel cell vehicle operation and maintenance decisions, enhances the degree of cross-layer data correlation, reduces the reliance on manual judgment and manual operation costs, improves the system's response speed to safety alarms, and improves the efficiency of unified dispatching of the entire fleet, thereby improving the overall effectiveness of hydrogen fuel cell vehicle operation and maintenance decisions and improving the overall operation and maintenance decision-making effect. Attached Figure Description
[0037] Figure 1 This is a flowchart of one embodiment.
[0038] Figure 2 This is a structural block diagram of a system in one embodiment.
[0039] Figure 3 This is a system architecture diagram of one embodiment.
[0040] Figure 4 This is a data link diagram from one embodiment.
[0041] Figure 5 This is an internal structural diagram of a computer device in one embodiment.
[0042] Figure 6This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation
[0043] The terms "first," "second," etc., used in the embodiments of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. The division of modules appearing in the embodiments of this application is only a logical division. In actual applications, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not performed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interface, and the indirect coupling or communication connection between modules may be electrical or other similar forms. None of these are limited in the embodiments of this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed among multiple circuit modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of this application.
[0044] Figure 1 This is a flowchart illustrating one embodiment, such as... Figure 1 As shown in the embodiments of this application, the method for full lifecycle operation and maintenance decision-making of hydrogen fuel cell vehicles based on edge-cloud collaboration includes: The S1100 acquires multi-source heterogeneous vehicle perception data based on the data perception layer devices deployed on hydrogen fuel cell vehicles, and aggregates the vehicle perception data to the edge computing gateway at the edge.
[0045] Among them, the data perception layer equipment refers to the combination of hardware modules on hydrogen fuel cell vehicles used to collect operational data such as electronic control, sensing, and positioning; multi-source heterogeneity refers to the different data sources, types, sampling frequencies, and interface protocols; and edge computing gateway refers to miniaturized computing nodes deployed on the vehicle side with local computing power.
[0046] In practical applications, after a hydrogen fuel cell vehicle starts, the OWA5X edge computing gateway enters working mode, reading operating parameters from the on-board electronic control unit via the CAN bus at a fixed sampling frequency; simultaneously, it reads the values from the added hydrogen concentration sensor via an independent analog interface. All raw data is cached within the edge gateway after being timestamped uniformly.
[0047] The S1200 performs data preprocessing and feature extraction on the vehicle perception data of the edge computing gateway to obtain feature-extracted data; calculates the priority information of each operation and maintenance decision task based on the priority preemptive task scheduler; performs millisecond-level real-time inference on the feature-extracted data according to the task execution order represented by the priority information, based on the pre-deployed edge inference models, and summarizes the feature data and alarm events of the inference process as edge inference result data; and uploads the edge inference result data to the cloud.
[0048] Feature extraction refers to the process of extracting numerical values or vectors with discriminative significance from preprocessed data; priority preemptive task scheduler refers to the scheduling program that preempts NPU computing power according to preset priorities; operation and maintenance decision tasks refer to specific processing tasks such as monitoring, identification, or prediction that need to be executed by the AI model; millisecond-level real-time inference refers to the process of completing the model's forward computation in milliseconds; edge inference result data refers to the collection of feature data and alarm events produced by the edge AI model inference.
[0049] In practical applications, the vehicle perception data is time-aligned, outlier filtered, and feature extracted; AI models are run for millisecond-level real-time inference; and the processed feature data and alarm events are uploaded to the cloud via 4G / 5G networks.
[0050] S1300 receives edge inference result data uploaded from the edge terminal based on the message queue, and constructs a time-series database based on the edge inference result data of each hydrogen energy vehicle; in response to the triggering event of cloud decision, it makes cloud decisions on the fleet operation data of the entire fleet in the time-series database based on the pre-trained cloud decision engines, and obtains cloud decision result data.
[0051] Among them, message queue refers to message middleware used for asynchronous buffering and reliable delivery of data streams; time series database refers to database organized and stored in chronological order and suitable for time series analysis; fleet operation data refers to the overall dataset composed of data uploaded by all vehicles and aggregated in the cloud; cloud decision engine refers to the algorithm model deployed in the cloud and responsible for solving specific decision problems; cloud decision result data refers to the executable decision conclusions produced by the cloud decision engine.
[0052] In practical applications, the cloud-based data receiving service extracts feature data packets uploaded from the edge devices from the message queue, parses them, and writes them into a time-series database. Upon receiving a trigger event, the corresponding optimization and solution process is initiated.
[0053] The S1400 sends the cloud-based decision-making results data to the mobile device, displaying the cloud-based decision-making results data on the mobile device for vehicle users associated with hydrogen fuel cell vehicles to make decisions.
[0054] Among them, mobile terminal refers to terminals such as mobile phones or tablets used by vehicle users or administrators; vehicle user refers to the driver or fleet manager associated with hydrogen fuel cell vehicles.
[0055] In practical applications, the decision results generated in the cloud are pushed to the corresponding mobile app in real time. When the hydrogen refueling optimization engine outputs a hydrogen refueling suggestion, the driver's app receives a push notification, displaying the recommended hydrogen refueling station, estimated waiting time, and recommended amount of hydrogen to be added. The driver can then start navigation with one click. When the predictive maintenance engine generates a maintenance work order, the management app receives the work order notification. The dispatcher can view the work order details, confirm the maintenance time, and assign maintenance personnel. When the driving behavior recognition model produces a real-time score, the driver's app's driving assistant interface is updated with the score and energy-saving tips.
[0056] Compared to traditional technologies, the technical solution of this application improves the real-time performance and collaborative efficiency of hydrogen fuel cell vehicle operation and maintenance decisions by constructing an end-to-end closed-loop decision-making architecture with edge-cloud collaboration and combining it with a priority-based preemptive task scheduling mechanism. This enhances the cross-layer data correlation, reduces the reliance on manual judgment and operation, improves the system's response speed to safety alarms, and improves the efficiency of unified scheduling across the entire fleet. As a result, it improves the overall effectiveness of hydrogen fuel cell vehicle operation and maintenance decisions and enhances the overall operation and maintenance decision-making results.
[0057] Optionally, in some embodiments of this application, millisecond-level real-time inference is performed on the feature-extracted data based on pre-deployed edge inference models, and the feature data and alarm events of the inference process are summarized as edge inference result data. This includes: extracting hydrogen concentration, hydrogen storage cylinder pressure, and hydrogen storage cylinder temperature data from the feature-extracted data as hydrogen storage safety monitoring input data; inferring the hydrogen storage safety monitoring input data based on the hydrogen storage safety monitoring model to obtain hydrogen storage safety monitoring output data; extracting voltage data of each section of the fuel cell stack from the feature-extracted data; inferring the voltage data of each section of the fuel cell stack based on the fuel cell stack anomaly detection and life prediction model to obtain fuel cell stack anomaly detection and life prediction data; extracting operating condition data from the feature-extracted data; inferring the operating condition data based on the energy consumption prediction model to obtain energy consumption prediction data; extracting driving operation data from the feature-extracted data; inferring the driving operation data based on the driving behavior recognition model to obtain driving behavior recognition data; and summarizing the data based on the hydrogen storage safety monitoring output data, fuel cell stack anomaly detection and life prediction data, energy consumption prediction data, and driving behavior recognition data to determine the edge inference result data.
[0058] Among them, hydrogen concentration refers to the volume percentage of hydrogen molecules in the hydrogen storage system or vehicle cabin; hydrogen storage cylinder pressure refers to the actual pressure value of hydrogen in the high-pressure hydrogen storage cylinder; hydrogen storage cylinder temperature refers to the measured temperature of the surface of the hydrogen storage cylinder; and hydrogen storage safety monitoring model refers to the edge AI model used to identify abnormal states of the hydrogen storage system.
[0059] Among them, the voltage of each section of the fuel cell stack refers to the voltage value across each individual cell in the fuel cell stack; the fuel cell stack anomaly detection and lifetime prediction model refers to the AI model used to diagnose fuel cell stack faults and estimate the remaining lifetime.
[0060] Among them, operating condition data refers to a set of parameters such as speed, load, and gradient that reflect the vehicle's operating status; energy consumption prediction model refers to a lightweight AI model that predicts the hydrogen consumption rate in the future; and energy consumption prediction data refers to the expected hydrogen consumption and deviation results output by the energy consumption prediction model.
[0061] Among them, driving operation data refers to the time sequence data of driver operation signals such as acceleration, braking, and steering; driving behavior recognition model refers to the edge AI model that classifies and scores driving styles; driving behavior recognition data refers to the result data such as driving style classification, scoring, and energy-saving suggestions.
[0062] In practical applications, this embodiment provides four AI inference channels during the inference stage at the edge: The hydrogen storage safety monitoring channel receives hydrogen concentration and hydrogen storage cylinder pressure / temperature data. It calculates the reconstruction error through an autoencoder. If the reconstruction error exceeds a preset dynamic threshold, a local audible and visual alarm is triggered within 10 milliseconds, and a safety alarm event data packet is generated to be uploaded to the cloud.
[0063] The fuel cell stack anomaly detection and lifetime prediction channel receives voltage data from each section of the fuel cell stack, uses the isolated forest algorithm to determine whether the current operating state falls into an anomaly cluster, calculates the overall health deviation of the fuel cell stack through an autoencoder, and updates the estimated remaining lifetime of the fuel cell stack.
[0064] The energy consumption prediction channel receives operating condition data such as current vehicle speed, load, and hydrogen consumption rate, predicts the expected hydrogen consumption curve for the next 5 minutes, and compares it with the ideal hydrogen consumption curve to calculate the energy consumption deviation.
[0065] The driving behavior recognition channel receives driving operation data such as accelerator pedal, brake pedal, and vehicle speed change rate. It uses a CNN-LSTM fusion model to identify the current driving style (aggressive / mild / economical) and combines it with energy consumption deviation to generate a driving behavior score and personalized energy-saving suggestions.
[0066] In this embodiment, by deploying four heterogeneous AI models—hydrogen storage, fuel cell stack, energy consumption, and driving—in parallel on the same edge gateway and scheduling them in a priority-based preemptive manner, the concurrent efficiency of multi-source data fusion inference and the real-time performance of safety alarms are improved. This also mitigates the data fragmentation caused by the independent deployment of each monitoring module and reduces the deployment costs of repeated investment in hardware terminals.
[0067] In a specific embodiment, the specific technical implementation process of the hydrogen storage safety monitoring channel during the inference phase is as follows.
[0068] The hydrogen storage safety monitoring model is based on an autoencoder architecture. Its core principle is to train the model using only sensor data from vehicles operating under normal conditions, enabling it to reconstruct "normal" data patterns. When abnormal data is input, the model cannot effectively reconstruct the data, resulting in a significant increase in reconstruction error, thus identifying the anomaly. The specific calculation process consists of the following five steps: Step 1: Model selection and architecture design.
[0069] Given that the input for hydrogen storage safety monitoring consists of multi-dimensional time-series sensor data (hydrogen concentration, hydrogen storage cylinder pressure, and hydrogen storage cylinder temperature), and that there are strict constraints on computing power and power consumption at the edge, a fully connected multilayer autoencoder (MLP) architecture is chosen. The model consists of three parts: Encoder: Composed of 2-3 fully connected layers, it compresses the input data from a high-dimensional space to a low-dimensional bottleneck representation layer by layer. The number of nodes in the bottleneck layer is much smaller than that in the input layer. Bottleneck layer: The low-dimensional representation layer forces the model to learn the most essential feature patterns in the normal data. Decoder: With a structure symmetrical to the encoder, it consists of 2-3 fully connected layers, reconstructing the low-dimensional representation back to the original input dimension layer by layer. The total number of model parameters is controlled within 3MB (after INT8 quantization) to meet the requirements for lightweight deployment at the edge.
[0070] Step 2: Training dataset construction.
[0071] The training data uses only historical operating data from vehicles under normal conditions and does not include any abnormal or fault samples. The dataset construction process is as follows: (1) Extract the historical records of the vehicle model under normal operating conditions from the cloud time-series database, covering typical operating scenarios such as cold start, idling, constant speed driving, acceleration and deceleration, and hill climbing; (2) Extract a three-dimensional input vector for each time point: x = [hydrogen concentration, hydrogen storage tank pressure, hydrogen storage tank temperature]; (3) Perform Z-Score standardization on each dimension to eliminate dimensional differences: ,in and These are the mean and standard deviation of the dimension on the training set, respectively; (4) Use the sliding window method to divide the continuous time series data into a fixed-length sample sequence (the window size is recommended to be 50~100 consecutive time points, corresponding to 0.5~1 second of data), and each window is used as an independent training sample; (5) Divide the processed normal sample set into a training set and a validation set in a ratio of 8:2.
[0072] Step 3: Model training.
[0073] The training objective is to minimize the reconstruction error, and the mean squared error (MSE) is used as the loss function.
[0074] Where N is the dimension of the input vector ( (corresponding to the three dimensions of hydrogen concentration, pressure, and temperature). The original input value, Reconstruct the output value for the decoder.
[0075] The training process uses the Adam optimizer, with a learning rate set to... The batch size is set to 64~128. The early stopping mechanism monitors the validation set loss and terminates training when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting.
[0076] Step 4: Determine the dynamic threshold.
[0077] After training, the reconstruction error for each sample is calculated on the validation set to obtain the distribution of the reconstruction error. A dynamic threshold is set using the 3σ criterion (mean plus three times the standard deviation).
[0078] in To verify the mean of the reconstruction error of the validation set, This represents the standard deviation of the reconstruction error of the validation set. Choosing a 3σ threshold (corresponding to a 99.7% confidence interval) can control the false alarm rate while ensuring a high detection rate.
[0079] Step 5: Online reasoning and anomaly detection.
[0080] During vehicle operation, the edge gateway collects real-time data on hydrogen concentration, hydrogen storage tank pressure, and temperature. After undergoing the same standardized processing, this data is input into a trained autoencoder model for forward inference, calculating the real-time reconstruction error. :
[0081] The exception detection logic is as follows: If If the condition is determined to be normal, no alarm will be triggered; if If the status is determined to be abnormal, a local audible and visual alarm will be triggered within 10 milliseconds, and a security alarm event data packet will be generated to prepare for uploading to the cloud.
[0082] Special note regarding the model update mechanism: Unlike the other three models, the training data for the hydrogen storage safety monitoring model is only trained offline once during the vehicle factory calibration stage and then fixed and deployed at the edge. The model is not retrained during operation. The reasons are as follows: (1) Safety-related tasks require the model to behave in a highly stable and predictable manner, avoiding the introduction of unknown risks by OTA updates; (2) The physical failure modes of the hydrogen storage system are relatively fixed, so there is no need to update the model frequently.
[0083] If the detection sensitivity needs to be adjusted due to environmental changes (such as significant changes in ambient temperature / barometric pressure caused by vehicle operation across regions), the cloud will issue new threshold parameters through the security rule base configuration method. or standardized parameters Instead of replacing the entire model file.
[0084] Optionally, in some embodiments of this application, the hydrogen storage safety monitoring input data is inferred based on the hydrogen storage safety monitoring model to obtain hydrogen storage safety monitoring output data, including: constructing an initial hydrogen storage safety monitoring model based on a fully connected multilayer autoencoder architecture; constructing a training dataset based on historical operating data under normal vehicle conditions; training the initial hydrogen storage safety monitoring model based on the training dataset, with minimizing the reconstruction error as the training objective and using mean square error as the loss function, to obtain a trained hydrogen storage safety monitoring model; after the model training is completed, calculating the distribution data of the reconstruction error on the validation set of the training dataset, and setting a dynamic threshold based on the distribution data of the reconstruction error; inferring the hydrogen storage safety monitoring input data based on the trained hydrogen storage safety monitoring model to obtain the real-time reconstruction error; performing anomaly judgment based on the real-time reconstruction error and the dynamic threshold to obtain anomaly judgment results; recording the hydrogen storage safety monitoring input data, the distribution data of the reconstruction error, the dynamic threshold, and the anomaly judgment results during the inference process as feature data, and generating alarm events based on the anomaly judgment results to determine the hydrogen storage safety monitoring output data.
[0085] Among them, the fully connected multilayer autoencoder refers to a symmetric encoder-decoder neural network composed of multiple fully connected layers; the initial hydrogen storage safety monitoring model refers to an autoencoder network with untrained parameters and random initialization; normal operating conditions refer to the set of typical operating conditions under which a vehicle operates without faults; historical operating data refers to the time-series records of vehicle sensors that have been collected and stored in the past; and the training dataset refers to the set of labeled or unlabeled samples used for model training and validation.
[0086] Minimizing the reconstruction error refers to the optimization objective of continuously reducing the difference between the input and the reconstructed output during the training process; mean squared error refers to the mean of the sum of squares of the differences between the input and the reconstructed values, and is used as the loss function.
[0087] Here, the validation set refers to the subset of data split from the training set used to evaluate the model's generalization ability; reconstruction error refers to the measure of the difference between the autoencoder input and output; distribution data refers to the statistical distribution information of the reconstruction error values on the validation set; and dynamic threshold... This refers to the anomaly detection boundary that is adaptively set based on the distribution of historical normal data.
[0088] Among them, real-time reconstruction error This refers to the reconstruction error value of the current sample during online inference; the anomaly determination result refers to the judgment conclusion on whether the current input is abnormal.
[0089] In practical applications, the hydrogen storage safety monitoring model is based on an autoencoder architecture. Its core principle is to train the model using only sensor data from vehicles operating under normal conditions, enabling the model to learn to reconstruct "normal" data patterns. When abnormal data is input, the model cannot effectively reconstruct the data, resulting in a significant increase in reconstruction error, thus identifying the anomaly.
[0090] In this embodiment, the unsupervised anomaly detection mechanism combining autoencoder reconstruction error with dynamic threshold improves the detection capability and discrimination accuracy of hydrogen storage safety monitoring for unknown anomaly patterns, improves the false alarm and false negative effects of traditional threshold alarms which are easily affected by operating condition drift, and reduces the cost of anomaly sample labeling and model iteration maintenance.
[0091] Optionally, in some embodiments of this application, a training dataset is constructed based on historical operating data under normal vehicle conditions, including: extracting historical operating data of the target vehicle model covering preset scenarios under normal operating conditions from a time-series database in the cloud; extracting a three-dimensional input vector including hydrogen concentration, hydrogen storage tank pressure, and hydrogen storage tank temperature for each time point in the historical operating data; standardizing each dimension of the three-dimensional input vector to obtain standardized data; using a sliding window method to divide the standardized data of continuous time series into a fixed-length sample sequence, and using each window of the fixed-length sample sequence as an independent training sample to obtain a segmented normal sample set; dividing the segmented normal sample set into a training set and a validation set according to a preset ratio to generate a training dataset.
[0092] Among them, the target vehicle model refers to the specific vehicle model selected according to the design parameters for training data collection; the preset scenario refers to the coverage of typical operating conditions such as cold start, constant speed, and hill climbing; the three-dimensional input vector refers to the column vector composed of three components: hydrogen concentration, cylinder pressure, and cylinder temperature; and the standardization process refers to Z-score normalization of the data in each dimension according to the mean and standard deviation.
[0093] Among them, the sliding window method refers to the sampling technique of dividing time series data along the time axis with a fixed step size; fixed length means that the number of consecutive time points contained in each sample sequence remains consistent; sample sequence refers to the model input unit composed of continuous data points of fixed length; normal sample set refers to the data set composed of sample sequences that have been screened and confirmed to be without anomalies.
[0094] The preset ratio refers to the percentage of the training set and validation set in the normal sample set, such as an 8:2 ratio. The training set is the subset of samples used for learning model parameters; the validation set is the subset of samples used to evaluate the model training effect and set thresholds.
[0095] In practical applications, the training data only uses historical operating data under normal vehicle conditions. The dataset construction process is as follows: (1) Extract the historical records of the vehicle model under normal operating conditions from the cloud time series database, covering typical operating scenarios such as cold start, idling, constant speed driving, acceleration and deceleration, and hill climbing; (2) Extract the three-dimensional input vector for each time point; (3) Perform Z-Score standardization on each dimension to eliminate the difference in dimensions; (4) Use the sliding window method to divide the continuous time series data into a fixed-length sample sequence, with each window serving as an independent training sample; (5) Divide the processed normal sample set into a training set and a validation set in an 8:2 ratio.
[0096] In this embodiment, by standardizing and using a sliding window segmentation mechanism that covers historical data from multiple scenarios, the representativeness of training samples to the diversity of vehicle operating conditions is improved, the model's generalization ability to different operating conditions is improved, the model's insufficient adaptability caused by single-condition training data is improved, and the false alarm rate caused by operating condition drift in actual deployment is reduced.
[0097] In a specific embodiment, the specific technical implementation process of the stack anomaly detection and lifetime prediction channel in the inference stage is as follows.
[0098] (1) Implementation process of the Isolation Forest algorithm
[0099] Step 1 (Offline Training Phase): Collect historical stack voltage data of the same type of hydrogen fuel cell stack under normal operating conditions (including voltage distribution characteristics under various operating conditions, such as idling, cruising, climbing, and acceleration / deceleration), and construct a normal operating condition feature vector training set. Use this training set to train the Isolation Forest model. The core principle of the Isolation Forest model is to recursively construct multiple binary trees by randomly selecting features and randomly selecting the split values of those features. It leverages the characteristic of "few and distinct" outlier data, meaning that outliers are more easily isolated at shallower tree depths under random partitioning, thus achieving unsupervised anomaly detection.
[0100] Step 2 (Online Inference Stage): The edge gateway collects voltage data for each section of the fuel cell stack in real time during the current operating cycle (e.g., the last N sampling windows), constructing a feature vector with the same dimensions as the training set. This feature vector is then input into the trained isolation forest model, which calculates anomaly scores for each feature vector. The anomaly score is calculated based on the ratio of the average path length of the sample in each binary tree to the baseline path length. Typically, the anomaly score ranges from [0,1], with values closer to 1 indicating a higher likelihood of an anomaly.
[0101] Step 3 (Abnormal Cluster Determination): Compare the abnormal score output by the isolated forest with the preset abnormal determination threshold. If the abnormal score is greater than or equal to the threshold, the current operating state is determined to "fall into an abnormal cluster" (i.e., the voltage distribution characteristics of the current section of the fuel cell stack deviate from the normal operating state characteristic space) and marked as an abnormal state; if it is less than the threshold, it is determined to be a normal state.
[0102] (2) How are the results of abnormal cluster judgment used to calculate the overall health deviation of the fuel cell stack?
[0103] The relationship between the anomaly cluster determination results of the isolated forest and the health deviation calculation of the autoencoder can be divided into two cases: Case A (Isolation Forest is judged to be normal): The reconstruction error output by the autoencoder itself reflects the slight deviation between the current state and the healthy baseline. In this case, the health deviation HD(t) is only derived from the reconstruction error of the autoencoder (refer to the HD(t) calculation formula framework in the previous question answer), and the isolation forest does not apply any additional correction.
[0104] Case B (Isolation Forest Detected as Anomaly): The anomaly score of the isolation forest is used as an amplification factor for the health deviation of the autoencoder. In this case: HD'(t) = HD(t) × α(AS(t)); where AS(t) is the anomaly score of the isolation forest, and α() is the amplification function. The purpose of introducing the isolation forest is that the autoencoder excels at capturing "gradual deviations," while the isolation forest excels at capturing "abrupt distributions." The two complement each other, improving the detection speed of sudden voltage drops in the fuel cell stack (such as a sudden voltage collapse in a single section).
[0105] (3) Feature data output by the fuel cell stack anomaly detection model
[0106] The feature data output by the fuel cell stack anomaly detection model is not a single updated remaining useful life estimate, but a set of feature data, including: the current value of each section voltage and the statistical characteristics (mean, standard deviation, minimum, maximum) of nearly N windows; the isolated forest anomaly score AS(t); the autoencoder reconstruction error RE(t); the health deviation HD(t); and the corrected remaining useful life estimate RUL(t).
[0107] All five items above are packaged as feature data and uploaded to the cloud. The cloud requires the first four original features for population comparison analysis and model retraining.
[0108] (4) How is alarm data generated?
[0109] The alarm data generation logic of the fuel cell stack anomaly detection model is as follows: Alarm triggering conditions: An alarm event is generated when one of the following conditions is met: Condition 1 (Abnormal Cluster Alarm): The isolated forest anomaly score AS(t) ≥ the preset anomaly alarm threshold; Condition 2 (Lifetime Degradation Alarm): RUL(t) ≤ Preset lifetime alarm threshold; Condition 3 (Health Deviation Alarm): HD(t) ≥ Preset deviation alarm threshold.
[0110] Alarm data packet content: alarm type (abnormal cluster alarm / lifetime alarm / deviation alarm), alarm level (urgent / important / general), alarm trigger timestamp, specific indicator values that triggered the alarm (specific values of AS(t), HD(t), RUL(t)), and real-time snapshot data of the voltage of each section of the fuel cell stack when the alarm was triggered.
[0111] (5) The basic lifespan needs to be determined based on the normal aging rate, and then dynamically corrected using the health deviation.
[0112] The dynamic estimation of the remaining useful life (RUL) of the fuel cell stack in this application adopts a two-stage approach of "base life duration + dynamic correction of health deviation". The overall technical route is as follows: collecting parameters such as stack voltage distribution, temperature field, and pressure difference (all from the vehicle's native electronic control system (FCU fuel cell controller, etc.) and acquiring them via CAN bus); performing anomaly detection using a local AI model (based on autoencoder / isolated forest algorithm); identifying the stack degradation trend and predicting the remaining life.
[0113] The computational logic framework is as follows: Remaining useful life estimate RUL(t) = RUL_base(t) f(HD(t)); RUL(t): Corrected remaining useful life estimate of the fuel cell at time t, in hours. RUL_base(t): The base remaining lifetime at time t, calculated based on the conventional aging model of the fuel cell stack (which can be derived from the factory aging rate curve or by fitting the relationship between the cumulative operating time of the fuel cell stack and performance degradation), in hours. This value reflects the remaining time for the fuel cell stack to reach the performance failure threshold (usually defined as the rated power dropping to 80% or 90% of the initial value) when operating at the normal aging rate under standard operating conditions.
[0114] HD(t): Deviation of overall health of the fuel cell stack at time t, obtained by normalizing the reconstruction error output by the autoencoder model, with a value range of [0,1] or [0, +∞). The larger the value, the more serious the deviation from the health baseline.
[0115] f(HD(t)): Health deviation correction function, which maps HD(t) to lifespan correction coefficient.
[0116] Optionally, in some embodiments of this application, the stack anomaly detection and lifetime prediction data are obtained by inferring the voltage data of each section of the stack based on the stack anomaly detection and lifetime prediction model. This includes: constructing an isolated forest model using the isolated forest algorithm; inputting the voltage data of each section of the stack into the isolated forest model to obtain anomaly cluster judgment results; outputting the stack reconstruction error using an autoencoder; calculating the overall health deviation of the stack based on the anomaly cluster judgment results and the stack reconstruction error; updating the estimated remaining lifetime of the stack based on the health deviation; recording the voltage data of each section of the stack, the anomaly cluster judgment results, the stack reconstruction error, the health deviation, and the updated estimated remaining lifetime of the stack as feature data; and generating alarm events based on the anomaly cluster judgment results, the health deviation, and the updated estimated remaining lifetime of the stack to determine the stack anomaly detection and lifetime prediction data.
[0117] Among them, the Isolation Forest algorithm refers to an ensemble learning algorithm that achieves unsupervised anomaly detection based on random partitioning; the Isolation Forest model refers to an anomaly detection model composed of multiple random isolation trees; and the anomaly cluster judgment result refers to the judgment conclusion of the Isolation Forest on whether the current sample belongs to the anomaly category.
[0118] Among them, the stack reconfiguration error RE(t) refers to the reconfiguration error of the autoencoder on the stack input; the health deviation HD(t) refers to the quantified value of the overall deviation of the current state of the stack from the health baseline; and the estimated remaining useful life of the stack RUL(t) refers to the estimated number of remaining operating hours of the stack from the current state to the performance failure threshold.
[0119] In practical applications, the characteristic data includes: the current value of each section voltage and the statistical characteristics (mean, standard deviation, minimum value, maximum value) of the nearly N windows; the isolated forest anomaly score AS(t); the autoencoder reconstruction error RE(t); the health deviation HD(t); and the corrected remaining useful life estimate RUL(t).
[0120] The alarm data generation logic of the fuel cell stack anomaly detection model is as follows: an alarm event is generated when one of the following conditions is met simultaneously: Condition 1 (anomaly cluster alarm): the isolated forest anomaly score AS(t) ≥ the preset anomaly alarm threshold; Condition 2 (lifetime decay alarm): RUL(t) ≤ the preset lifetime alarm threshold; Condition 3 (health deviation alarm): HD(t) ≥ the preset deviation alarm threshold.
[0121] In this embodiment, the joint detection capability of gradual deviation and sudden anomaly of the fuel cell stack and the accuracy of remaining lifetime estimation are improved by using a dual-model collaborative anomaly detection and lifetime prediction mechanism of isolated forest and autoencoder.
[0122] Optionally, in some embodiments of this application, an isolated forest model is constructed using the isolated forest algorithm. The voltage data of each section of the fuel cell stack is input into the isolated forest model to obtain the abnormal cluster judgment result. This includes: collecting historical section voltage data of the target model of hydrogen fuel cell stack under normal operating conditions to construct a normal operating condition feature vector training set; training the initial isolated forest model based on the normal operating condition feature vector training set to obtain a trained isolated forest model; constructing feature vectors of the same dimension as the training set based on the voltage data of each section of the fuel cell stack, inputting the feature vectors of the same dimension as the training set into the trained isolated forest model to determine the abnormal score; and comparing the abnormal score output by the isolated forest with a preset abnormal judgment threshold to determine the abnormal cluster judgment result.
[0123] Among them, normal operating condition refers to the healthy operating condition in which the various performance indicators of the fuel cell stack are within the design range; terminal voltage data refers to the terminal voltage sampling value of each cell in the fuel cell stack; normal operating condition feature vector training set refers to the normal sample set organized according to a unified dimension; feature vector refers to the numerical input representation composed of multiple feature dimensions.
[0124] Here, the anomaly score refers to the quantitative score given by the isolation forest to the degree of sample anomaly, with a value range of [0,1]. The anomaly determination threshold is the critical score used to distinguish between normal and abnormal samples. The anomaly cluster judgment result refers to the conclusion of whether the current sample falls into an abnormal distribution.
[0125] In practical applications, Step 1 (Offline Training Phase): Collect historical segment voltage data of the same type of hydrogen fuel cell stack under normal operating conditions to construct a normal operating condition feature vector training set. Use this training set to train the isolated forest model. The core principle of the isolated forest is to recursively construct multiple binary trees by randomly selecting features and randomly selecting the segmentation value of those features, utilizing the characteristics of abnormal data to achieve unsupervised anomaly detection. Step 2 (Online Inference Phase): The edge gateway collects data in real time for the current operating cycle and constructs a feature vector with the same dimensions as the training set. Input this feature vector into the trained isolated forest model, and the model calculates anomaly scores for each feature vector. Step 3 (Anomaly Cluster Judgment): Compare the anomaly scores output by the isolated forest with a preset anomaly judgment threshold. If the anomaly score is greater than or equal to the threshold, it is marked as an abnormal state; if it is less than the threshold, it is judged as a normal state.
[0126] In this embodiment, the training of the unsupervised isolated forest algorithm on historical voltage data under multiple operating conditions and the abnormal score threshold determination mechanism improve the discrimination accuracy of abnormal stack status identification and reduce the labeling cost of abnormal samples and the false alarm and false negative rates.
[0127] Optionally, in some embodiments of this application, the overall health deviation of the fuel cell stack is calculated based on the abnormal cluster judgment result and the fuel cell stack reconstruction error, including: when the abnormal cluster judgment result is that the isolated forest is judged to be normal, the health deviation is obtained by mapping the fuel cell stack reconstruction error; when the abnormal cluster judgment result is that the isolated forest is judged to be abnormal, the abnormal score is used as an amplification factor for the health deviation to calculate the health deviation.
[0128] Here, "mapping" refers to converting the reconstruction error into a health deviation value through a preset function; "amplification factor" refers to the coefficient variable that amplifies and corrects the basic health deviation.
[0129] In practical applications, the relationship between the anomaly cluster determination result of the isolated forest and the health deviation calculation of the autoencoder can be divided into two cases: Case A (isolated forest is judged as normal): The reconstruction error output by the autoencoder itself reflects the slight deviation of the current state from the health benchmark. In this case, the health deviation HD(t) is only derived from the reconstruction error of the autoencoder. Case B (isolated forest is judged as abnormal): The anomaly score of the isolated forest is used as an amplification factor for the health deviation of the autoencoder. In this case: HD'(t) = HD(t) × α(AS(t)).
[0130] In this embodiment, the branching processing of the abnormal cluster determination results improves the comprehensiveness of the fuel cell health deviation calculation and the sensitivity of the abnormal response.
[0131] Optionally, in some embodiments of this application, generating an alarm event based on the abnormal cluster judgment result, health deviation degree, and updated estimated remaining lifespan of the fuel cell stack includes: generating an abnormal cluster alarm when the abnormal cluster judgment result indicates an abnormal score greater than or equal to a preset abnormal alarm threshold; generating a lifespan decay alarm when the estimated remaining lifespan of the fuel cell stack is less than or equal to a preset lifespan alarm threshold; generating a health deviation alarm when the health deviation degree is greater than or equal to a preset health deviation alarm threshold; and using the abnormal cluster alarm, lifespan decay alarm, and health deviation alarm as alarm events.
[0132] Among them, the abnormal cluster alarm refers to the alarm triggered when the voltage distribution of the current section of the fuel cell stack falls into an abnormal cluster; the lifetime degradation alarm refers to the preventive maintenance alarm triggered when the fuel cell stack RUL is lower than the threshold; and the health deviation alarm refers to the alarm triggered when the overall health deviation of the fuel cell stack exceeds the limit.
[0133] In practical applications, alarm triggering conditions are as follows: An alarm event is generated when one of the following conditions is met simultaneously: Condition 1 (abnormal cluster alarm): Isolation forest abnormal score AS(t) ≥ preset abnormal alarm threshold; Condition 2 (lifetime decay alarm): RUL(t) ≤ preset lifetime alarm threshold; Condition 3 (health deviation alarm): HD(t) ≥ preset deviation alarm threshold.
[0134] In this embodiment, a multi-channel alarm mechanism that is independently triggered by three types of indicators—abnormal score, remaining lifetime, and health deviation—improves the completeness of stack abnormal event detection and the accuracy of alarm classification. It also improves the problem that single threshold alarms are prone to missing potential abnormalities and reduces the response cost for maintenance personnel to miss or misjudge abnormal events.
[0135] Optionally, in some embodiments of this application, updating the estimated remaining useful life of the fuel cell stack based on the health deviation includes: obtaining a conventional aging model of the fuel cell stack by fitting the factory aging rate curve or the relationship between the cumulative operating time of the fuel cell stack and performance degradation; obtaining the base lifetime based on the conventional aging rate of the conventional aging model of the fuel cell stack; calculating the lifetime correction coefficient based on the health deviation correction function and the health deviation; and calculating the estimated remaining useful life of the fuel cell stack based on the base lifetime and the lifetime correction coefficient.
[0136] Among them, the factory aging rate curve refers to the standard aging rate reference curve obtained from the factory test of the fuel cell stack; the cumulative operating time of the fuel cell stack refers to the total number of operating hours of the fuel cell stack since it was put into use; the performance degradation relationship refers to the statistical relationship between the cumulative operating time of the fuel cell stack and the power decay; the fuel cell stack conventional aging model refers to the standard aging trend model fitted based on factory data or historical data; and the conventional aging rate refers to the natural degradation rate of fuel cell stack performance over time under standard operating conditions.
[0137] Among them, the base lifetime duration RUL_base(t) refers to the number of remaining hours from the current time to the failure threshold, calculated according to the conventional aging model; the health deviation correction function f(HD(t)) refers to the preset function that maps the health deviation to the lifetime correction coefficient; the lifetime correction coefficient refers to the proportional coefficient for dynamically correcting the base lifetime duration; and the estimated remaining lifespan of the fuel cell stack RUL(t) refers to the estimated remaining operating hours of the current fuel cell stack after correction.
[0138] In practical applications, the remaining useful life estimate RUL(t) = RUL_base(t) f(HD(t)); RUL(t): Corrected remaining useful life estimate of the fuel cell at time t; RUL_base(t): Base remaining useful life at time t calculated based on the conventional aging model of the fuel cell; HD(t): Overall health deviation of the fuel cell at time t; f(HD(t)): Health deviation correction function, which maps HD(t) to a life correction coefficient.
[0139] In this embodiment, a two-stage RUL estimation mechanism that combines basic lifetime duration with dynamic correction of health deviation improves the accuracy of remaining useful life estimation of the fuel cell stack and its adaptability to actual health status, thereby enhancing the accuracy of predictive maintenance of the fuel cell stack.
[0140] In a specific embodiment, the specific technical implementation process of the energy consumption prediction channel in the inference stage is as follows.
[0141] (1) Implementation process
[0142] Step 1 (Operating Condition Data Acquisition): The edge gateway acquires operating condition parameters such as vehicle speed v(t), motor output power P(t), instantaneous hydrogen consumption rate H2_rate(t), accelerator pedal opening Acc(t), current gear / load status Load(t), and road gradient (obtained from GPS + map data) from the CAN bus at a fixed frequency.
[0143] Step 2 (Energy Consumption Prediction Inference): Input the operating data sequence of the past N seconds into the lightweight Transformer prediction model. The model outputs the expected hydrogen consumption rate sequence H2_pred(t+Δt) for the next 5 minutes, where Δt∈[1, 300] seconds is the sampling interval. The prediction model is trained offline in the cloud based on the historical operating data of the entire fleet and the corresponding hydrogen consumption records. After training, it is quantized by INT8 and sent to the edge gateway.
[0144] Step 3 (Obtaining the Ideal Hydrogen Consumption Curve): Based on the physical model (polarization curve model) of the hydrogen fuel cell system, under the current operating conditions (vehicle speed, load, and gradient), and constrained by the optimal stack operating point (maximum efficiency point) and the optimal driving strategy (economic mode), the theoretical minimum hydrogen consumption curve is obtained through offline simulation or online calculation.
[0145] Step 4, the computational logic framework for calculating energy consumption deviation: Energy consumption deviation E_dev = (t, t+5min) [H2_pred(τ) - H2_ideal(τ)] dτ / (t,t+5min) H2_ideal(τ) dτ; The integral of the numerator represents the total difference between the predicted hydrogen consumption and the ideal hydrogen consumption over the next 5 minutes (i.e., the "excess hydrogen consumption"), while the denominator represents the total ideal hydrogen consumption. E_dev is dimensionless and takes a value ≥0, where 0 indicates that the predicted hydrogen consumption is equal to or lower than the ideal hydrogen consumption (optimal state), and a larger value indicates a more severe deviation.
[0146] (2) How feature data is generated
[0147] The characteristic data output by the energy consumption prediction channel includes: 1. Current operating condition snapshot (vehicle speed, load, instantaneous hydrogen consumption rate, accelerator pedal opening, road slope); 2. Expected hydrogen consumption curve for the next 5 minutes; 3. Energy consumption deviation E_dev; 4. Operating condition dimension with the largest deviation from the ideal curve and the corresponding deviation value (used to locate the cause of high energy consumption).
[0148] (3) How is alarm data generated?
[0149] Energy consumption-related alarms are not emergency safety alarms but rather operational management alarms. Alarm trigger condition: An energy consumption anomaly alarm event is generated when E_dev ≥ a preset alarm threshold. Alarm data includes: alarm type (energy consumption anomaly), vehicle ID, time period, E_dev value, the dimension of the operating condition with the largest deviation, and the corresponding driving behavior snapshot (obtained from the driving behavior channel). This alarm does not trigger local audible and visual alarms; it is only uploaded to the cloud for fleet administrators to view in the energy consumption report of the management app.
[0150] Optionally, in some embodiments of this application, inference is performed on the operating condition data based on the energy consumption prediction model to obtain energy consumption prediction data, including: predicting the operating condition data based on the lightweight converter prediction model to obtain the expected hydrogen consumption rate sequence within a future preset time period; solving for the theoretical minimum hydrogen consumption curve based on the physical model of the hydrogen fuel cell system, with the optimal stack operating point and optimal driving strategy as constraints; calculating the energy consumption deviation degree based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve; determining the operating condition dimension with the largest deviation from the theoretical minimum hydrogen consumption curve and the corresponding maximum deviation value based on the energy consumption deviation degree; recording the operating condition data, the expected hydrogen consumption rate sequence, the theoretical minimum hydrogen consumption curve, the operating condition dimension with the largest deviation, and the maximum deviation value as feature data; generating an alarm event based on the magnitude relationship between the energy consumption deviation degree and the preset energy consumption deviation alarm threshold to determine the energy consumption prediction data.
[0151] The lightweight converter prediction model refers to a time-series prediction model based on the lightweight Transformer architecture, i.e., the lightweight Transformer prediction model; the future preset time period is within the next 5 minutes, and the expected hydrogen consumption rate sequence H2_pred(t+Δt) refers to the time-series value of hydrogen consumption rate predicted by the model for the future period.
[0152] Among them, the optimal fuel cell stack operating point refers to the power and voltage combination at the point where the fuel cell stack efficiency is the highest; the optimal driving strategy refers to the driving mode constraints that minimize the hydrogen consumption of the entire vehicle; and the theoretical minimum hydrogen consumption curve H2_ideal(τ) refers to the ideal hydrogen consumption time series curve obtained by the physical model under optimal constraints.
[0153] Among them, the energy consumption deviation refers to the degree of relative deviation between the actual predicted hydrogen consumption and the ideal hydrogen consumption; the preset energy consumption deviation alarm threshold refers to the deviation threshold used to trigger an abnormal energy consumption alarm.
[0154] In practical applications, the characteristic data output by the energy consumption prediction channel includes: 1. Current operating condition snapshot; 2. Expected hydrogen consumption curve for the next 5 minutes; 3. Energy consumption deviation E_dev; 4. The operating condition dimension with the largest deviation from the ideal curve and its corresponding deviation value. Alarm triggering condition: When E_dev ≥ the preset deviation alarm threshold, an energy consumption anomaly alarm event is generated.
[0155] In this embodiment, the energy consumption deviation measurement mechanism, which compares the predicted hydrogen consumption of a lightweight Transformer with the ideal hydrogen consumption of a physical model, improves the ability to identify the deviation between the actual hydrogen consumption and the optimal hydrogen consumption of hydrogen-powered vehicles, thereby reducing the additional hydrogen consumption cost.
[0156] Optionally, in some embodiments of this application, the energy consumption deviation is calculated based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve, including: performing a first integral operation based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve within a future preset time period to obtain the total difference between the predicted hydrogen consumption and the ideal hydrogen consumption; performing a second integral operation based on the theoretical minimum hydrogen consumption curve within a future preset time period to obtain the total ideal hydrogen consumption; and using the ratio of the total difference to the total ideal hydrogen consumption as the energy consumption deviation.
[0157] The first integral operation refers to integrating the difference between the two curves over time over a future period; the total difference refers to the cumulative value of the difference between the predicted hydrogen consumption and the ideal hydrogen consumption over a future period; the second integral operation refers to integrating the ideal hydrogen consumption curve over time over a future period; the total ideal hydrogen consumption refers to the cumulative value of the theoretical minimum hydrogen consumption curve over a future period; and the energy consumption deviation refers to the ratio of the total difference to the total ideal hydrogen consumption.
[0158] In practical applications, energy consumption deviation E_dev = (t, t+5min) [H2_pred(τ) - H2_ideal(τ)] dτ / (t, t+5min) H2_ideal(τ) dτ, where the integral of the numerator is the total difference between the predicted hydrogen consumption and the ideal hydrogen consumption in the next 5 minutes, and the denominator is the total ideal hydrogen consumption.
[0159] In this embodiment, the energy consumption deviation is calculated by the two-curve integral ratio method, which improves the objectivity of the measurement of the deviation of actual hydrogen consumption from the optimal level and improves the evaluation effect of instantaneous hydrogen consumption index.
[0160] In a specific embodiment, the specific technical implementation process of the driving behavior recognition channel in the inference stage is as follows.
[0161] (1) Implementation process
[0162] Step 1 (Behavioral Data Acquisition): Real-time acquisition of driving operation signals such as accelerator pedal opening change rate, brake pedal action frequency and depth, vehicle speed change rate (acceleration distribution), and steering wheel angle change rate from the CAN bus to form a driving operation feature vector.
[0163] Step 2 (Driving Style Recognition): Input the driving operation feature vector sequence over a past period (e.g., 3 minutes) into the CNN-LSTM fusion model. The CNN layer extracts local operation pattern features (e.g., the "rapid acceleration-rapid deceleration" combination pattern), and the LSTM layer captures the temporal dependencies of the operation sequence (e.g., the mode transition from "continuous rapid acceleration followed by smooth cruising"). The model outputs the current driving style classification result (aggressive / mild / economical categories) and the corresponding confidence score.
[0164] Step 3, driving behavior score generation, calculation logic framework: score = w1 × acceleration smoothness score + w2 × braking smoothness score + w3 × vehicle speed stability score + w4 × energy consumption deviation score.
[0165] Meaning of each score: Acceleration smoothness score: Based on the statistical distribution of the accelerator pedal opening change rate, assesses whether frequent rapid acceleration exists. Braking smoothness score: Based on the frequency and depth of brake pedal action, assesses whether frequent sudden braking exists. Vehicle speed stability score: Based on the standard deviation of vehicle speed and the proportion of dwell time in the economic speed range. Energy consumption deviation score: Inversely mapped from E_dev; the smaller the E_dev, the higher the score.
[0166] Step 4 (Generate Personalized Energy-Saving Suggestions): Example 1: When the acceleration smoothness score is low, it is suggested that "Frequent rapid acceleration has been detected. It is recommended to keep the pedal opening change no more than 30% / s, which can reduce hydrogen consumption by about 8%-12%." Example 2: When the vehicle speed stability score is low, it is suggested that "The current vehicle speed fluctuates greatly. It is recommended to maintain cruise control within the economic speed range of 60-80km / h, which can improve the driving range by about 10%-15%." Example 3: When the energy consumption deviation score is low, it is suggested that "Under current operating conditions, your energy consumption is 15% higher than the optimal level. Please pay attention to the real-time energy consumption indicator on the instrument panel and keep it in the green zone."
[0167] At the same time, when the score is higher than 90, positive incentive prompts are given (such as "Current driving economy is excellent, score 92").
[0168] (2) How feature data is generated
[0169] The characteristic data output by the driving behavior channel includes: 1. Current driving style classification results and confidence level; 2. Driving behavior score and scores of each dimension; 3. Types of energy-saving suggestions triggered and trigger times; 4. Cumulative driving behavior statistics for this trip (frequency distribution of various operation events).
[0170] (3) How is alarm data generated?
[0171] The driving behavior channel does not generate safety alarms, only operational management-related early warning events. Warning trigger conditions include multiple consecutive trip scores below 40 points or serious dangerous driving behavior during a single trip (such as extreme acceleration / deceleration frequency exceeding a threshold). These warnings are then generated and uploaded to the cloud. Warning data includes: vehicle ID, driver ID, trip time period, score change trend, and a statistics list of specific violations. This warning is pushed to the driving behavior score ranking module in the management app for fleet administrators to monitor and intervene in.
[0172] Optionally, in some embodiments of this application, reasoning is performed on driving operation data based on a driving behavior recognition model to obtain driving behavior recognition data, including: constructing a driving operation feature vector based on the driving operation data; inputting the driving operation feature vector into a trained fusion model to output the current driving style classification result and the corresponding driving style confidence score; performing weighted summation based on various scores and weights to obtain a driving behavior score; generating personalized energy-saving suggestions based on the current driving style classification result; generating feature data based on the current driving style classification result, driving style confidence score, driving behavior score, personalized energy-saving suggestions, and driving operation data; and generating an alarm event based on the driving behavior score to determine the driving behavior recognition data.
[0173] Among them, the driving operation feature vector refers to the numerical input representation composed of various dimensions of driving operation data; the fusion model refers to the hybrid deep learning classification model that combines CNN and LSTM; the current driving style classification result refers to the determination of the driving style category as aggressive, mild or economical; and the driving style confidence score refers to the confidence probability value of the fusion model for the style classification result.
[0174] Among them, the driving behavior score refers to the quantitative score of driving economy obtained by combining multiple dimensions of factors; personalized energy-saving suggestions refer to energy-saving operation prompts generated based on the current driving style.
[0175] In practical applications, a sequence of driving operation feature vectors over a past period is input into the fusion model. The model outputs the current driving style classification result and the corresponding confidence score.
[0176] In this embodiment, the accuracy and real-time performance of driving style classification and behavior scoring are improved by using a CNN-LSTM fusion model for joint recognition of driving operation temporal features and a multi-factor weighted scoring mechanism.
[0177] Optionally, in some embodiments of this application, feature data is generated based on the current driving style classification result, driving style confidence level, driving behavior score, personalized energy-saving suggestions, and driving operation data. This includes: determining the suggestion type and trigger time associated with the personalized energy-saving suggestions; performing trip-cumulative driving behavior statistical analysis on the driving operation data to obtain driving behavior statistical data; and generating feature data based on the current driving style classification result, driving style confidence level, driving behavior score, suggestion type, trigger time, and driving behavior statistical data.
[0178] Among them, the suggestion type refers to the category label to which the personalized energy-saving suggestion belongs; the trigger time refers to the timestamp when the personalized energy-saving suggestion is generated; and the cumulative driving behavior statistical analysis of the trip refers to the summary statistics of the driving operations of a single trip.
[0179] In practical applications, the characteristic data output by the driving behavior channel includes: 1. Current driving style classification results and confidence level; 2. Driving behavior score and scores of each dimension; 3. Types of energy-saving suggestions triggered and triggering times; 4. Cumulative driving behavior statistics for this trip.
[0180] Optionally, in some embodiments of this application, cloud-based decision-making is performed on the fleet operation data in the time-series database based on pre-trained cloud-based decision engines to obtain cloud-based decision result data. This includes: extracting historical order data, vehicle utilization rate, maintenance cycle, and operating cost structure from the fleet operation data as input data for fleet size optimization; solving the fleet size optimization input data based on the fleet size optimization engine to obtain a vehicle configuration scheme; extracting hydrogen refueling station distribution, real-time traffic, weather forecast, and road restrictions from the fleet operation data as input data for route network planning; and solving the route network planning input data based on the route network planning engine to obtain a route. The network planning scheme involves extracting vehicle current health status, hydrogen reserves, hydrogen refueling station queue status, and hydrogen price from the entire fleet's operational data as input data for hydrogen refueling optimization. The hydrogen refueling optimization engine then solves this input data to obtain a hydrogen allocation scheme. Conversely, the fleet's operational data also includes extracting fuel cell stack and component degradation trends, group comparison anomaly reports, historical maintenance records, and spare parts inventory status as input data for predictive maintenance decisions. The predictive maintenance decision engine then solves this input data to obtain maintenance work orders. Finally, the cloud-based decision-making results are determined based on the vehicle configuration scheme, route network planning scheme, hydrogen allocation scheme, and maintenance work orders.
[0181] Among them, the vehicle configuration scheme refers to the number of vehicles of each model corresponding to the optimal solution; the route network planning scheme refers to the Pareto optimal path set that satisfies multiple objectives; and the hydrogen refueling allocation scheme refers to the hydrogen refueling station, timing and amount of hydrogen refueling allocated to each vehicle.
[0182] Among them, the degradation trend of fuel cell stack and its components refers to the performance degradation trend of components such as fuel cell stack over time; the group comparison anomaly report refers to the deviation analysis results of the health indicators of a single vehicle and the distribution of the group; and the maintenance work order refers to a standardized task order that includes maintenance time, spare parts list and working hours.
[0183] In practical applications, when the cloud-based large model decision sublayer receives a trigger event, it initiates the corresponding optimization solution process: First, the fleet size optimization solution process is as follows: the fleet size optimization engine runs a genetic algorithm to solve for the optimal vehicle configuration scheme under demand forecasting constraints; Second, the route network planning solution process is as follows: the route network planning engine runs a multi-objective search or ant colony algorithm to solve for the Pareto optimal path under the three-dimensional objectives of shortest time / lowest hydrogen consumption / optimal safety; Third, the hydrogen refueling optimization solution process is as follows: the hydrogen refueling optimization engine takes the vehicle health state vector, hydrogen reserve vector, hydrogen refueling station state vector, and hydrogen price vector as inputs and runs a mixed integer programming model to solve for the optimal hydrogen allocation scheme; Fourth, the predictive maintenance solution process is as follows: the predictive maintenance decision engine automatically generates maintenance work orders based on component life prediction values, including suggested maintenance time windows, required spare parts lists, and estimated working hours.
[0184] In this embodiment, the parallel optimization mechanism of four cloud-based decision engines under a unified triggering framework enhances the collaborative decision-making capabilities for fleet size configuration, route planning, hydrogen refueling scheduling, and predictive maintenance, thereby strengthening the intelligent decision-making capabilities for the entire lifecycle management of hydrogen fuel cell fleets.
[0185] In a specific embodiment, the specific technical implementation process of the fleet size optimization solution process in the decision-making stage is as follows.
[0186] Fleet size optimization is solved using a genetic algorithm. The objective function is to minimize total operating cost and maximize revenue. Constraints include: total capacity ≥ predicted demand, vehicle utilization rate 60%~90%, maintenance queue not exceeding the workstation limit, purchase cost not exceeding the budget, and the number of vehicle types being a non-negative integer. The genetic algorithm employs real-vector encoding, tournament selection, SBX crossover, and polynomial mutation. A penalty function handles the constraints, and a reinforcement learning policy network is invoked every 10 generations to dynamically adjust the crossover / mutation probabilities, accelerating convergence.
[0187] The implementation steps of the genetic algorithm for optimizing fleet size are as follows: Step 1 (Encoding and Initialization): Using the number of vehicles of each model as the decision variable, real number vector encoding (each gene position represents the number of vehicles of one model) is used to randomly generate the initial population within the range of satisfying the constraints.
[0188] Step 2 (Fitness Assessment): Calculate the objective function (minimize total operating cost, maximize revenue) for each individual. Use the penalty function method to multiply individuals that violate the constraints (total capacity ≥ predicted demand, vehicle utilization rate 60%~90%, maintenance queue ≤ workstation limit, purchase cost ≤ budget, number of vehicles is a non-negative integer) by a penalty coefficient and include them in the fitness, so that feasible solutions are given priority.
[0189] Step 3 (Tournament Selection): Each time, k individuals are randomly selected from the population, and the one with the best fitness is selected to enter the mating pool. This process is repeated until a new generation of parents is generated.
[0190] Step 4 (SBX Crossing): Perform simulated binary crossing (SBX) on the parent real vector according to the crossover probability to generate offspring that inherit the distribution characteristics of the parent.
[0191] Step 5 (Polynomial Mutation): Apply polynomial mutation perturbation to the offspring gene loci according to the mutation probability to maintain population diversity and enhance local search.
[0192] Step 6 (Adaptive Parameter Tuning for Reinforcement Learning): The reinforcement learning policy network is invoked once every 10 generations. Based on the current population diversity, convergence progress, and other states, the crossover and mutation probabilities are adjusted dynamically to accelerate convergence and avoid premature convergence.
[0193] Step 7 (Constraint Repair and Elite Retention): Repair individuals that have gone out of bounds after crossover / mutation, and retain the historical best individuals (elite strategy).
[0194] Step 8 (Termination Decision and Output): Stop when the maximum number of algebras or the convergence condition is reached, and output the optimal fleet size recommendation and vehicle configuration ratio.
[0195] Steps two through seven constitute the main evolutionary cycle (evaluation → selection → crossover → mutation → repair → retention), while step six acts as the "outer loop," intervening once every 10 generations to adjust the genetic operator parameters based on the feedback from the main cycle state. The two form a two-layered relationship of "main evolution + parameter self-regulation."
[0196] Optionally, in some embodiments of this application, the vehicle configuration scheme is obtained by solving the fleet size optimization input data based on the fleet size optimization engine, including: constructing a fleet size optimization model with the objective function of minimizing total operating cost and maximizing revenue, and with constraints such as total capacity constraint, vehicle utilization rate constraint, maintenance queue upper limit constraint, purchase cost constraint and vehicle type quantity constraint; and solving the fleet size optimization model using a fleet size optimization genetic algorithm to obtain the vehicle configuration scheme.
[0197] Here, the objective function refers to the evaluation index expression that needs to be maximized or minimized in the optimization model; the fleet size optimization genetic algorithm refers to a variant of the genetic algorithm used to solve the fleet size optimization problem.
[0198] In practical applications, fleet size optimization is solved using a genetic algorithm. The objective function is to minimize total operating costs and maximize revenue. Constraints include: total capacity ≥ predicted demand, vehicle utilization rate 60%~90%, maintenance queue not exceeding the workstation limit, purchase cost not exceeding the budget, and the number of vehicle types being a non-negative integer. The genetic algorithm employs real-vector encoding, tournament selection, SBX crossover, and polynomial mutation. A penalty function handles the constraints, and every 10 generations, a reinforcement learning policy network is invoked to dynamically adjust the crossover / mutation probabilities, accelerating convergence.
[0199] In this embodiment, the economy and business matching of the fleet size configuration scheme are improved by using an operations research optimization model that combines dual objectives and multiple constraints and a genetic algorithm solution mechanism. This reduces the operating loss costs caused by fleet idleness or insufficient capacity and enhances the scientific nature of fleet size planning.
[0200] Optionally, in some embodiments of this application, a fleet size optimization genetic algorithm is used to solve the fleet size optimization model to obtain a vehicle configuration scheme, including: using the number of vehicles of each type as the decision variable, using real number vector encoding, and randomly generating an initial population within the range of satisfying constraints; calculating the objective function for each individual in the initial population, and using the penalty function method to multiply individuals that violate the constraints by a penalty coefficient and include them in the fitness; randomly selecting a preset number of individuals from the initial population, selecting the one with the best fitness to enter the mating pool, until a new generation of parent set is generated; performing simulated binary crossover based on the new generation of parent set to generate crossover offspring; applying polynomial mutation perturbation to the gene loci of the offspring to generate perturbed offspring; performing dynamic parameter tuning operation every preset number of generations; repairing individuals that exceed the limits in the crossover offspring and perturbed offspring, and retaining the historical best individuals; stopping the iteration when the maximum number of generations or the convergence condition is reached, and outputting the vehicle configuration scheme based on the historical best individuals.
[0201] Here, decision variables refer to the unknowns that need to be solved in the optimization model; real vector encoding refers to the encoding method of chromosome genes using real vectors; initial population refers to the set of randomly generated candidate solutions; penalty function method refers to the method of reducing the fitness of solutions that violate constraints by a penalty coefficient; penalty coefficient refers to the fitness scaling factor applied to solutions that violate constraints.
[0202] Among them, fitness refers to the quantitative evaluation index of the superiority or inferiority of an individual in the population; mating pool refers to the set of parent individuals that participate in crossover to produce offspring; multinomial mutation refers to the continuous variable mutation perturbation operator based on multinomial distribution; dynamic parameter tuning refers to the operation of adaptively adjusting the algorithm parameters according to the evolutionary state; elite preservation refers to the strategy of directly copying the best individuals from previous generations to the next generation; convergence condition refers to the stopping criterion when the population no longer improves significantly.
[0203] In this embodiment, the global search capability and convergence speed of the genetic algorithm for optimizing fleet size are improved, the solution efficiency is increased, and the solution time and computational cost of large-scale fleet size optimization are reduced.
[0204] In a specific embodiment, the specific technical implementation process of the route network planning solution process in the decision-making stage is as follows.
[0205] Multi-objective A Search process: ① Load road network topology and hydrogen refueling station distribution, overlaying real-time traffic, weather, and road restriction data; ② Calculate the three-dimensional cost (time, hydrogen consumption, and safety risk) for each candidate path, where hydrogen consumption is corrected for slope / load, and the safety score is overlaid with weather / time-of-day risk factors; ③ Multi-objective A Search for expanded non-dominated path labels, and force the path to pass through at least one hydrogen refueling station; ④ After the search is completed, remove the dominated paths from the set of non-dominated paths and output the Pareto optimal path set (usually 2 to 5 paths); ⑤ The dispatcher selects on the mobile terminal or automatically recommends based on operational preferences.
[0206] The road network topology and hydrogen refueling station distribution, overlaid with real-time traffic, weather, and road restriction data, are external data acquired on demand by the cloud when the route planning engine is triggered. Specific data acquisition methods include: Road network topology and hydrogen refueling station distribution: static map data persisted locally in the cloud (third-party map service providers / hydrogen refueling station management systems); Real-time traffic: real-time retrieval via cloud-connected third-party traffic data service APIs; Weather forecast: cloud-connected meteorological data service APIs; Road restrictions: data interfaces from map service providers or traffic management departments (height restrictions, traffic restrictions, temporary controls, etc.).
[0207] In summary, these data are obtained in real time from external interfaces by the cloud when the route planning request is triggered. They are cloud-side data sources and are independent of the feature data stream from the edge to the cloud.
[0208] Optionally, in some embodiments of this application, the route network planning input data is solved based on the route network planning engine to obtain a route network planning scheme, including: acquiring external loading data; calculating three-dimensional cost data for each candidate path based on the external loading data; performing multi-objective search based on the three-dimensional cost data to determine a set of non-dominated paths; performing dominant path elimination on the non-dominated path set after the search is completed to output a Pareto optimal path set; determining the target path from the Pareto optimal path set according to the dispatcher's selection operation or the recommendation operation of the operation preference; and generating a route network planning scheme based on the target path.
[0209] Among them, external loading data refers to the set of real-time and static data from external third parties required for route planning; candidate paths refer to all possible driving paths available from the starting point to the destination; three-dimensional cost data refers to a three-dimensional evaluation vector consisting of driving time, hydrogen consumption, and safety risks; multi-objective search refers to a path search method that simultaneously seeks the optimal solution under multiple conflicting objectives; non-dominated path set refers to the set of paths that are not completely dominated by any other path under multiple objectives; Pareto optimal path set refers to the non-dominated path subset after filtering by dominance relations; and target path refers to the final execution path selected from the Pareto optimal path set.
[0210] In this embodiment, the route planning mechanism of three-dimensional cost multi-objective search and Pareto optimal selection improves the comprehensive optimization capability of the path scheme under multiple objectives of time, hydrogen consumption and safety, and improves the path selection effect that the single objective shortest path cannot take into account both economy and safety.
[0211] Optionally, in some embodiments of this application, obtaining externally loaded data includes: obtaining road network topology data and hydrogen refueling station distribution data based on static map data persisted locally in the cloud; pulling real-time traffic data in real time through a third-party traffic data service interface; obtaining weather forecast data through a cloud-connected meteorological data service interface; obtaining road restriction data through a map service provider interface; and generating externally loaded data based on the road network topology data, hydrogen refueling station distribution data, real-time traffic data, weather forecast data, and road restriction data.
[0212] Among them, static map data refers to geographic information data such as road network structure and landmarks that do not change over time; road network topology data refers to road connection relationship data composed of nodes and edges; hydrogen refueling station distribution data refers to hydrogen refueling station location and facility information data.
[0213] Among them, the third-party traffic data service interface refers to the external data service API that provides real-time traffic conditions; real-time traffic data refers to real-time information reflecting the current road congestion; meteorological data service interface refers to the external data interface that provides weather forecast information; weather forecast data refers to the weather forecast for a period of time in the future; and road restriction data refers to road traffic restriction information such as height restrictions, traffic restrictions, and temporary traffic control.
[0214] In this embodiment, the completeness and timeliness of the input data of the route planning engine are improved by pulling external data from multiple sources on demand in real time.
[0215] In a specific embodiment, the specific technical implementation process of the hydrogen replenishment optimization solution process in the decision-making stage is as follows.
[0216] A mixed-integer programming model is used: decision variables are vehicle-station allocation, hydrogen refueling quantity (continuous), and arrival time (integer). The objective function minimizes the hydrogen refueling cost + detour cost + waiting cost + unallocated penalty. Constraints include: hydrogen reserve constraint (hydrogen refueling quantity must not be less than the next task's requirement and must not exceed the safe available quantity), unique allocation per vehicle per station, capacity limits for hydrogen refueling stations at different times, health status limit constraints, and hydrogen refueling station inventory limit. A phased strategy is employed: branch and bound to solve for allocation → fixed allocation followed by LP to solve for hydrogen refueling quantity → backtracking iteration; for large-scale scenarios, urgency sorting + greedy allocation + local search are used.
[0217] The steps of the mixed integer programming (MILP) solution algorithm used in the staged approach are as follows: Step 1 (Modeling): Construct a MILP model with decision variables including vehicle-hydrogen station allocation (0-1 integer variables), hydrogen refueling quantity (continuous variable), and arrival time (integer variable). The objective is to minimize hydrogen refueling cost + detour cost + waiting cost + unallocated penalty.
[0218] Step 2 (Branch and Bound for Allocation): First, run branch and bound on the integer variable (vehicle-station allocation) to search for the optimal allocation scheme through the branch and pruning system.
[0219] Step 3 (Calculating the hydrogen refueling amount using LP after fixed allocation): After the allocation scheme is fixed, the original MILP degenerates into a pure linear programming (LP) program. The optimal hydrogen refueling amount for each vehicle (continuous variable) is directly solved using the simplex / interior point method.
[0220] Step 4 (Backtracking Iteration): If the LP is not feasible under the fixed allocation or violates other constraints, backtrack to the previous node in the branch and bound tree, adjust the allocation scheme and solve again until a feasible and better complete solution is obtained.
[0221] Step 5 (Output): Output the optimal timing, location, and amount of hydrogen added.
[0222] (1) The criteria for large-scale identification are: when the number of vehicles to be dispatched N or the variable scale of "number of vehicles × number of stations" exceeds the preset threshold, or when the branch delimitation cannot converge within the preset time limit, the switch is made (the specific threshold / time limit is determined by R&D based on the fleet size, such as N>50, number of variables exceeding 100,000, and solution time exceeding 30 seconds).
[0223] (2) Heuristic solution steps: Step 1 (Urgency Ranking): Calculate the urgency score for each vehicle (considering factors such as remaining hydrogen, health status (RUL / safety risk level), and distance to the nearest hydrogen refueling station), sort them from highest to lowest urgency, and determine the order of processing.
[0224] Step 2 (Greedy Allocation): Greedily select the hydrogen refueling station with the lowest marginal cost for each vehicle in order of urgency (shortest waiting time, lowest cost, and convenient route), while checking and updating the capacity / inventory constraints of each station in real time.
[0225] Step 3 (Local Search): Using the greedy solution as the initial solution, iteratively improve it through neighborhood operations (swapping the stops of two trains, adjusting one train to an adjacent time slot, etc.) until the total cost no longer decreases.
[0226] Step 4 (Output): Output near-optimal recommendations for hydrogenation timing, site, and hydrogenation volume.
[0227] In this process, step one, sorting, determines "who is allocated first," step two, greedily constructing feasible initial solutions in sequence, and step three, locally optimizing within the neighborhood of the initial solution, are all interconnected. The more accurate the sorting, the closer the greedy solution is to the optimal solution, and the smaller the space for improvement in local search.
[0228] Optionally, in some embodiments of this application, the hydrogen refueling optimization input data is solved based on the hydrogen refueling optimization engine to obtain a hydrogen refueling allocation scheme, including: using vehicle station allocation, hydrogen refueling amount, and arrival time as decision variables, constructing an objective function based on hydrogen refueling cost, detour cost, waiting cost, and unallocated penalty, and constructing a mixed integer programming model with hydrogen surplus constraints, station allocation constraints, hydrogen refueling station capacity limits for each time period, health status limit constraints, and hydrogen refueling station inventory limits as constraints; and solving the mixed integer programming model according to the mixed integer solving algorithm to obtain the hydrogen refueling allocation scheme.
[0229] Among them, mixed integer programming model refers to an optimization model in which decision variables contain both integer and continuous variables.
[0230] In this embodiment, a refined modeling mechanism that comprehensively considers cost, detour, waiting and health constraints through a hybrid integer programming model is used to improve the overall optimality of the hydrogen refueling allocation scheme under economic and vehicle health constraints, and improve the scheduling effect that takes into account multiple costs and constraints.
[0231] Optionally, in some embodiments of this application, the mixed integer programming model is solved according to the mixed integer solving algorithm, including: when the number of vehicles to be scheduled or the scale of the preset variables exceeds a preset threshold, or the branch and bound cannot converge within a preset time limit, the solution scale is determined as the target scale, and heuristic solving steps are performed to solve the model.
[0232] Among them, the number of vehicles to be dispatched, N, refers to the number of vehicles that need to be allocated a hydrogen refueling plan within the current hydrogen refueling dispatching cycle; the preset variable scale, i.e., the variable scale of "number of vehicles × number of stations", refers to the number of decision variables or the scale of the product of the number of vehicles and the number of stations.
[0233] Among them, the preset threshold refers to the scale or time critical value that triggers the switching of the solution strategy; branch and bound refers to the system search and pruning algorithm for solving integer programming; target size refers to the size of the problem that triggers heuristic solution, i.e. the judgment criterion for large scale; heuristic solution steps refer to the approximate optimization solution process based on empirical rules and local search.
[0234] In practical applications, the criteria for large-scale identification are: when the number of vehicles to be dispatched N or the variable scale of "number of vehicles × number of stations" exceeds a preset threshold, or when the branch delimitation cannot converge within a preset time limit, the switch is made.
[0235] In this embodiment, a hybrid strategy mechanism of large-scale scene recognition and heuristic solution switching is adopted to improve the solution efficiency and feasibility of hydrogen refueling optimization under different fleet sizes.
[0236] In a specific embodiment, the specific technical implementation process of the predictive maintenance solution process in the decision-making phase is as follows.
[0237] Regarding the fuel cell degradation trend, each round of inference at the edge generates a RUL(t) snapshot value and uploads it to the cloud. After accumulating these snapshots in the cloud time-series database, the "fuel cell degradation trend" needs to be obtained through two processing steps: Step 1 (Denoising and Smoothing): The original RUL(t) sequence is affected by sensor noise and short-term operating condition fluctuations, resulting in high-frequency jitter. A sliding window filter (such as Exponentially Weighted Moving Average (EWMA) or Savitzky-Golay filter) is used in the cloud to denoise the original RUL sequence, yielding a smoothed RUL trend curve.
[0238] The second step (trend inflection point detection): run the change point detection algorithm on the smooth curve to identify whether the RUL decay rate changes abruptly (e.g., from "slow linear decay" to "accelerated exponential decay"). The abrupt change point means that the fuel cell stack may have entered the accelerated aging stage.
[0239] Therefore, "pile decay trend" = RUL snapshot sequence uploaded from the edge → cloud-based noise reduction and smoothing → structured result after trend inflection point detection (including current decay rate, decay mode, and whether an acceleration inflection point was detected), which is a trend feature refined through cloud analysis.
[0240] Optionally, in some embodiments of this application, extracting the degradation trend of the fuel cell stack and its components includes: extracting the original remaining lifetime sequence from the entire fleet's operating data; denoising and smoothing the original remaining lifetime sequence based on sliding window filtering to obtain a smoothed remaining lifetime trend curve; and detecting trend inflection points based on the smoothed remaining lifetime trend curve to obtain the degradation trend of the fuel cell stack and its components.
[0241] Among them, sliding window filtering refers to a denoising method that uses a fixed window to perform a weighted average on the sequence along the time axis; denoising smoothing refers to the process of eliminating high-frequency random fluctuations in the original sequence to obtain a stable trend; the remaining lifetime trend curve refers to the time series change curve of RUL after denoising and smoothing; and trend inflection point detection refers to identifying the point in time when the curve decay rate changes abruptly.
[0242] In this embodiment, the two-step processing mechanism of sliding window denoising and smoothing and trend inflection point detection improves the availability and reliability of the degradation trend characteristics of the fuel cell stack and its components for cloud-based decision-making.
[0243] Optionally, in some embodiments of this application, multi-source heterogeneous vehicle perception data is acquired based on the data perception layer device deployed on the hydrogen energy vehicle, including: acquiring vehicle control data, battery management data, fuel cell control data, and hydrogen management data based on the vehicle's native electronic control system to form the vehicle's native electronic control data; acquiring hydrogen storage tank temperature and pressure data, hydrogen concentration sensing data, and vehicle camera data based on independently installed sensor groups to form sensor data; acquiring positioning data based on the positioning module; and obtaining vehicle perception data based on the vehicle's native electronic control data, sensor data, and positioning data.
[0244] The positioning module refers to the GPS or Beidou terminal module used to obtain vehicle location information; the positioning data refers to the location information such as latitude, longitude, speed, and direction output by the positioning module.
[0245] In practical applications, the vehicle data perception layer is deployed on hydrogen fuel cell vehicles and includes the vehicle's native electronic control system (VCU vehicle controller, BMS battery management system, FCU fuel cell controller, HMS hydrogen management system), independently installed sensor groups (hydrogen storage tank temperature / pressure sensor, hydrogen concentration sensor, vehicle camera), and GPS / BeiDou positioning module.
[0246] In this embodiment, a multi-source heterogeneous perception mechanism that integrates three types of data sources—the vehicle's native electronic control system, independently installed sensors, and positioning modules—improves the completeness and accuracy of the coverage of vehicle operating status by the onboard perception data.
[0247] Optionally, in some embodiments of this application, the method further includes: displaying alarm events in the hydrogen storage safety monitoring output data at the edge end through an in-vehicle display screen; and outputting personalized energy-saving suggestions in the driving behavior recognition data at the edge end in real time through a local audio-visual device.
[0248] Among them, the vehicle-mounted display screen refers to the visual display device deployed in the cockpit; the local audio-visual device refers to the alarm device at the vehicle end used for sound and light signal output.
[0249] In practical applications, driving suggestions and warning signals are output in real time through in-vehicle displays or audio-visual devices. If an abnormal state is detected, a local audio-visual alarm is triggered within 10 milliseconds, and a safety alarm event data packet is generated to prepare for uploading to the cloud.
[0250] In this embodiment, the real-time output mechanism of the vehicle display screen and the local audio-visual device improves the timeliness of safety alarms and energy-saving suggestions to the driver and the speed of execution response, thereby enhancing the safety monitoring and energy-saving guidance capabilities during the operation of hydrogen fuel cell vehicles.
[0251] Optionally, in some embodiments of this application, the method further includes: constructing a component lifespan prediction model in the cloud, and performing lightweight deployment on the edge based on the component lifespan prediction model to obtain a fuel cell stack anomaly detection and lifespan prediction model; constructing an energy consumption benchmark model in the cloud, and performing lightweight deployment on the edge based on the energy consumption benchmark model to obtain an energy consumption prediction model; constructing a driving behavior group profile model in the cloud, and performing lightweight deployment on the edge based on the driving behavior group profile model to obtain a driving behavior recognition model; configuring a configuration file in the cloud through a security rule base, and deploying a hydrogen storage safety monitoring model on the edge based on the configuration file; and using the fuel cell stack anomaly detection and lifespan prediction model, the energy consumption prediction model, the driving behavior recognition model, and the hydrogen storage safety monitoring model as an edge inference model.
[0252] Among them, the component life prediction model refers to the multi-component life prediction model trained in the cloud based on the data of the entire fleet; the energy consumption benchmark model refers to the theoretically optimal hydrogen consumption benchmark model established in the cloud based on the data of the entire fleet; and the driving behavior group profile model refers to the group behavior benchmark model obtained by clustering the data of the entire fleet in the cloud.
[0253] Among them, the safety rule base refers to the set of hydrogen storage safety thresholds and standardized parameter configurations maintained in the cloud; the configuration file refers to the parameter file generated by the safety rule base for deployment at the edge.
[0254] In practical applications, the offline model is trained and then deployed in a lightweight manner. The cloud-based full-precision model undergoes structural pruning, knowledge distillation, and INT8 quantization. Operator fusion and format conversion are performed. After integrity verification, the model is distributed via OTA. The old model is backed up at the edge and loaded on the next startup.
[0255] After each model training iteration, the cloud-based basic AI sublayer performs INT8 quantization and format conversion on the updated model parameters to generate model files suitable for OWA5X edge gateway NPU inference. These model files are then distributed to designated vehicles or the entire fleet via OTA (Over-The-Air) communication.
[0256] Regarding the hydrogen storage safety monitoring model: There is no corresponding version in the cloud. Hydrogen storage safety is a hard real-time safety task (<10ms). The model is pre-deployed at the edge and does not rely on cloud training and distribution. The cloud only manages it through safety rule base configuration; OTA delivers configuration files, not model files.
[0257] In this embodiment, a dual-track deployment mechanism combining cloud training and lightweight OTA delivery with security rule base configuration is adopted to improve the edge AI model's ability to balance accuracy and lightweight design, as well as the flexibility of continuous model iteration, thereby enhancing the stability of AI inference capabilities throughout the entire lifecycle of hydrogen fuel cell vehicles.
[0258] Optionally, in some embodiments of this application, the method further includes: obtaining the decision execution result obtained by the vehicle user after making a decision based on the decision result data in the cloud, and sending the decision execution result back to the cloud; updating the vehicle status in the time series database according to the decision execution result, and calibrating each cloud decision engine according to the decision execution result.
[0259] Among them, decision execution results refer to the actual effect information after vehicle users take actions based on cloud-based decisions; vehicle status refers to the latest status values of various indicators of the vehicle after the decision is executed; and engine calibration refers to the process of correcting model parameters or constraints based on actual execution results.
[0260] In practical applications, during the execution phase: Decision results generated in the cloud are pushed to the corresponding mobile app in real time. When the hydrogen refueling optimization engine outputs a hydrogen refueling suggestion, the driver's app receives a push notification displaying recommended refueling stations, estimated waiting times, and recommended hydrogen refueling volume. The driver can then initiate navigation with a single click. When the predictive maintenance engine generates a maintenance work order, the management app receives the work order notification. The dispatcher can view the work order details, confirm the maintenance time, and assign maintenance personnel. When the driving behavior recognition model produces a real-time score, the driver's app's driving assistant interface synchronously updates the score and energy-saving prompts. After execution, the results are transmitted back to the cloud via the app for updating vehicle status and calibrating the decision model.
[0261] In this embodiment, a closed-loop feedback mechanism of decision execution result feedback and cloud decision engine calibration is used to improve the adaptability and prediction accuracy of the cloud decision model to the actual execution effect.
[0262] Optionally, in some embodiments of this application, the method further includes: obtaining difficult example inference results with model confidence below a threshold and abnormal working condition identification results with deviations from cloud predictions exceeding a threshold from the edge inference result data; determining the difficult example inference results and abnormal working condition identification results as difficult example data in the inference process; uploading the difficult example data to the cloud to construct the incremental dataset required for the next model training.
[0263] Among them, model confidence refers to the quantitative measure of the certainty of the AI model's own reasoning results; confidence threshold refers to the critical value used to filter out low-confidence reasoning results; and difficult example reasoning results refer to sample results whose model confidence is below the threshold and which the model has difficulty accurately identifying.
[0264] Among them, the difficult example data refers to the sample set consisting of the reasoning results of difficult examples and the identification results of abnormal working conditions; the incremental dataset refers to the sample set added to supplement the original training set for the next round of model training.
[0265] In practical applications, edge gateways identify and cache difficult example data during inference, and upload it to the cloud along with the regular data stream to supplement the incremental dataset for the next model training. Different models use different confidence measurement methods.
[0266] In this embodiment, the model evolution mechanism, which involves automatic identification of difficult cases and construction of incremental datasets, enhances the continuous learning ability and prediction accuracy of the edge AI model on long-tail conditions and boundary samples.
[0267] The technical research process and other technical details of this application are described below with reference to a specific embodiment.
[0268] Hydrogen fuel cell vehicles, with their advantages of zero emissions, long range, and fast refueling, are becoming an important technological route to replace fossil fuels in the medium and heavy-duty commercial vehicle sector. However, in stark contrast to the rapid deployment of vehicles, the operation and maintenance management technology for hydrogen fuel cell vehicles is severely lagging behind. The industry generally still adopts a rudimentary model of "data transmission via onboard TBOX + manual review of reports in the cloud," which is insufficient to meet the rigid requirements of safety, economy, and efficiency for large-scale commercial operation.
[0269] Currently, the industry offers several representative existing technical solutions for the operation and maintenance management of hydrogen fuel cell vehicles: First, single-function fault monitoring solutions. Some companies have deployed fault warning systems based on edge computing terminals on hydrogen fuel cell vehicles. These solutions use AI inference models running on the onboard terminal to monitor key parameters of the fuel cell (such as cell voltage, temperature, and pressure) in real time, triggering alarms when anomalies are detected. Representative technical approaches include: using time-series prediction models such as LSTM / CNN to predict future cell voltage values and triggering warnings when the predicted values deviate from the normal range; and using rule engines or simple machine learning models (such as isolated forests) to detect anomalies in sensor data. However, these solutions have two fundamental limitations: first, their functionality is limited, covering only fault warning scenarios and not addressing operational needs such as driving behavior optimization, energy consumption management, and hydrogen refueling scheduling; second, the collaboration between the edge and cloud remains superficial, with the cloud passively receiving alarm data for post-event analysis and report display, rather than proactively making business decisions based on global data and updating the edge model accordingly.
[0270] Second, single-function hydrogen refueling scheduling solutions. Some technical solutions focus on solving the problems of queue management and vehicle scheduling at hydrogen refueling stations. For example, by collecting information such as the number of vehicles queuing at hydrogen refueling stations, the hydrogen demand per vehicle, and the occupancy status of hydrogen refueling nozzles, they calculate and predict waiting times to guide vehicles to choose the optimal hydrogen refueling station. The drawback of this type of solution is that it is completely decoupled from the vehicle side. The hydrogen refueling scheduling decision relies solely on information from the hydrogen refueling station side (queue size, refueling speed, etc.), without knowing the actual health status of the vehicles (whether the fuel cell stack is in a period of accelerated degradation, whether there is a risk of micro-leakage in the hydrogen storage tank, etc.). Therefore, it is impossible to make refined decisions on "hydrogen refueling strategies based on vehicle health status," such as whether to prioritize refueling vehicles with poor fuel cell stack condition to avoid them operating in an inefficient range.
[0271] Third, discrete fleet management solutions. Some connected vehicle platforms offer basic fleet management functions such as vehicle location tracking, driving trajectory playback, and mileage and fuel consumption statistics. However, these functions are mainly "post-event statistics" and lack real-time AI analysis capabilities, nor do they involve predictive maintenance. That is, they cannot predict the remaining service life of parts and failure time windows based on real-time vehicle operation data. Maintenance work still relies on fixed-cycle planned maintenance or reactive repairs after a failure occurs.
[0272] The three types of solutions mentioned above collectively constitute the current technological status quo of hydrogen fuel cell vehicle operation and maintenance management. Their common shortcomings can be summarized in three levels: At the data level, vehicle health data, hydrogen refueling scheduling data, driving behavior data, and fleet operation and maintenance data are stored in isolated systems, forming "data silos"; at the decision-making level, each system makes locally optimal decisions independently (e.g., the fault warning system only cares about whether an alarm is triggered, and the hydrogen refueling scheduling system only cares about queuing time), lacking a global AI engine to integrate multi-dimensional information and make overall optimal decisions; at the execution level, there is a lack of an end-to-end automated closed-loop link from "AI diagnosis" to "decision output" to "work order execution" to "effect feedback," with many links relying on manual judgment and manual operation. These shortcomings are tolerable when the fleet size is small, but when the fleet size reaches hundreds of vehicles, the operational efficiency loss caused by data silos and fragmented decision-making will be amplified exponentially. Therefore, the industry urgently needs an integrated solution that can connect the entire chain of "vehicle edge monitoring - cloud AI decision-making - hydrogen refueling scheduling - mobile operation and maintenance."
[0273] Based on this, this application provides a method for full lifecycle operation and maintenance decision-making of hydrogen fuel cell vehicles based on edge-cloud collaboration, also known as a method for full lifecycle operation and maintenance decision-making of hydrogen fuel cell vehicles based on edge-cloud collaborative AI, which relates to the field of intelligent monitoring and operation and maintenance management technology for hydrogen fuel cell vehicles. Details are as follows.
[0274] This application aims to design a hydrogen fuel cell vehicle full lifecycle operation and maintenance decision-making system based on edge-cloud collaborative AI, in order to solve the fundamental problems in existing technologies such as vehicle health monitoring, hydrogen refueling scheduling, driving behavior optimization, and fleet operation and maintenance management being independent, data not being shared, and lacking a global AI decision engine.
[0275] This application, leveraging the in-vehicle AI inference capabilities of the OWA5X edge computing gateway and the large-scale model and operations optimization capabilities of the Corvina cloud IoT platform, creatively establishes a five-stage operation and maintenance decision-making closed-loop paradigm: "multi-scenario joint real-time monitoring at the edge → data aggregation and feature extraction → global AI decision-making in the cloud → decision command issuance and execution feedback → bidirectional closed-loop model update." Unlike existing solutions where the edge only performs data transmission or simple threshold judgment, and the cloud only performs passive data storage and report display, this application positions the edge as a millisecond-level response layer of "real-time perception - immediate control - data preprocessing," and the cloud as a minute- to hour-level decision-making layer of "global analysis - strategic decision-making - model evolution." The two form a deep bidirectional collaboration through structured data flow and model update flow.
[0276] This application has the following technical highlights: 1. For the first time, a closed-loop decision-making system for the operation and maintenance of hydrogen fuel cell vehicles, covering the entire system from "vehicle-station-fleet-personnel," has been established. By integrating edge-based vehicle health monitoring and driving behavior analysis, cloud-based fleet scheduling and hydrogen refueling strategy optimization, and mobile-based multi-role collaborative management into a unified system, the system solves the pain points of information asymmetry in the industry, such as "vehicles not knowing where to refuel, stations not knowing who urgently needs hydrogen refueling, and fleets not knowing which vehicle needs maintenance." It achieves a closed-loop system with no human intervention throughout the entire process, from AI-based automatic diagnosis to automatic decision-making to automatic dispatch.
[0277] 2. Pioneering the use of real-time vehicle health status as the core input variable for hydrogen refueling dispatch decisions.
[0278] Unlike existing hydrogen refueling scheduling systems that rely solely on location and hydrogen reserves for simple queuing, this application uses deep diagnostic indicators such as the predicted remaining lifespan of the fuel cell stack generated by edge AI, the safety risk level of the hydrogen storage cylinder, and the fuel cell health score as constraints for the hydrogen refueling scheduling model. This enables a refined hydrogen refueling strategy based on vehicle health status (such as prioritizing vehicles on low-load refueling routes during the accelerated fuel cell stack degradation phase to avoid high-load operation that accelerates fuel cell stack aging), filling a gap in existing technologies.
[0279] 3. Implementing parallel secure inference for heterogeneous AI models on the same edge gateway
[0280] Under the strict resource constraints (computing power, power consumption, heat dissipation) and real-time requirements (millisecond-level response for safety tasks) of the vehicle environment, this solution addresses the problems of high cost, difficult integration, and inability to merge data caused by each monitoring function belonging to an independent hardware system in existing solutions. This is achieved by using priority preemptive scheduling and NPU hardware resource isolation technology to simultaneously deploy and run four types of heterogeneous AI models for stack health, hydrogen storage safety, driving behavior, and energy consumption prediction.
[0281] 4. A two-way closed-loop update mechanism for edge-cloud AI models has been established.
[0282] The high-precision model trained on the entire fleet's historical data is regularly deployed to the edge to replace the old model. Difficult example data accumulated at the edge during inference is uploaded to the cloud for incremental dataset expansion for model retraining, forming a continuous evolution loop of "cloud training → edge inference → difficult example back transmission → cloud optimization → model deployment", ensuring that the AI model maintains high prediction accuracy throughout its entire lifecycle.
[0283] The key technical points of the edge-cloud collaborative hydrogen fuel cell vehicle full lifecycle operation and maintenance decision-making method proposed in this application are as follows.
[0284] 1. The overall system architecture consists of five stages: "Edge-end multi-scenario joint real-time monitoring layer → Data aggregation layer → Cloud-based global AI decision-making layer → Mobile execution layer → Model bidirectional closed-loop update bus", as well as the connection relationships and data flow between the layers.
[0285] In this architecture, the edge is positioned for millisecond-level real-time perception and instant control (embedded with four parallel AI inference channels: fuel cell health, hydrogen storage safety, driving behavior, and energy consumption prediction), while the cloud is positioned for minute- to hour-level global analysis and strategic decision-making (embedded with four decision engines: fleet size optimization, route planning, hydrogen refueling optimization, and predictive maintenance). The two layers form a deep bidirectional collaboration through the uplink flow of structured feature data and the downlink update flow of model parameters. This is different from the shallow collaboration mode in existing solutions where the edge only performs data transmission or single fault detection, and the cloud only performs passive data storage.
[0286] 2. A cross-process data linkage mechanism that uses the real-time health status of vehicles as a constraint for hydrogen refueling scheduling decisions.
[0287] Health status indicators such as the predicted remaining lifespan of the fuel cell stack, the safety risk level of the hydrogen storage tank, and the comprehensive health score of the fuel cell generated by the edge AI inference layer are transmitted to the cloud-based hydrogen refueling optimization engine via a data flow bus. These indicators serve as input variables for hydrogen refueling timing selection, hydrogen refueling station recommendation, and hydrogen refueling quantity decision-making, enabling refined scheduling that "dynamically adjusts the hydrogen refueling strategy based on the vehicle's health status." This differs from the extensive mode of hydrogen refueling scheduling in existing technologies that only rely on vehicle location and remaining hydrogen.
[0288] The hydrogen reserve is collected in real time at the edge and uploaded to the cloud along with the feature data stream. It comes from the FCU hydrogen management system and is calculated by reading the pressure / temperature of the hydrogen storage tank via the CAN bus. The cloud always keeps the latest values. The queuing status of the hydrogen refueling station is collected by the cloud through the API of the hydrogen refueling station management system. It includes the number of vehicles currently in the queue, the occupancy status of each hydrogen refueling gun, and the estimated waiting time. The cloud caches and refreshes it periodically (typically every 1 minute). The hydrogen price is collected by the cloud through the API of the hydrogen refueling station management system or a third-party price interface to obtain the real-time listed price of each station. The price is actively pushed to update when there is a change. If the hydrogen refueling station system does not support the real-time price interface, the price table can be manually maintained daily as an initial solution, and automatic synchronization can be achieved later through API integration.
[0289] 3. Priority-based preemptive parallel inference scheduling mechanism for heterogeneous AI models within the edge gateway.
[0290] A priority-based preemptive task scheduler is deployed on the NPU processor of the OWA5X edge gateway. The hydrogen storage safety monitoring model is set to the highest priority (hard real-time, response latency constraint less than 10 milliseconds), the stack anomaly detection model is set to high priority (hard real-time, response latency constraint less than 10 milliseconds), the energy consumption prediction model is set to medium priority (soft real-time, response latency constraint less than 50 milliseconds), and the driving behavior recognition model is set to normal priority (soft real-time, response latency constraint less than 30 milliseconds). At any time, the scheduler prioritizes allocating NPU computing resources to the highest priority inference task that is currently in the ready state, ensuring that the response latency of safety-critical tasks is not affected by the inference load of other models.
[0291] 4. Heterogeneous data joint reasoning and cross-subsystem associated fault identification mechanism for fuel cell stack health and hydrogen storage safety.
[0292] The outputs of the fuel cell stack anomaly detection model (voltage deviation values of each section of the fuel cell stack, and temperature difference anomalies between the fuel cell stack inlet and outlet) and the outputs of the hydrogen storage safety monitoring model (temperature change rate anomalies of the hydrogen storage tank) are subjected to cross-model feature cross-analysis at the edge. When a causal relationship is detected between the fuel cell stack cooling-related anomaly indicators and the hydrogen storage tank temperature anomaly indicators in the time series, a cross-subsystem correlation alarm of "fuel cell stack cooling failure leading to thermal risk of hydrogen storage system" is generated, triggering joint safety protection measures (simultaneously limiting the fuel cell stack output power and activating the auxiliary heat dissipation device of the hydrogen storage tank).
[0293] 5. A two-way closed-loop OTA update and incremental learning mechanism for AI models from the cloud to the edge.
[0294] The cloud periodically distributes retrained and calibrated model parameters (including fuel cell lifetime prediction model, energy consumption baseline model, and driving style classifier weights) based on the entire fleet's historical data to the edge gateway via OTA channel after INT8 quantization compression and integrity verification. After receiving the new model, the edge gateway saves a backup of the old model and loads the new model parameters the next time the vehicle starts. Difficult case data (samples with model confidence below a preset threshold or prediction deviation exceeding a preset threshold) identified by the edge gateway during inference are marked and uploaded to the cloud, and included in the incremental dataset for the next round of model training. The entire mechanism forms a continuous evolutionary closed loop of "cloud training → edge inference → difficult case collection → cloud incremental learning → model distribution".
[0295] The technical effects of the edge-cloud collaborative hydrogen fuel cell vehicle full lifecycle operation and maintenance decision-making method proposed in this application are as follows.
[0296] This application creatively integrates edge multi-scenario joint real-time monitoring, cloud-based global AI decision-making, and mobile multi-role collaborative management into a five-stage closed-loop architecture, bringing the following significant and substantial technical effects: 1. A paradigm shift in the operation and maintenance of hydrogen fuel cell vehicles has been achieved, moving from "single-point passive response" to "global proactive decision-making".
[0297] In traditional solutions, vehicle fault detection, driving behavior assessment, hydrogen refueling scheduling, and maintenance planning operate in four independent systems with data that is not interconnected. This application, through a unified edge gateway data collection standard, a unified cloud data fusion engine, and a unified AI decision-making entry point, enables, for the first time, data from these four stages to be cross-referenced and jointly optimized, producing technical effects that cannot be achieved by a single subsystem. For example, the health status of the fuel cell stack automatically influences the hydrogen refueling strategy, and driving behavior scores are automatically incorporated into maintenance risk prediction. This cross-stage automated synergy is something that no existing single system possesses.
[0298] 2. Significantly improved the real-time safety response and reliability of hydrogen fuel cell vehicles.
[0299] By allocating the highest priority preemptive scheduling authority and an independent hardware watchdog to the hydrogen storage safety monitoring model at the edge, local millisecond-level response to dangerous events such as hydrogen leaks is achieved (without round trips to the cloud), eliminating safety blind spots caused by network latency and cloud service interruptions. Simultaneously, cross-model joint inference of stack health and hydrogen storage safety enables the system to identify faults associated with each subsystem (such as thermal risks in hydrogen storage cylinders caused by abnormal stack cooling), which cannot be detected independently by any single system in traditional discrete monitoring architectures.
[0300] 3. Significantly reduced fleet operating costs and improved vehicle availability.
[0301] The cloud-based predictive maintenance decision engine reduces unplanned downtime by 40%-60%, shifting maintenance from "reactive repair" to "prevention." The hydrogen refueling optimization engine, based on vehicle health status, performs refined hydrogen refueling scheduling, reducing accelerated fuel cell stack aging caused by inappropriate refueling strategies. The real-time driving behavior feedback system guides drivers to develop economical driving habits, expected to reduce hydrogen consumption by 12%-18%. For a fleet of 100 hydrogen-powered commercial vehicles, these three optimizations can save over 1.8 million yuan in operating costs annually.
[0302] 4. It solves the performance bottleneck of parallel inference of multiple AI models in resource-constrained environments at the edge.
[0303] By employing a three-pronged approach—INT8 quantization compression, NPU operator fusion acceleration, and priority-based preemptive task scheduling—this application achieves the parallel deployment and stable operation of four heterogeneous AI models on a single edge gateway, within constraints that limit the total model size to no more than 40MB and the worst-case safety task response latency to no more than 50 milliseconds. Compared to the traditional approach of deploying independent hardware terminals for each monitoring function, this application reduces hardware costs by approximately 60% and significantly lowers system integration complexity.
[0304] 5. A technological foundation has been established for the continuous self-evolution of AI models during system operation.
[0305] Through a four-step closed-loop mechanism of "automatic labeling and back-transmission of difficult examples → incremental learning in the cloud → OTA model deployment → online updates at the edge," the prediction accuracy of the AI model does not decay over time or with environmental changes, but rather continues to optimize as the amount of operational data increases. This mechanism is particularly suitable for real-world scenarios where hydrogen fuel cell vehicles operate under diverse conditions (across temperature zones, altitudes, and load ranges), ensuring that the system maintains high accuracy under various operating conditions even after large-scale deployment.
[0306] It should be noted that any technical feature in any of the above embodiments provided in this application is also applicable to any of the following embodiments provided in this application; the technical features in the relevant embodiments of the various methods provided in this application are also applicable to the relevant embodiments of the various devices, systems or equipment provided in this application; the same, related or corresponding technical features in the various embodiments provided in this application can be referenced and explained to each other, and similarities will not be repeated.
[0307] Figure 2 Here is a structural block diagram of the system in one embodiment, with reference to Figure 2 The edge-cloud collaborative hydrogen fuel cell vehicle lifecycle operation and maintenance decision-making system includes: The vehicle data sensing module 201 is used to acquire multi-source heterogeneous vehicle sensing data based on the data sensing layer equipment deployed on hydrogen energy vehicles, and to aggregate the vehicle sensing data to the edge computing gateway at the edge. The edge inference module 202 is used to preprocess and extract features from the vehicle perception data of the edge computing gateway to obtain feature-extracted data; calculate the priority information of each operation and maintenance decision task based on the priority preemptive task scheduler; perform millisecond-level real-time inference on the feature-extracted data according to the task execution order represented by the priority information and based on the pre-deployed edge inference models; summarize the feature data and alarm events of the inference process as edge inference result data; and upload the edge inference result data to the cloud. The cloud decision module 203 is used to receive edge inference result data uploaded from the edge terminal based on the message queue, and construct a time series database based on the edge inference result data of each hydrogen energy vehicle; in response to the triggering event of cloud decision, it performs cloud decision on the fleet operation data of the entire fleet in the time series database based on the pre-trained cloud decision engines, and obtains cloud decision result data. The mobile collaboration module 204 is used to send cloud decision-making results data to the mobile terminal and display the cloud decision-making results data on the mobile terminal for vehicle users associated with hydrogen energy vehicles to make decisions.
[0308] In this embodiment of the application, based on, as follows Figure 2 The connections between the various modules or units shown in the diagram enhance the overall efficiency of operation and maintenance decision-making for hydrogen-powered vehicles and improve the effectiveness of overall operation and maintenance decisions through the cooperation between these modules or units.
[0309] The following section describes the system composition and technical principles of the edge-cloud collaborative full lifecycle operation and maintenance decision-making system for hydrogen fuel cell vehicles from the perspective of other technical descriptions.
[0310] The edge-cloud collaborative hydrogen fuel cell vehicle full lifecycle operation and maintenance decision system provided in this application is also known as the edge-cloud collaborative AI hydrogen fuel cell vehicle full lifecycle operation and maintenance decision system, or a hydrogen fuel cell vehicle intelligent operation and maintenance decision system that covers the entire lifecycle of "vehicle monitoring - fleet management - hydrogen refueling scheduling - mobile operation and maintenance" based on the collaborative work of edge computing gateway and cloud artificial intelligence platform.
[0311] Figure 3 This is a system architecture diagram of one embodiment. The structure of the system is as follows: Figure 3 As shown, its structure and the functions of each part are as follows.
[0312] The system comprises four core components: an in-vehicle data perception layer, an edge AI inference layer, a cloud AI decision-making layer, and a mobile collaboration layer, as well as a data flow bus and model update bus that run through these four layers. Among them: (1) Vehicle Data Sensing Layer: Deployed on hydrogen fuel cell vehicles, this layer includes the vehicle's native electronic control system (VCU vehicle controller, BMS battery management system, FCU fuel cell controller, HMS hydrogen management system), independently installed sensor groups (hydrogen storage tank temperature / pressure sensor, hydrogen concentration sensor, vehicle camera), and GPS / BeiDou positioning module. This layer aggregates multi-source heterogeneous data to the OWA5X edge computing gateway via CAN bus and analog / digital interfaces. The core function of this layer is to complete the raw acquisition of the entire vehicle's operating status data, providing a data foundation for upper-layer AI inference.
[0313] (2) Edge AI Inference Layer: Based on the OWA5X edge computing gateway, this layer incorporates a multi-core heterogeneous AI processor (CPU+NPU+GPU). This layer deploys four types of lightweight AI models: a driving behavior recognition model (based on a MobileNet feature extraction backbone superimposed with an LSTM time-series classification head, approximately 10MB), an energy consumption prediction model (based on a lightweight Transformer architecture, approximately 20MB), a fuel cell stack anomaly detection and lifetime prediction model (a hybrid model based on the isolated forest algorithm and a deep autoencoder, approximately 3MB), and a hydrogen storage safety monitoring model (an anomaly detection model based on autoencoder reconstruction error analysis, approximately 3MB). All models are quantized using INT8 and then accelerated for inference on the NPU.
[0314] This layer also runs a priority-based preemptive task scheduler, with safety-related tasks (hydrogen leak detection) having the highest priority, and energy consumption optimization and driving behavior analysis having normal priority.
[0315] The core functions of this layer are: time alignment, outlier filtering, and feature extraction of raw sensor data; running AI models for millisecond-level real-time inference; uploading processed feature data and alarm events to the cloud via 4G / 5G networks; and outputting driving suggestions and alarm signals in real time through in-vehicle displays or audio-visual devices.
[0316] (3) Cloud AI Decision Layer: Deployed on the Corvina cloud IoT platform, it consists of a basic AI sublayer and a large model decision layer.
[0317] The basic AI sublayer includes: a time-series database (stores vehicle feature data and historical operation records), a message queue (receives real-time data streams uploaded from the edge), and an offline model training engine (trains and updates component life prediction models, energy consumption benchmark models, and driving behavior group profile models based on historical data from the entire fleet).
[0318] The historical operation records are collected during the inference phase of the edge AI inference layer. After inference is completed at the edge, the "feature data and alarm events" are packaged and uploaded to the cloud, where they are written into a time-series database and stored as historical operation records. Therefore, historical operation records explicitly include alarm events. Alarm events are uploaded and stored as independent fields along with the feature data, and are used for fault statistical analysis, model training, and operation and maintenance decisions.
[0319] The cloud-based component lifespan prediction model corresponds to the edge-based fuel cell stack anomaly detection and lifespan prediction model. The edge-based model is a lightweight INT8-quantized version (approximately 3MB) used for real-time inference on a single vehicle; the cloud-based model is a full-precision version, retrained based on historical data from the entire fleet, and replaced on the edge-based model via OTA updates after version iterations. The edge-based model is a lightweight real-time model for a single fuel cell stack component; the cloud-based model covers multiple components such as the fuel cell stack, hydrogen storage tank, motor, and tires. The two models form a 'cloud-based training - edge-based inference' relationship through OTA updates.
[0320] The cloud-based energy consumption benchmark model corresponds to the edge-based energy consumption prediction model. The edge outputs real-time hydrogen consumption prediction and energy consumption deviation for each vehicle; the cloud establishes theoretically optimal hydrogen consumption benchmarks for each vehicle type / route / operating condition based on data from the entire fleet, for comparative analysis.
[0321] The cloud-based driving behavior group profile model corresponds to the edge-based driving behavior recognition model. The edge-based model identifies the driving style of an individual vehicle in a single instance (aggressive / moderate / economical); the cloud-based model aggregates data from the entire fleet to form a group profile (typical driving behavior distribution along the same route at the same time), which is used to establish benchmarks and identify anomalies.
[0322] Regarding the hydrogen storage safety monitoring model: There is no corresponding version in the cloud. Hydrogen storage safety is a hard real-time safety task (<10ms). The model is pre-deployed at the edge and does not rely on cloud training and distribution. The cloud only manages it through safety rule base configuration (such as adjusting alarm thresholds, delay parameters, etc.). OTA delivers configuration files, not model files.
[0323] The lightweight deployment process after offline model training includes the following steps: structural pruning of the full-precision model in the cloud (removing redundant connections / channels); knowledge distillation (using the large cloud model as the teacher network to train small edge models); INT8 quantization (converting weights / activations from FP32 to INT8); operator fusion and format conversion (generating model files adapted to OWA5X NPU); OTA deployment after integrity verification; and backup of the old model at the edge for loading on the next startup.
[0324] Specifically, the lightweight deployment process from the energy consumption baseline model to the energy consumption prediction model involves: the cloud-based energy consumption baseline model learns the optimal model, establishing the theoretically optimal hydrogen consumption baseline for each vehicle type / route / operating condition based on historical data from the entire fleet; when deployed to the edge, it is distilled / compressed into a lightweight Transformer prediction model, shifting its role to real-time prediction of "actual" consumption, outputting the expected hydrogen consumption curve for the next 5 minutes. The two work together: the edge compares the "predicted actual hydrogen consumption" with the "ideal baseline" deployed from the cloud to calculate the energy consumption deviation E_dev. In other words, "the cloud learns the baseline, the edge makes predictions, and the deviation is obtained through comparison."
[0325] Specifically, the lightweight deployment process from the driving behavior group profiling model to the driving behavior recognition model involves: The cloud-based group profiling model acts as a cluster model, clustering the entire fleet's driving data to obtain the typical driving behavior distribution and classification boundaries (group baseline) for each route / time period. When deployed to the edge, the cluster centers and classification boundaries of the group profiling are used as the calibration basis / soft label for the edge-based CNN-LSTM classifier, compressed into a real-time style classifier for each vehicle, outputting "aggressive / moderate / economical" and confidence levels. In other words, "the cloud sets the group baseline, the edge performs individual vehicle recognition, and the scoring aligns with the group."
[0326] The large model decision sublayer includes: a fleet size optimization engine (based on genetic algorithms and reinforcement learning, taking historical order data, vehicle utilization, maintenance cycle, and operating cost structure as input, and outputting optimal fleet size suggestions and vehicle configuration ratios), a route network planning engine (combining multi-dimensional data such as hydrogen refueling station distribution, real-time traffic, weather forecasts, and road restrictions to perform multi-objective path optimization), a hydrogen refueling optimization engine (taking the vehicle's current health status, hydrogen reserves, hydrogen refueling station queuing status, and hydrogen price as input, and outputting optimal hydrogen refueling timing, station, and refueling quantity suggestions), and a predictive maintenance decision engine (combining three inputs: time-series denoising and trend inflection point detection of the fuel cell stack and various components uploaded from the edge, a group comparison anomaly report generated by the cloud-based basic AI sublayer, and the vehicle's historical maintenance records and spare parts inventory status, to predict the failure time window of each component; when the predicted failure time is within a preset threshold, a maintenance work order is automatically generated and pushed to the management app).
[0327] The core functions of this layer are: to aggregate the operational data of the entire fleet for macro-analysis and model training; to output fleet-level business decisions based on global data and multi-dimensional constraints through operations research optimization and large model inference; and to distribute the updated AI model parameters to the edge.
[0328] (4) Mobile collaboration layer: includes two differentiated interfaces, the management app and the driver app, which obtain real-time data and decision results through the Corvina cloud API.
[0329] The management app is designed for fleet administrators and dispatchers, providing functions such as fleet map overview, alarm message center, task assignment and tracking, data reports (energy consumption statistics, cost analysis, driving behavior score ranking), and maintenance management (maintenance reminders, repair work orders, spare parts inventory).
[0330] The driver-facing app is designed for drivers and provides functions such as a task center (today's tasks, one-click navigation), vehicle status (hydrogen reserve, health status), driving assistant (real-time driving score and energy-saving suggestions, navigation to nearby hydrogen refueling stations), and hydrogen refueling services (hydrogen refueling station queue estimation, hydrogen refueling records and expense reimbursement).
[0331] The two apps share the same cloud data source, but differentiated data views and function access are achieved through role-based access control. The core function of this layer is to push cloud-based AI decision-making results to end users with different roles in a visualized and actionable format in real time, completing the final execution loop of operational decisions. Specifically: The data flow relationship between the above four layers is as follows: The vehicle data perception layer generates raw sensor data → it is aggregated to the edge AI inference layer for local preprocessing and real-time inference → feature data and alarm events are uploaded to the cloud AI decision layer through the data flow bus → the cloud AI decision layer performs model training and decision optimization based on global data, and sends the decision results to the mobile collaboration layer for user execution → at the same time, the updated model parameters are sent back to the edge AI inference layer through the model update bus to complete the continuous evolution loop of the model.
[0332] In one specific embodiment, the role of the group comparison anomaly report of the cloud-based basic AI sublayer in the decision-making stage is as follows.
[0333] The group comparison anomaly report is a key preliminary analysis step in the predictive maintenance decision engine. Its core function is to compare the component health indicators of a single vehicle with the component health indicator distribution of the entire fleet of the same model and mileage segment, and identify "abnormal individuals" that significantly deviate from the group mean or normal distribution. This is a "group reference" mechanism that differs from the traditional single-vehicle threshold judgment.
[0334] For example, a vehicle's fuel cell stack RUL is estimated at 3000 hours, which might be considered "normal" if judged solely on a single vehicle basis. However, if the median RUL of the entire fleet for the same mileage segment is 5000 hours, and this vehicle's RUL is at the P5 quantile (lower than 95% of similar vehicles), then the group comparison anomaly report should mark this vehicle as having a "fuel cell stack RUL deviating from the normal distribution of the group," indicating a potential problem that may not have been detected by the single-vehicle monitoring model (such as accelerated aging caused by long-term operation under extreme conditions). Conversely, if a vehicle's indicator slightly exceeds the single-vehicle threshold but is still within the normal distribution range of the group, group comparison can reduce the false alarm rate.
[0335] The group comparison anomaly report has two uses: Use 1 (internal consumption): as input to the predictive maintenance decision engine—when the report marks a vehicle or component as having a "group anomaly deviation," the predictive maintenance decision engine automatically generates a preventative inspection work order for that vehicle or component, even if the component's individual vehicle health indicators have not yet triggered the regular alarm threshold; Use 2 (external display): pushed to the management app (fleet administrator) to display the data in the form of visual comparison charts (such as the percentile position of each vehicle's health indicator in the fleet distribution), for administrator manual analysis. It is not pushed to the driver's app because group comparison information is a macro-management tool for the fleet management team; drivers only need to focus on their own vehicle's real-time individual indicators.
[0336] The working principle of the edge-cloud collaborative hydrogen fuel cell vehicle full life cycle operation and maintenance decision-making system of this application is as follows.
[0337] Figure 4 Here is a data link diagram for one embodiment. The core working principle and steps of the system are as follows: Figure 4 As shown. The system as a whole follows a five-stage closed-loop workflow of "perception-reasoning-decision-execution-evolution". Among them: (1) Perception Phase (Execution by Onboard Data Perception Layer): After the hydrogen fuel cell vehicle starts, the OWA5X edge computing gateway enters the working state and reads operating parameters such as accelerator pedal opening, brake pedal status, vehicle speed, motor output power, voltage of each section of the fuel cell stack, stack inlet and outlet temperature / pressure, instantaneous hydrogen consumption rate, hydrogen storage tank pressure / temperature, and ambient temperature from the onboard electronic control units such as VCU, BMS, FCU, and HMS via the CAN bus at a fixed sampling frequency (typically 100Hz). At the same time, it reads the values of the installed hydrogen concentration sensor through an independent analog interface. All raw data is cached after being timestamped uniformly inside the edge gateway.
[0338] (2) Inference phase (execution of edge AI inference layer): The edge inference engine retrieves the latest batch of time-aligned sensor data from the cache, first performs data preprocessing (outlier removal, missing value interpolation, normalization), and then distributes the preprocessed data to the four AI inference channels simultaneously: First, the hydrogen storage safety monitoring channel (safety channel, highest priority) receives hydrogen concentration and hydrogen storage cylinder pressure / temperature data. It calculates the reconstruction error through an autoencoder. If the reconstruction error exceeds the preset dynamic threshold (based on the 3σ range of historical normal data distribution), a local audible and visual alarm is triggered within 10 milliseconds, and a safety alarm event data packet is generated to prepare for uploading to the cloud.
[0339] Second, the fuel cell stack anomaly detection and lifetime prediction channel (health channel, high priority) receives voltage data from each section of the fuel cell stack, uses the isolated forest algorithm to determine whether the current operating status falls into an anomaly cluster, calculates the overall health deviation of the fuel cell stack through an autoencoder, and updates the estimated remaining lifetime of the fuel cell stack.
[0340] Third, the energy consumption prediction channel (energy consumption channel, medium priority) receives operating condition data such as current vehicle speed, load, and hydrogen consumption rate, predicts the expected hydrogen consumption curve in the next 5 minutes, and compares it with the ideal hydrogen consumption curve to calculate the energy consumption deviation.
[0341] Fourth, the driving behavior recognition channel (behavior channel, normal priority) receives driving operation data such as accelerator pedal, brake pedal, and vehicle speed change rate. It identifies the current driving style (aggressive / mild / economical) through a CNN-LSTM fusion model and generates a driving behavior score and personalized energy-saving suggestions by combining energy consumption deviation.
[0342] After all inference results are aggregated at the edge, feature data and alarm events are packaged and compressed, and then uploaded to the Corvina cloud platform via 4G / 5G network.
[0343] (3) Decision-making stage (executed by cloud AI decision-making layer): The cloud data receiving service extracts the feature data packets uploaded by the edge terminal from the message queue, parses them and writes them into the time series database. The cloud basic AI sub-layer regularly (default early morning every day) executes offline batch analysis tasks, including: summarizing the driving behavior data of the entire fleet from the previous day, updating the driving behavior group profile and energy consumption benchmark model; retraining the fuel cell life prediction model based on the historical operating data and fault records of the same type of fuel cell stacks in the entire fleet; analyzing the time series trends of the health indicators of each vehicle component and generating a group comparison anomaly report.
[0344] When the cloud-based big data model decision sublayer receives a triggering event (such as a dispatcher initiating a route planning request, the system detecting a concentrated surge in hydrogen refueling demand from multiple vehicles, or the maintenance engine predicting that a vehicle component will reach its replacement threshold within 72 hours), it initiates the corresponding optimization solution process: First, the solution process for fleet size optimization is as follows: the fleet size optimization engine runs a genetic algorithm to solve for the optimal vehicle configuration scheme under the constraint of demand prediction; Second, the solution process for route network planning is as follows: The route network planning engine runs a multi-objective A... Search or ant colony algorithm to find Pareto optimal path under the three-dimensional objectives of shortest time / lowest hydrogen consumption / optimal safety; Third, the hydrogen refueling optimization solution process is as follows: The hydrogen refueling optimization engine takes the vehicle health state vector, hydrogen reserve vector, hydrogen refueling station state vector and hydrogen price vector as inputs, and runs a mixed integer programming model to solve the optimal hydrogen allocation scheme. Fourth, the predictive maintenance solution process is as follows: The predictive maintenance decision engine automatically generates maintenance work orders based on the predicted component lifespan, including suggested maintenance time windows, a list of required spare parts, and estimated working hours.
[0345] (4) Execution Phase (Mobile Collaboration Layer Execution): The decision results generated in the cloud are sent to the corresponding mobile app through a real-time push channel. When the hydrogen refueling optimization engine outputs a hydrogen refueling suggestion, the driver's app receives a push notification, displaying the recommended hydrogen refueling station, estimated queuing time, and recommended hydrogen refueling amount. The driver can start navigation with one click. When the predictive maintenance engine generates a maintenance work order, the management app receives the work order notification. The dispatcher can view the work order details, confirm the maintenance time, and assign maintenance personnel. When the driving behavior recognition model produces a real-time score, the driver's app's driving assistant interface updates the score and energy-saving prompts synchronously. After execution is completed, the execution results (such as hydrogen refueling completion time, actual hydrogen refueling amount, and maintenance completion status) are sent back to the cloud through the app for updating vehicle status and calibrating the decision model.
[0346] (5) Evolutionary Stage (Closed Loop of Model Update from Cloud to Edge): After each model training, the cloud-based basic AI sublayer performs INT8 quantization and format conversion on the updated model parameters (including the retrained fuel cell lifetime prediction model, the calibrated energy consumption baseline model, and the optimized driving style classifier weights) to generate a model file suitable for OWA5X edge gateway NPU inference. The model file is distributed to designated vehicles or the entire fleet via OTA (Over-the-Air) channel. The edge gateway receives and verifies the integrity of the model file when the vehicle is off or the network is idle. After verification, the old model is replaced and the new model parameters are loaded on the next startup.
[0347] Meanwhile, the "difficult case data" identified and cached by the edge gateway during inference (such as inference results with model confidence below the threshold, and abnormal operating condition data with deviations from cloud predictions exceeding the threshold) are uploaded to the cloud along with the regular data stream and added to the incremental dataset for the next model training. This completes a full evolution cycle of "cloud training → edge inference → difficult case backhaul → cloud optimization → model deployment," ensuring that the system's AI capabilities continue to improve throughout its lifecycle.
[0348] In the process of edge gateways identifying and caching "difficult case data" during inference, different models employ different confidence measurement methods. The hydrogen storage safety monitoring model uses: Confidence = 1 - Reconstruction Error / Threshold; the smaller the reconstruction error, the higher the confidence. The stack anomaly detection model uses: Confidence = 1 - Normalized Anomaly Score; the lower the anomaly score (closer to the center of the normal cluster), the higher the confidence. The driving behavior recognition model uses: Confidence = Maximum Class Probability Value Output by Softmax; the higher the probability, the more certain the model is about the classification result. The energy consumption prediction model uses: Confidence = Normalized Complement of the Information Entropy of the Prediction Distribution; the lower the entropy, the more concentrated / certain the prediction.
[0349] Among them, the deviation from the cloud prediction refers to the difference between the edge inference result and the prediction result of the cloud model (or cloud service interface) on the same input sample. It can also be changed to compare the edge prediction with the subsequent actual observation value (such as actual hydrogen consumption, actual remaining life, etc.). When the deviation exceeds the threshold, it is also marked as a difficult case.
[0350] In addition, the edge gateway model undergoes offline evaluation in the cloud. The process is as follows: Before deploying the model at the edge, the cloud runs a full-precision model using historical feature data uploaded from the edge, and performs offline comparison and verification with the output of the edge's INT8 model to ensure that the quantization accuracy loss is within an acceptable range; after verification, it is deployed via OTA. If a real-time inference interface is configured in the cloud during edge operation, the deviation from the real-time inference result in the cloud is continuously calculated. When the deviation exceeds a threshold, it is marked as a difficult example for the next round of incremental training.
[0351] In another embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, it includes a processor, memory, input / output interfaces, and a communication interface. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface is connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores relevant data. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. The computer program can be executed by the processor to implement the various methods described in the above embodiments.
[0352] In yet another embodiment, a computer device is provided, such as a terminal, whose internal structure diagram may be as follows: Figure 6As shown, it includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. The computer program can be executed by the processor to implement the various methods described in the above embodiments.
[0353] Those skilled in the art will understand that Figure 5 and Figure 6 The structure shown is only a block diagram of a part of the structure related to the present application and does not constitute a limitation on the computer device on which the present application is applied. It may also include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements, in order to realize the function of the computer device.
[0354] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0355] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the systems, devices, equipment, modules or units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0356] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, devices, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.
[0357] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0358] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.
[0359] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.
[0360] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, optical fiber) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium (e.g., a solid-state drive), etc.
[0361] The technical solutions provided by the embodiments of this application have been described in detail above. Specific examples have been used in the embodiments of this application to illustrate the principles and implementation methods of the embodiments of this application. The description of the above embodiments is only for the purpose of helping to understand the methods and core ideas of the embodiments of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the embodiments of this application. Therefore, the content of this specification should not be construed as a limitation on the embodiments of this application.
Claims
1. A method for full lifecycle operation and maintenance decision-making for hydrogen fuel cell vehicles based on edge-cloud collaboration, characterized in that, The method includes: Based on the data perception layer device deployed on hydrogen fuel cell vehicles, multi-source heterogeneous vehicle perception data is acquired, and the vehicle perception data is aggregated to the edge computing gateway at the edge. The vehicle perception data of the edge computing gateway is preprocessed and feature extracted to obtain feature-extracted data; priority information of each operation and maintenance decision task is calculated based on a priority preemptive task scheduler; according to the task execution order represented by the priority information, the feature-extracted data is inferred in millisecond-level real time based on each pre-deployed edge inference model, and the feature data and alarm events of the inference process are summarized as edge inference result data; the edge inference result data is uploaded to the cloud. The edge inference result data uploaded by the edge terminal is received based on the message queue, and a time-series database is constructed based on the edge inference result data of each hydrogen energy vehicle; in response to the triggering event of cloud decision, cloud decision is performed on the fleet operation data of the entire fleet in the time-series database based on the pre-trained cloud decision engines, and cloud decision result data is obtained. The cloud-based decision-making results data are sent to a mobile device, and displayed on the mobile device for vehicle users associated with the hydrogen fuel cell vehicle to make decisions.
2. The method according to claim 1, characterized in that, The pre-deployed edge inference models perform millisecond-level real-time inference on the feature-extracted data, and summarize the feature data and alarm events of the inference process as edge inference result data, including: Hydrogen concentration, hydrogen storage cylinder pressure and hydrogen storage cylinder temperature data are extracted from the data after feature extraction as input data for hydrogen storage safety monitoring. Based on the hydrogen storage safety monitoring model, the hydrogen storage safety monitoring input data are inferred to obtain hydrogen storage safety monitoring output data. The voltage data of each section of the fuel cell stack is extracted from the data after feature extraction. The fuel cell stack anomaly detection and lifetime prediction model is used to infer the voltage data of each section of the fuel cell stack to obtain fuel cell stack anomaly detection and lifetime prediction data. Operating condition data is extracted from the data after feature extraction, and energy consumption prediction data is obtained by reasoning based on the energy consumption prediction model. Driving operation data is extracted from the feature-extracted data, and driving behavior recognition data is obtained by reasoning based on the driving behavior recognition model. The edge inference result data is determined by summarizing the hydrogen storage safety monitoring output data, the stack anomaly detection and lifetime prediction data, the energy consumption prediction data, and the driving behavior recognition data.
3. The method according to claim 2, characterized in that, The hydrogen storage safety monitoring output data is obtained by reasoning from the hydrogen storage safety monitoring input data based on the hydrogen storage safety monitoring model, including: The initial hydrogen storage safety monitoring model was constructed based on a fully connected multilayer autoencoder architecture. The training dataset was constructed based on historical operating data of the vehicle under normal operating conditions. Based on the training dataset, with minimizing the reconstruction error as the training objective, the mean squared error is used as the loss function to train the initial hydrogen storage safety monitoring model, resulting in the trained hydrogen storage safety monitoring model. After the model training is completed, the distribution data of the reconstruction error is calculated on the validation set of the training dataset, and a dynamic threshold is set based on the distribution data of the reconstruction error. The real-time reconstruction error is obtained by inferring from the hydrogen storage safety monitoring input data based on the trained hydrogen storage safety monitoring model. Anomaly determination is performed based on the real-time reconstruction error and the dynamic threshold to obtain the anomaly determination result; The hydrogen storage safety monitoring input data, the reconstruction error distribution data, the dynamic threshold, and the anomaly determination result recorded during the reasoning process are used as the feature data. The alarm event is generated based on the anomaly determination result to determine the hydrogen storage safety monitoring output data.
4. The method according to claim 2, characterized in that, The method of reasoning about the voltage data of each section of the fuel cell stack based on the fuel cell stack anomaly detection and lifetime prediction model to obtain fuel cell stack anomaly detection and lifetime prediction data includes: An isolated forest model is constructed using the isolated forest algorithm. The voltage data of each section of the fuel cell stack is input into the isolated forest model to obtain the abnormal cluster judgment result. The stack reconstruction error is output through an autoencoder; Based on the abnormal cluster judgment results and the fuel cell reconfiguration error, the overall health deviation of the fuel cell stack is calculated. Update the remaining useful life estimate of the fuel cell stack based on the health deviation; The voltage data of each section of the fuel cell stack, the abnormal cluster judgment result, the fuel cell stack reconstruction error, the health deviation, and the updated estimated remaining lifespan of the fuel cell stack are recorded as the feature data. The alarm event is generated based on the abnormal cluster judgment result, the health deviation, and the updated estimated remaining lifespan of the fuel cell stack to determine the fuel cell stack anomaly detection and lifespan prediction data.
5. The method according to claim 2, characterized in that, The process of reasoning from the operating condition data based on the energy consumption prediction model to obtain energy consumption prediction data includes: Based on the lightweight converter prediction model, the operating data is predicted to obtain the expected hydrogen consumption rate sequence within a future preset time period. Based on the physical model of the hydrogen fuel cell system, and with the optimal stack operating point and optimal driving strategy as constraints, the theoretical minimum hydrogen consumption curve is obtained by solving. The energy consumption deviation is calculated based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve. The operating condition dimension with the largest deviation from the theoretical minimum hydrogen consumption curve and the corresponding maximum deviation value are determined based on the energy consumption deviation. The operating condition data, the expected hydrogen consumption rate sequence, the theoretical minimum hydrogen consumption curve, the operating condition dimension with the largest deviation, and the maximum deviation value are recorded as the feature data. The alarm event is generated based on the relationship between the energy consumption deviation degree and the preset energy consumption deviation alarm threshold to determine the energy consumption prediction data. The calculation of energy consumption deviation based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve includes: Based on the expected hydrogen consumption rate sequence and the theoretical minimum hydrogen consumption curve, a first integral operation is performed within the future preset time period to obtain the total difference between the predicted hydrogen consumption and the ideal hydrogen consumption. Based on the theoretical minimum hydrogen consumption curve, a second integral operation is performed within the future preset time period to obtain the ideal total hydrogen consumption. The ratio of the total difference to the total ideal hydrogen consumption is taken as the energy consumption deviation.
6. The method according to claim 2, characterized in that, The process of reasoning from the driving operation data based on the driving behavior recognition model to obtain driving behavior recognition data includes: A driving operation feature vector is constructed based on the driving operation data; The driving operation feature vector is input into the trained fusion model, and the current driving style classification result and the corresponding driving style confidence score are output. The driving behavior score is obtained by weighting and summing the scores of each item according to their respective weights. Personalized energy-saving suggestions are generated based on the current driving style classification results; Based on the current driving style classification result, the driving style confidence score, the driving behavior score, the personalized energy-saving suggestion, and the driving operation data, the feature data is generated, and the alarm event is generated based on the driving behavior score to determine the driving behavior identification data; Based on the current driving style classification result, the driving style confidence score, the driving behavior score, the personalized energy-saving suggestion, and the driving operation data, the feature data is generated, including: Determine the suggestion type and trigger time associated with the personalized energy-saving suggestion; The driving operation data is subjected to cumulative driving behavior statistical analysis to obtain driving behavior statistics. The feature data is generated based on the current driving style classification result, the driving style confidence level, the driving behavior score, the suggestion type, the trigger time, and the driving behavior statistics.
7. The method according to claim 1, characterized in that, The pre-trained cloud-based decision engines perform cloud-based decision-making on the fleet operation data in the time-series database to obtain cloud-based decision result data, including: Historical order data, vehicle utilization rate, maintenance cycle, and operating cost structure are extracted from the entire fleet operation data as input data for fleet size optimization; the fleet size optimization engine solves the fleet size optimization input data to obtain the vehicle configuration scheme; The distribution of hydrogen refueling stations, real-time traffic, weather forecasts, and road restrictions are extracted from the entire fleet's operational data as input data for route network planning. The route network planning engine is used to solve the route network planning input data to obtain a route network planning scheme. The current health status of vehicles, hydrogen reserves, hydrogen refueling station queue status, and hydrogen price are extracted from the entire fleet's operational data as input data for hydrogen refueling optimization. The hydrogen refueling optimization engine solves the hydrogen refueling optimization input data to obtain a hydrogen allocation scheme. The fuel cell stack and individual component attenuation trends, group comparison anomaly reports, historical maintenance records, and spare parts inventory status are extracted from the entire fleet's operational data as input data for predictive maintenance decisions. The predictive maintenance decision engine then solves the predictive maintenance decision input data to obtain maintenance work orders. The cloud-based decision-making results data are determined based on the vehicle configuration scheme, the route network planning scheme, the hydrogen refueling allocation scheme, and the maintenance work order.
8. The method according to claim 7, characterized in that, The process of solving the fleet size optimization input data using the fleet size optimization engine to obtain the vehicle configuration scheme includes: With the objective function of minimizing total operating costs and maximizing revenue, and with constraints such as total capacity, vehicle utilization rate, maintenance queue limit, purchase cost, and number of vehicle types, a fleet size optimization model is constructed. The vehicle configuration scheme is obtained by solving the fleet size optimization model using a fleet size optimization genetic algorithm. The process of using a genetic algorithm for fleet size optimization to solve the fleet size optimization model and obtain the vehicle configuration scheme includes: Using the number of vehicles of each model as the decision variable, and employing real-number vector encoding, an initial population is randomly generated within the constraints. For each individual in the initial population, the objective function is calculated, and the penalty function method is used to multiply the individuals that violate the constraints by a penalty coefficient and include them in the fitness. A predetermined number of individuals are randomly selected from the initial population, and those with the best fitness are selected to enter the mating pool until a new generation of parent parents is generated. Based on the new generation parent set, simulated binary crossover is performed to generate the crossover offspring; Apply a polynomial mutation perturbation to the gene loci of the offspring to generate perturbed offspring; Dynamic parameter tuning is performed at each preset algebra iteration. Repair the crossover offspring and the perturbation offspring that exceed the boundary, and retain the historical best individual; The iteration stops when the maximum number of generations or the convergence condition is reached, and the vehicle configuration scheme is output based on the historical best individual.
9. The method according to claim 7, characterized in that, The route network planning engine solves the input data for route network planning to obtain a route network planning scheme, including: Retrieve externally loaded data; Three-dimensional cost data is calculated for each candidate path based on the externally loaded data; Perform a multi-objective search based on the aforementioned three-dimensional cost data to determine the set of non-dominated paths; Dominant path elimination is performed on the non-dominated path set after the search is completed, and the Pareto optimal path set is output. The target path is determined from the Pareto optimal path set, corresponding to the dispatcher's selection operation or the recommended operation based on operational preferences. Generate the route network planning scheme based on the target path; The acquisition of externally loaded data includes: Road network topology data and hydrogen refueling station distribution data are obtained from static map data that is persisted locally in the cloud. Real-time traffic data is retrieved from a third-party traffic data service interface. Obtain weather forecast data by connecting to the cloud-based meteorological data service interface; Obtain road restriction data from the map service provider's interface; The external loading data is generated based on the road network topology data, the hydrogen refueling station distribution data, the real-time traffic data, the weather forecast data, and the road restriction data.
10. The method according to claim 7, characterized in that, The process of solving the hydrogen refueling optimization input data using a hydrogen refueling optimization engine to obtain a hydrogen allocation scheme includes: Using vehicle station allocation, hydrogen refueling volume, and arrival time as decision variables, an objective function is constructed based on hydrogen refueling cost, detour cost, waiting cost, and unallocated penalty. A mixed integer programming model is constructed with constraints such as hydrogen surplus, station allocation, capacity limits of hydrogen refueling stations for each time period, health status limits, and inventory limits of hydrogen refueling stations. The mixed-integer programming model is solved using a mixed-integer algorithm to obtain the hydrogenation allocation scheme; The mixed-integer programming model is solved using a mixed-integer solution algorithm, including: If the number of vehicles to be dispatched or the size of the preset variables exceeds the preset threshold, or if the branch and bound cannot converge within the preset time limit, the solution size is determined as the target size, and heuristic solution steps are executed to solve the problem.
11. The method according to claim 7, characterized in that, The extraction of the degradation trend of the fuel cell stack and its components includes: The original sequence of remaining lifespan was extracted from the entire fleet's operational data; The original remaining lifetime sequence is denoised and smoothed using sliding window filtering to obtain a smoothed remaining lifetime trend curve. Based on the smoothed remaining lifetime trend curve, trend inflection point detection is performed to obtain the degradation trend of the fuel cell stack and its components.
12. A decision-making system for the entire lifecycle operation and maintenance of hydrogen-powered vehicles based on edge-cloud collaboration, characterized in that, The system includes: The vehicle data sensing module is used to acquire multi-source heterogeneous vehicle sensing data based on the data sensing layer device deployed on hydrogen energy vehicles, and to aggregate the vehicle sensing data to the edge computing gateway at the edge. The edge inference module is used to preprocess and extract features from the vehicle perception data of the edge computing gateway to obtain feature-extracted data; calculate the priority information of each operation and maintenance decision task based on the priority preemptive task scheduler; perform millisecond-level real-time inference on the feature-extracted data according to the task execution order represented by the priority information and based on the pre-deployed edge inference models; summarize the feature data and alarm events of the inference process as edge inference result data; and upload the edge inference result data to the cloud. The cloud decision module is used to receive edge inference result data uploaded from the edge terminal based on the message queue, and construct a time series database based on the edge inference result data of each hydrogen energy vehicle; in response to the triggering event of cloud decision, it performs cloud decision on the fleet operation data of the entire fleet in the time series database based on the pre-trained cloud decision engines to obtain cloud decision result data. The mobile collaboration module is used to send the cloud decision result data to the mobile terminal and display the cloud decision result data on the mobile terminal for vehicle users associated with the hydrogen energy vehicle to make decisions.
13. A computer device, characterized in that, The computer device includes: At least one processor and memory; The memory is used to store program code, and the processor is used to call the program code stored in the memory to execute the method as described in any one of claims 1 to 11.
14. A computer storage medium, characterized in that, It includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 11.
