Fuel cell hydrogen-oxygen cross predictive diagnosis method and system

Abstract

The invention discloses a predictive diagnosis method and system for hydrogen and oxygen intercrossing of a fuel cell, and relates to the technical field of fuel cells, and the method comprises the steps of parameter initialization, data real-time monitoring, fault prediction diagnosis, early warning grade analysis decision and model parameter feedback learning, performing fusion processing on multi-source data obtained through real-time monitoring, updating a probability prediction result of hydrogen-oxygen cross fault occurrence based on fused data features, and regularly re-evaluating the remaining service life of the galvanic pile so as to dynamically control the fault risk and the service life state; and the early warning grade analysis decision determines a corresponding early warning grade according to the probability prediction result obtained in the prediction diagnosis stage and the residual service life result of the electric pile, and generates maintenance suggestions and processing measures according to the early warning grade. The problems that in the prior art, hydrogen-oxygen inter-string fault early warning time is short, multi-source data fusion is difficult, fault positioning precision is low, predictive maintenance capacity is lacked, and false alarm and missing alarm are likely to happen under the dynamic working condition are solved.

Description

Technical Field

[0001] This invention relates to the field of fuel cell technology, and specifically to a method and system for predictive diagnosis of hydrogen-oxygen crosstalk in fuel cells. Background Technology

[0002] With the rapid development of the hydrogen energy industry, the commercialization of proton exchange membrane fuel cells in fields such as vehicles and energy storage is accelerating. However, hydrogen-oxygen crosstalk has become a core safety hazard. When the proton exchange membrane is microscopically damaged, oxygen from the cathode can permeate to the anode and form an explosive mixture with hydrogen, which may cause catastrophic accidents. Therefore, efficient monitoring and early warning technologies are urgently needed.

[0003] Current technologies suffer from several key shortcomings, making it difficult to meet practical needs. Voltage monitoring (CVM) methods can only detect voltage anomalies after a fault occurs, resulting in short warning times; electrochemical impedance spectroscopy (EIS) scans at low frequencies, failing to capture rapid fault processes; anode exhaust oxygen analysis is limited by sensor response delays and accuracy, leading to poor early warning effects. Existing systems largely rely on fixed empirical threshold alarms, which cannot adapt to the dynamic operating conditions of fuel cells, and single parameters are insufficient to distinguish similar fault symptoms, easily resulting in false alarms and missed alarms. Current digital twin technologies in the fuel cell field primarily focus on performance optimization and control strategy verification, lacking deep integration with multi-source real-time data, thus failing to achieve fault prediction. Different sensor data vary significantly in temporal scale (millisecond-level CVM to minute-level EIS) and spatial resolution (single cell level to stack level). Existing fusion technologies are mostly designed for lithium battery systems and lack physical mechanism support, making effective data integration difficult. Current technologies largely remain at the level of fault detection and diagnosis, unable to achieve predictive maintenance; data-driven models have poor generalization ability and high false alarm rates when training data is insufficient. Summary of the Invention

[0004] To address these issues, the present invention provides a predictive diagnostic method and system for hydrogen-oxygen inter-connection in fuel cells, which solves the problems of short early warning time, difficulty in multi-source data fusion, low fault location accuracy, lack of predictive maintenance capability, and inability to adapt to dynamic operating conditions, which can easily lead to false alarms and missed alarms.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predictive diagnostics of hydrogen-oxygen crosstalk in a fuel cell, comprising the following steps:

[0006] (1) Initialization: Load the fuel cell stack design parameters to build and initialize the digital twin model, calibrate and benchmark the sensors for data acquisition to establish the stack health status baseline, pre-train the AI ​​model for fault analysis and load the historical fault case library;

[0007] (2) Real-time monitoring: Synchronously collect multi-source sensor data and perform time dimension alignment processing, perform quality assessment on the collected data to remove abnormal measurement point data, and update the state parameters of the digital twin model in real time based on the processed valid data so that the digital twin model is consistent with the state of the fuel cell stack.

[0008] (3) Predictive diagnosis: Regularly perform fusion processing on the multi-source data obtained from real-time monitoring, update the probability prediction results of hydrogen-oxygen inter-connection faults based on the characteristics of the fused data, and regularly reassess the remaining service life of the stack in order to dynamically control the fault risk and life status.

[0009] (4) Early warning decision: Based on the probability prediction results and the remaining service life of the fuel cell stack obtained in the prediction and diagnosis stage, the corresponding early warning level is determined. Maintenance suggestions and handling measures are generated according to the early warning level, and instructions are sent to the fuel cell system through the control interface to execute equipment adjustment operations.

[0010] (5) Feedback learning: Collect the implementation results of maintenance measures and fault confirmation information in the early warning decision-making stage, update the AI ​​model parameters and data feature weights based on the implementation results of the maintenance measures and the fault confirmation information, optimize the early warning threshold and diagnostic rules, and realize the iterative improvement of the diagnostic capabilities of the AI ​​model.

[0011] As a preferred method for predictive diagnostics of hydrogen-oxygen inter-flow in fuel cells, the digital twin model is constructed by employing intrinsic orthogonal decomposition (POD), real-time parameter identification, and surrogate model construction techniques to achieve model order reduction and real-time processing.

[0012] 500 dominant modes were extracted through intrinsic orthogonal decomposition, increasing the computational degrees of freedom of the model from 10. 6 Dropped to 500;

[0013] Real-time parameter identification uses an unscented Kalman filter algorithm to update model parameters, with a parameter update period of 1 second, so that the model parameters match the actual state of the fuel cell stack.

[0014] The surrogate model is constructed based on the Kriging method to build a real-time response surface as a surrogate model. The Kriging method describes the spatial correlation of data by constructing a variogram function, thereby achieving accurate prediction of data at unmeasured points.

[0015] The digital twin model integrates multiphysics governing equations, including electrochemical reaction equations, proton conduction equations, gas diffusion equations, and energy conservation equations. Specified parameters from these multiphysics governing equations are used as model input variables. Specifically, the specified parameter for the electrochemical reaction equation is the exchange current density. Transmission coefficient limiting current The proton conduction equation specifies the water diffusion coefficient as a parameter. Electroosmotic drag coefficient The gas diffusion equation specifies the binary diffusion coefficient as the parameter. Component mole fraction The energy conservation equation specifies the thermal conductivity as the parameter. Heat of reaction .

[0016] As a preferred method for predictive diagnostics of hydrogen-oxygen crosstalk in fuel cells, in step (2), the multi-source sensor data includes electrical performance data, gas concentration data and thermal data; the electrical performance data is collected by an ultra-high-speed CVM system and a multi-band EIS system, the gas concentration data is collected by a laser oxygen analyzer, and the thermal data is collected by a distributed temperature and pressure sensing network.

[0017] Data quality assessment includes data integrity checks, data consistency checks, and outlier detection. Data integrity checks confirm that there are no missing data, data consistency checks verify the logical matching between data from different sensors, and outlier detection uses the 3σ criterion to identify and remove data that exceeds the normal range.

[0018] As a preferred method for predictive diagnostics of hydrogen-oxygen crosstalk in fuel cells, the multi-source data fusion in step (3) adopts a multi-level fusion strategy, including the DS evidence theory fusion algorithm, the attention mechanism neural network and the spatiotemporal alignment algorithm.

[0019] The implementation methods of the DS evidence theory fusion algorithm include:

[0020] The support function is calculated based on the distance between data feature values ​​and rank feature values, and a basic probability value is assigned to each sensor data.

[0021] Dempster's combination rule is used to process multi-source evidence. The combination formula is as follows:

[0022]

[0023] In the formula, Assign values ​​to the basic probability of proposition A after the evidence is synthesized. , The basic probability assignments for propositions B and C are given by the two sources of evidence, respectively.

[0024] Based on the synthesized trust function and likelihood function, a quantitative assessment of the failure hypothesis is achieved;

[0025] In the spatiotemporal alignment algorithm, time registration uses sliding window averaging and linear interpolation to unify the time reference, and spatial registration is based on the three-dimensional mesh model of the electric stack, and the spatial continuity mapping of discrete measurement points is achieved through radial basis functions.

[0026] As a preferred scheme for predictive diagnostics of hydrogen-oxygen interconnection in fuel cells, in step (3), the predictive diagnostics employs a physical information neural network, and the loss function of the physical information neural network is:

[0027]

[0028]

[0029]

[0030]

[0031] In the formula, For the data fitting term, the mean squared error is used for calculation. These are the model's predicted values. These are actual measured values; For physical constraint terms, The model's predicted value for the physical quantity. Equations describing the physical processes of fuel cells include the Butler-Volmer equation and Fick's law; The term represents a monotonicity constraint, and SOH represents the stack health state. For time, It is a linear rectified function; , These are the weighting coefficients.

[0032] As a preferred method for predictive diagnostics of hydrogen-oxygen crosstalk in fuel cells, in step (4), the warning level is divided into four levels, including Level 1 Attention Level, Level 2 Warning Level, Level 3 Danger Level and Level 4 Emergency Level.

[0033] The Level 1 trigger condition is a single abnormal data feature, a hydrogen-oxygen crosstalk fault probability of <10%, and a remaining lifespan of the fuel cell stack >500h. The system response is to log, increase the monitoring frequency to twice the original frequency, and perform trend tracking. The expected processing time is >24 hours.

[0034] Level 2 triggering conditions are multiple data feature anomalies, a failure probability of 10%-30%, and a remaining lifespan of 200-500 hours for the fuel cell stack. The system response is to issue an early warning, appropriately reduce the rated load, and prepare for maintenance. The expected processing time is 8-24 hours.

[0035] Level 3 triggering conditions are: the digital twin model predicts an anomaly, the probability of failure is 30%-60%, and the remaining service life of the fuel cell stack is 50-200 hours. The system response is to issue an audible and visual alarm, significantly reduce the rating, and schedule emergency maintenance. The expected processing time is 1-8 hours.

[0036] Level 4 trigger conditions are: multiple data sources jointly confirm the anomaly, the probability of failure is >60%, and the remaining service life of the fuel cell stack is <50 hours. The system response is to start the emergency shutdown procedure, activate the explosion suppression system, and require immediate on-site handling. The expected handling time is <1 hour.

[0037] As a preferred method for predictive diagnostics of hydrogen-oxygen inter-flow in fuel cells, the time-dimensional alignment process employs differentiated processing methods for different data types: CVM voltage inspection data is averaged using a 100ms sliding window; temperature and pressure data are mapped to a unified timestamp through linear interpolation; and EIS electrochemical impedance spectroscopy data are first feature-extracted and then fused at the decision layer. The extracted EIS features include high-frequency impedance. Charge transfer resistance Mass transfer impedance These correspond to the gas diffusion state, membrane hydration state, and catalyst activity, respectively.

[0038] As a preferred method for predictive diagnostics of hydrogen-oxygen inter-passive transmission in fuel cells, the predictive diagnostic process employs an LSTM sequence prediction model and a Cox proportional hazards model. The LSTM sequence prediction model includes an LSTM layer, an attention mechanism layer, and a regressor. The LSTM layer encodes the historical performance data sequence of the fuel cell stack to capture long-term data dependencies. The LSTM layer contains an input gate, a forget gate, and an output gate, which are used to control data input, memory update, and output, respectively. The attention mechanism layer calculates the weight of each time point in the historical sequence to predict key time point data for faults. The regressor consists of a fully connected layer and outputs the remaining lifespan of the fuel cell stack and the corresponding confidence interval.

[0039] The Cox proportional hazards model provides RUL prediction results at different confidence levels by calculating the time distribution of failure probability.

[0040] As a preferred method for predictive diagnostics of hydrogen-oxygen interconnection in fuel cells, in step (5), the online parameter update of feedback learning has no fixed cycle and is triggered in real time as maintenance results and fault confirmation information are collected. The updated content includes AI model weights, feature importance coefficients, and warning threshold ranges.

[0041] This invention also provides a fuel cell hydrogen-oxygen crosstalk predictive diagnostic system for implementing the above-mentioned fuel cell hydrogen-oxygen crosstalk predictive diagnostic method, comprising:

[0042] The initialization module is used to load fuel cell stack design parameters to build and initialize a digital twin model, calibrate and benchmark the sensors that collect data to establish a baseline for the health status of the stack, pre-train an AI model for fault analysis, and load a historical fault case library.

[0043] The real-time monitoring module is used to synchronously collect data from multiple sources of sensors and perform time-dimensional alignment processing. It performs quality assessment on the collected data to remove abnormal measurement point data and updates the state parameters of the digital twin model in real time based on the processed valid data, so that the digital twin model is consistent with the state of the fuel cell stack.

[0044] The predictive diagnostic module is used to periodically perform fusion processing on multi-source data acquired by real-time monitoring, update the probability prediction results of hydrogen-oxygen inter-connection faults based on the characteristics of the fused data, and periodically reassess the remaining service life of the fuel cell stack in order to dynamically control the fault risk and life status.

[0045] The early warning decision module is used to determine the corresponding early warning level based on the probability prediction results and the remaining service life of the fuel cell stack obtained in the prediction and diagnosis stage, generate maintenance suggestions and handling measures based on the early warning level, and issue instructions to the fuel cell system through the control interface to execute equipment adjustment operations.

[0046] The feedback learning module is used to collect the implementation results of maintenance measures and fault confirmation information during the early warning decision-making stage. Based on the implementation results of the maintenance measures and the fault confirmation information, the AI ​​model parameters and data feature weights are updated to optimize the early warning threshold and diagnostic rules, thereby achieving iterative improvement of the diagnostic capabilities of the AI ​​model.

[0047] The present invention has the following advantages:

[0048] First, the warning time has been extended from the traditional 0-5 minutes to 30 minutes to 24 hours, allowing sufficient time for emergency response and completely avoiding the risk of delay; it can identify 0.1% level membrane micro-damage, improves detection sensitivity by an order of magnitude, and has a fault detection confidence of over 95%, reducing misjudgments.

[0049] Second, it achieves single-cell-level positioning with an error of less than two cell spacings, eliminating the need to disassemble the entire stack for inspection and reducing maintenance time by more than 50%; the fault type identification accuracy is >96%, and it can distinguish eight types of faults such as membrane damage and water flooding, providing a basis for targeted maintenance.

[0050] Third, unplanned downtime is reduced by 85%, and the fuel cell stack life is extended from 8,000 hours to 10,800 hours; maintenance costs are reduced by 60%, spare parts inventory is reduced by 45%, maintenance personnel requirements are reduced by 30%, and the total life cycle cost is optimized by 25%-30%.

[0051] Fourth, the accidental failure rate of the fuel cell stack has been reduced from 15% to below 2%, with an average failure interval of 15,000 hours and an availability of 99.2%, meeting industrial-grade requirements; it is adapted to dynamic operating conditions, avoids false alarms and missed alarms, and can be stably applied to multiple scenarios such as vehicle-mounted and energy storage.

[0052] Fifth, it reduces reliance on expert experience; the model can learn and optimize itself, and after 6 months of operation, the early warning accuracy rate rises to 96% and the false alarm rate drops to 3%. It is compatible with multiple communication protocols and federated learning, and has strong scalability. Attached Figure Description

[0053] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0054] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0055] Figure 1 This is a schematic diagram of the fuel cell hydrogen-oxygen inter-traffic prediction diagnostic method provided in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the workflow of the fuel cell hydrogen-oxygen inter-traffic prediction diagnostic method provided in this embodiment of the invention.

[0057] Figure 3 This is a schematic diagram of the application architecture of the fuel cell hydrogen-oxygen inter-flow predictive diagnostic method provided in this embodiment of the invention.

[0058] Figure 4 This is a schematic diagram of the fuel cell hydrogen-oxygen inter-tandem predictive diagnostic system architecture provided in an embodiment of the present invention. Detailed Implementation

[0059] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1

[0061] See Figure 1 and Figure 2This invention provides a predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells, comprising the following steps:

[0062] S1. Initialization: Load fuel cell stack design parameters to build and initialize the digital twin model, calibrate and benchmark the sensors that collect data to establish a stack health status baseline, pre-train the AI ​​model for fault analysis and load the historical fault case library.

[0063] Among these measures, loading fuel cell stack design parameters, such as film thickness, electrode area, and flow channel structure, provides the physical basis for the digital twin model, ensuring that the digital twin model can accurately map the structure and initial state of the real fuel cell stack; sensor calibration corrects sensor errors through standard signal sources, while benchmark testing collects initial data under normal fuel cell stack operating conditions to form a health status baseline, providing a normal reference system for fault diagnosis; the pre-trained AI model uses historical fault data to train model parameters, and loading a fault case library provides the basis for fault feature matching for the AI ​​model, ensuring that the model has a preliminary fault identification capability.

[0064] S2. Real-time monitoring: Synchronously collect data from multiple sensor sources and perform time-dimensional alignment processing. Perform quality assessment on the collected data to remove abnormal measurement point data. Update the state parameters of the digital twin model in real time based on the processed valid data to make the digital twin model consistent with the state of the fuel cell stack.

[0065] Among these measures, multi-source sensor synchronous acquisition is necessary because different types of data require time consistency to establish a correlation. Time alignment eliminates deviations caused by data acquisition delays through a unified timestamp. Data quality assessment filters invalid data, integrity checks prevent data loss due to sensor failures, consistency checks verify the logical rationality of data from different sensors, and the 3σ criterion, based on the principle of normal distribution, eliminates outliers exceeding the mean ± 3 times the standard deviation to avoid extreme data interfering with diagnosis. The state parameters of the digital twin model are updated based on valid data, and virtual-real synchronization is achieved through data-driven model iteration, ensuring that the model can replicate the operating state of the fuel cell stack in real time.

[0066] S3. Predictive Diagnosis: Periodically perform fusion processing on multi-source data acquired by real-time monitoring, update the probability prediction results of hydrogen-oxygen inter-connection faults based on the characteristics of the fused data, and periodically reassess the remaining service life of the fuel cell stack in order to dynamically control the fault risk and life status.

[0067] Among these measures, the regular fusion of multi-source data is necessary because single data is easily affected by fluctuations in operating conditions. Fusion can integrate fault characteristics from different dimensions, such as voltage fluctuations reflecting abnormal electrical performance and increased oxygen concentration directly related to hydrogen-oxygen crosstalk, thereby improving the reliability of the characteristics. The fault probability is updated based on the fused characteristics, and the matching degree between the current characteristics and fault characteristics is calculated through an AI model to output the probability of the fault occurring. The remaining useful life (RUL) is regularly assessed based on the performance degradation trajectory prediction, combined with the aging law of the fuel cell stack and the fault development trend, to determine the remaining time from the current state to the point where the fuel cell stack can no longer operate normally.

[0068] S4. Early Warning Decision: Based on the probability prediction results and remaining lifespan of the fuel cell stack obtained in the prediction and diagnosis phase, the corresponding early warning level is determined. Maintenance suggestions and handling measures are generated according to the early warning level, and instructions are sent to the fuel cell system through the control interface to execute equipment adjustment operations.

[0069] The early warning level classification is based on two indicators: failure probability and remaining lifespan. The combination of these two indicators can comprehensively reflect the urgency of the risk. Failure probability reflects the likelihood of a failure occurring, while remaining lifespan reflects the time window for the failure to develop. The four early warning levels correspond to different risk levels, avoiding over-response or under-response. Maintenance recommendations are generated based on the early warning level, and targeted solutions are formulated by combining the failure type, risk level, and operation and maintenance costs. By issuing equipment adjustment commands through the control interface, the load on the faulty battery can be reduced, the failure development can be delayed, and time can be bought for maintenance.

[0070] S5. Feedback Learning: Collect the implementation results of maintenance measures and fault confirmation information during the early warning decision-making stage. Based on the implementation results of the maintenance measures and the fault confirmation information, update the AI ​​model parameters and data feature weights, optimize the early warning threshold and diagnostic rules, and achieve iterative improvement of the diagnostic capabilities of the AI ​​model.

[0071] Among these steps, collecting maintenance results and fault confirmation information provides real feedback for model diagnosis; updating AI model parameters and feature weights adjusts the model's sensitivity to different features based on feedback information, correcting model biases; optimizing early warning thresholds and diagnostic rules adapts to stack performance degradation and new fault modes, avoiding false alarms and missed alarms after long-term operation; and finally, through continuous iteration, the model's diagnostic accuracy and early warning accuracy improve over time, ensuring long-term reliable operation of the system.

[0072] In this embodiment, in step S1, during the construction of the digital twin model, intrinsic orthogonal decomposition (POD), real-time parameter identification, and surrogate model construction techniques are used to achieve model order reduction and real-time processing. 500 dominant modes are extracted through intrinsic orthogonal decomposition, reducing the model's computational degrees of freedom from... The number has been reduced to 500. The degrees of freedom are a measure of a model's computational complexity; the higher the degrees of freedom, the longer the computation time. The original degrees of freedom of the electric stack multiphysics model reached... The level of complexity implies a need to solve millions of equations, which ordinary hardware cannot perform in real time. Intrinsic Orthogonal Decomposition (POD) uses singular value decomposition on historical computational or experimental data to select 500 dominant modes that best characterize the dynamic properties of the system. It can restore the core physical behavior of the system with the fewest dimensions, ignore irrelevant minor fluctuations, and meet the computational efficiency requirements of real-time diagnostics.

[0073] Among them, the real-time parameter identification uses an unscented Kalman filter algorithm to update the model parameters, with a parameter update period of 1 second, so that the model parameters match the actual state of the fuel cell stack.

[0074] Specifically, the Unscented Kalman Filter (UKF) algorithm does not require linearization of the nonlinear system. By sampling sigma points to cover the system state space, it can more accurately handle the nonlinear characteristics of the model. Its update process is divided into two steps: prediction and update. In the prediction stage, the parameter state at the next moment is inferred based on the model equation. In the update stage, the predicted value is corrected using real-time sensor data, so that the parameters continuously approach the true value. The 1-second update cycle ensures that the parameters can quickly respond to changes in the state of the fuel cell stack, maintaining the consistency between the virtual and real models and the physical fuel cell stack.

[0075] Among them, the surrogate model is constructed based on the Kriging method to build a real-time response surface as a surrogate model. The Kriging method describes the spatial correlation of data by constructing a variogram function to achieve accurate prediction of data for unmeasured points.

[0076] Specifically, the Kriging method (Kriging interpolation) is a spatial interpolation technique based on statistical theory. By analyzing the spatial correlation of known data, it predicts the values ​​of unknown points. In digital twin models, the number of sensor measurement points inside the fuel cell stack is limited and cannot cover all areas. The Kriging method, by constructing a variogram function, quantifies the correlation strength between known measurement points and then calculates the physical quantities of unmeasured points through weighted interpolation. Its advantages include making full use of the spatial structure information of the data, having small prediction errors, and requiring far less computation than directly solving multiphysics equations. It can serve as a real-time response surface proxy for complex physical models, achieving accurate mapping of the entire state of the fuel cell stack.

[0077] The digital twin model integrates multiphysics control equations, including electrochemical reaction equations, proton conduction equations, gas diffusion equations, and energy conservation equations. Specified parameters from these multiphysics control equations are used as model input variables. Specifically, the specified parameter for the electrochemical reaction equation is the exchange current density. Transmission coefficient limiting current The proton conduction equation specifies the water diffusion coefficient as a parameter. Electroosmotic drag coefficient The gas diffusion equation specifies the binary diffusion coefficient as the parameter. Component mole fraction The energy conservation equation specifies the thermal conductivity as the parameter. Heat of reaction .

[0078] Specifically, electrochemical equations (such as the Butler-Volmer equation) describe the electrochemical reaction rates of hydrogen oxidation and oxygen reduction at the electrode, and their parameters directly determine the electrical performance output, including the exchange current density. Reflecting catalyst activity and transfer coefficient Characterizing reaction reversibility and limiting current Reflects the reactant supply limit; the Nernst-Planck equation describes the proton transport process in the proton exchange membrane; water diffusivity... Affected by membrane hydration state and electroosmotic drag coefficient The degree to which water molecules migrate with protons is reflected, and both determine the proton conduction efficiency; the gas diffusion equation describes the transport process of hydrogen and oxygen in the porous electrode medium, and the binary diffusion coefficient... Reflects gas diffusion rate and component mole fraction Reflecting gas concentration distribution, it is directly related to the sufficiency of reactant supply; the energy conservation equation describes the internal temperature change of the fuel cell stack, and the thermal conductivity... Determines heat transfer efficiency and reaction heat The exothermic / endothermic effects of the electrochemical reaction determine the temperature distribution of the fuel cell stack, preventing localized overheating that could lead to membrane damage. Using these parameters as input variables allows the model to reflect changes in the stack's state through parameter variations, providing mechanistic support for fault diagnosis.

[0079] In this embodiment, in step S2, the multi-source sensor data includes electrical performance data, gas concentration data, and thermal data. The electrical performance data is acquired by the ultra-high-speed CVM system and the multi-band EIS system; the gas concentration data is acquired by a laser oxygen analyzer; and the thermal data is acquired by a distributed temperature and pressure sensing network. Electrical performance data is an indirect characterization of faults. The ultra-high-speed CVM system can capture millisecond-level fluctuations in single-cell voltage. When membrane damage leads to abnormal local reactions, the corresponding cell voltage will exhibit irregular fluctuations. The multi-band EIS system reflects the electrochemical state of the membrane electrode through impedance spectroscopy. For example, membrane damage can lead to high-frequency impedance... Changes and catalyst poisoning can affect charge transfer resistance. Gas concentration data directly characterizes hydrogen-oxygen interferometry. Laser oxygen analyzers, based on TDLAS technology, can accurately detect trace amounts of oxygen in the anode exhaust gas. Under normal circumstances, the anode contains only hydrogen; an increase in oxygen concentration directly indicates oxygen permeation from the cathode. Thermal data provides an environmentally relevant characterization of faults. Distributed temperature and pressure sensing networks can construct the internal temperature and pressure distribution of the fuel cell stack. When membrane damage leads to abnormal local reactions, it is accompanied by local temperature increases and pressure fluctuations, providing spatial information for fault location.

[0080] The data quality assessment includes data integrity checks, data consistency checks, and outlier detection. Data integrity checks confirm that there are no missing data, data consistency checks verify the logical matching between data from different sensors, and outlier detection uses the 3σ criterion to identify and remove data that exceeds the normal range.

[0081] Specifically, data integrity checks are performed by verifying data frame identifiers and timestamp continuity to ensure no data loss due to sensor disconnections or communication interruptions, thus avoiding diagnostic biases caused by incomplete data. Data consistency checks are based on the fuel cell operating mechanism to verify the logical matching of data from different sensors. For example, when the stack power increases, the temperature should show an upward trend, and the voltage should show a stable or slightly downward trend. If the power increases while the temperature drops sharply, the data is inconsistent, and logically contradictory data is eliminated. Outlier detection uses the 3σ criterion, which is based on the principle of normal distribution. 99.73% of normal data will be distributed within the range of mean ± 3 standard deviations. Data outside this range is likely to be outliers caused by sensor interference or sudden noise. Eliminating these outliers can prevent extreme data from misleading the model's judgment and ensure the authenticity of the input data.

[0082] In this embodiment, the multi-source data fusion in step S3 adopts a multi-level fusion strategy, including the DS evidence theory fusion algorithm, the attention mechanism neural network, and the spatiotemporal alignment algorithm; the implementation of the DS evidence theory fusion algorithm includes:

[0083] The support function is calculated based on the distance between data feature values ​​and level feature values, and a basic probability value is assigned to each sensor data. First, fault level feature values ​​need to be preset. Then, the Euclidean or Mahalanobis distance between the feature values ​​collected in real time by each sensor and the feature values ​​of each level is calculated. The closer the distance, the higher the support of the sensor data for the fault level. The support function is constructed based on the distance, and the distance is converted into a basic probability value in the interval [0, 1]. This value directly reflects the degree of trust of a single sensor data for a certain fault proposition.

[0084] Dempster's combination rule is used to process multi-source evidence. The combination formula is as follows:

[0085]

[0086] In the formula, Assign values ​​to the basic probability of proposition A after the evidence is synthesized. , These represent the basic probability assignments for propositions B and C from two sources of evidence. The numerator represents the sum of the confidence products of all evidence combinations that satisfy the condition "the intersection of proposition B and proposition C is A," reflecting the coordinated support of multiple sources of evidence for the same proposition A; the denominator is the complement of the conflict coefficients, where... This represents the sum of confidence levels for conflicting evidence from multiple sources. Normalizing the denominator can eliminate the interference of conflicting evidence on the fusion result, ensuring the assignment of the basic probability after synthesis. It can reasonably reflect the overall level of trust in proposition A from multiple sources of evidence, avoiding the one-sidedness of a single piece of evidence.

[0087] Based on the synthesized trust function and likelihood function, a quantitative assessment of the failure hypothesis is achieved. The trust function Bel(A) is the sum of the basic probability values ​​of all subset propositions supporting A, reflecting the "minimum degree of trust" (completely confirmed trust) in proposition A; the likelihood function reflects the "maximum degree of trust" (not ruled out trust) in proposition A. The two constitute the trust interval [Bel(A), Pl(A)] of proposition A. The reliability of the failure hypothesis can be quantitatively assessed through this interval: if the interval is narrow (Bel(A)≈Pl(A)), it indicates sufficient evidence and a high degree of credibility of the failure hypothesis; if the interval is wide, it indicates insufficient evidence and further data collection is needed. Based on this quantitative assessment, the current failure status can be clarified, providing a basis for early warning.

[0088] In the spatiotemporal alignment algorithm, time registration uses sliding window averaging and linear interpolation to unify the time reference, and spatial registration is based on the three-dimensional mesh model of the electric stack, and the spatial continuity mapping of discrete measurement points is achieved through radial basis functions.

[0089] Specifically, in terms of time registration, different sensors have significantly different sampling rates. Sliding window averaging is used to process high-frequency data, smoothing noise while reducing data volume. Linear interpolation is used to map low-frequency or asynchronous data to a unified timestamp, linearly estimating the value at the intermediate moment based on the values ​​of two adjacent sampling points, ensuring that all data are synchronized in the time dimension with a synchronization accuracy of <1ms. In terms of spatial registration, a three-dimensional mesh model of the fuel cell stack is first constructed, mapping discrete sensor measurement points to the nearest mesh node. The radial basis function is an interpolation function that calculates the spatial distance between the measurement point and the mesh node to construct weights, interpolating the discrete measurement point data to the entire mesh, achieving continuous mapping of spatial data, thereby obtaining the physical field distribution of the entire fuel cell stack.

[0090] In this embodiment, in step S3, the prediction and diagnosis employs a physical information neural network, and the loss function of the physical information neural network is:

[0091]

[0092]

[0093]

[0094]

[0095] In the formula, For the data fitting term, the mean squared error is used for calculation. These are the model's predicted values. These are actual measured values; For physical constraint terms, The model's predicted value for the physical quantity. Equations describing the physical processes of fuel cells include the Butler-Volmer equation and Fick's law; The term represents a monotonicity constraint, and SOH represents the stack health state. For time, It is a linear rectified function; , These are the weighting coefficients. The ReLU function is used to penalize predictions of rising SOH, ensuring that the model's predicted degradation trajectory matches reality. and The weighting coefficients are used to balance the importance of the three losses, so that the model can fit the data while strictly following the laws of physics.

[0096] In this embodiment, in step S4, the warning level is divided into four levels, including Level 1 Attention Level, Level 2 Warning Level, Level 3 Danger Level and Level 4 Emergency Level.

[0097] The Level 1 triggering conditions are a single data feature anomaly and a hydrogen-oxygen crosstalk failure probability. Furthermore, the remaining service life of the fuel cell stack is >500h, and the system response is to log, increase the monitoring frequency to twice the original frequency and perform trend tracking, with an expected processing time of >24 hours;

[0098] Specifically, for single data feature anomalies, such as slight fluctuations at a certain temperature point without forming multi-dimensional corroboration, a failure probability of <10% indicates that the possibility of failure is extremely low, and a remaining lifespan of >500h indicates that even if there are potential risks, there is sufficient time for observation; the system response adopts an enhanced monitoring strategy, doubling the monitoring frequency to more intensively capture data trends, determine whether the anomaly continues or intensifies, record logs for easy follow-up, and an expected processing time of >24 hours provides maintenance personnel with sufficient observation window to avoid blind maintenance.

[0099] Among them, the Level 2 triggering conditions are multiple data features co-abnormal, failure probability of 10%-30% and remaining lifespan of the fuel cell stack of 200-500 hours. The system response is to issue an early warning, appropriately reduce the rated and prepare for maintenance. The expected processing time is 8-24 hours.

[0100] Specifically, Level 2 represents a moderate risk level, triggered by multi-feature coordinated anomalies, indicating that the anomalies have formed a logical correlation and are not random fluctuations. A failure probability of 10%-30% suggests a certain possibility of failure, and a remaining lifespan of 200-500 hours indicates the need for advance maintenance planning. Appropriate system derating involves reducing the power load on the fuel cell stack to decrease the factors that could trigger failure; issuing early warnings and preparing for maintenance; and an expected processing time of 8-24 hours, which balances operational efficiency with production continuity and avoids unplanned downtime.

[0101] Among them, the Level 3 triggering conditions are that the digital twin model predicts an anomaly, the failure probability is 30%-60%, and the remaining service life of the fuel cell stack is 50-200 hours. The system response is to issue an audible and visual alarm, significantly reduce the rating, and schedule emergency maintenance. The expected processing time is 1-8 hours.

[0102] Specifically, Level 3 (Dangerous) is a high-risk level. Anomalies predicted by the digital twin model in the triggering conditions indicate a confirmed fault trend at the mechanistic level. A fault probability of 30%-60% indicates a high probability of failure, and a remaining lifespan of 50-200 hours suggests the need for urgent handling. System-wide audible and visual alarms quickly attract the attention of maintenance personnel, and significant depreciation can greatly slow the rate of fault development, preventing it from escalating into a catastrophic incident. Emergency maintenance is scheduled with an expected processing time of 1-8 hours to ensure repairs are completed before the fault worsens, minimizing losses.

[0103] Among them, the Level 4 triggering conditions are that multiple sources of data jointly confirm the anomaly, the failure probability is >60% and the remaining service life of the fuel cell stack is <50h. The system response is to start the emergency shutdown procedure, activate the explosion suppression system and require immediate on-site handling, with an expected handling time of <1 hour.

[0104] Specifically, Level 4 (Emergency) is an extreme risk level, triggered by multiple sources of data confirming an anomaly. This indicates that all data points to a severe fault, with a fault probability >60%, suggesting a high probability that a fault has already occurred or is about to occur. Remaining lifespan <50 hours indicates that there is no maintenance buffer time. The system's emergency shutdown procedure can immediately shut down the fuel cell stack to prevent further deterioration of the fault. Activating the explosion suppression system is a specific safety measure against hydrogen-oxygen crosstalk, which can quickly suppress the risk of explosion. The expected handling time is <1 hour, requiring maintenance personnel to immediately handle the situation on-site, completely eliminate safety hazards, and ensure the safety of personnel and equipment.

[0105] In one possible embodiment, the time-dimensional alignment process employs differentiated processing methods for different types of data: CVM voltage inspection data is averaged using a 100ms sliding window; temperature and pressure data are mapped to a unified timestamp via linear interpolation; and EIS electrochemical impedance spectroscopy data undergoes feature extraction before fusion at the decision layer. The extracted EIS features include high-frequency impedance. Charge transfer resistance Mass transfer impedance These correspond to the gas diffusion state, membrane hydration state, and catalyst activity, respectively.

[0106] Specifically, CVM data is sampled at a high frequency of 10kHz and contains a large amount of noise. A 100ms sliding window averaging method calculates the mean of the data within the window, which can smooth out the noise, preserve the core fluctuation characteristics, and reduce the data volume to 10Hz for easier subsequent processing. Temperature and pressure data are sampled at 100Hz, and linear interpolation can accurately map them to a unified timestamp. Based on the assumption of "linear change between adjacent points," this ensures time synchronization without distorting the data trend. EIS data is a 30-second full-spectrum scan with high data dimensionality (containing impedance values ​​at different frequencies), making direct synchronization meaningless. Therefore, the core feature is extracted first: high-frequency impedance. The corresponding high-frequency impedance reflects the diffusion state of the gas diffusion layer, and the charge transfer resistance. The corresponding mid-frequency impedance reflects the membrane hydration state and charge transfer efficiency, and the mass transfer impedance. The corresponding low-frequency impedance reflects the catalyst activity and reactant mass transfer efficiency. The extracted features are single-valued variables that can be fused with other data at the decision-making level to achieve dimensionality reduction of high-dimensional data and preservation of effective features.

[0107] In one possible embodiment, the predictive diagnosis process employs an LSTM-based sequence prediction model and a Cox proportional hazards model. The LSTM sequence prediction model includes an LSTM layer, an attention mechanism layer, and a regressor. The LSTM layer encodes the historical performance data sequence of the fuel cell stack to capture long-term data dependencies. The LSTM layer contains an input gate, a forget gate, and an output gate, which are used to control data input, memory update, and output, respectively. The attention mechanism layer focuses on the key time point data for fault prediction by calculating the weight of each time point in the historical sequence. The regressor consists of a fully connected layer and outputs the remaining service life of the fuel cell stack and the corresponding confidence interval.

[0108] Specifically, the LSTM sequence prediction model captures the long-term temporal dependencies of fuel cell stack performance data. Because fault degradation is a gradual process, future trends need to be predicted through historical sequences. The input gate controls the input weights of the current data, the forget gate determines whether historical memories are retained or forgotten, and the output gate controls the output of the current memories. The three work together to enable the model to effectively process long sequence data and avoid gradient vanishing or gradient exploding. The attention mechanism layer automatically identifies key time points for fault prediction by calculating the attention weights at each time point in the historical sequence and assigns them higher weights, improving the model's sensitivity to key information. The regressor consists of fully connected layers that map the features processed by the LSTM and attention mechanism to the remaining useful life (RUL) and confidence interval. The confidence interval is determined based on the statistical analysis of the model's prediction error, reflecting the reliability of the prediction results.

[0109] The Cox proportional hazards model provides Remaining Life (RUL) prediction results at different confidence levels by calculating the time distribution of failure probabilities. This model can calculate the failure probability distribution at different time points, thus obtaining the remaining useful life at different confidence levels. Complementing the LSTM model, the Cox model is better at analyzing failure development probabilities based on risk rates, while the LSTM model is better at predicting specific lifetime values ​​based on sequence trends. Combining the two can improve the accuracy and reliability of RUL prediction.

[0110] In one possible embodiment, in step S5, the online parameter updates for feedback learning have no fixed cycle and are triggered in real time as maintenance results and fault confirmation information are collected. The updated content includes AI model weights, feature importance coefficients, and warning threshold ranges.

[0111] Specifically, maintenance results and fault confirmation information serve as the model's "true labels." Real-time updates enable the model to quickly correct biases. For example, if the model misjudges a feature as a key fault feature, its weight is reduced after feedback. Updating AI model weights involves adjusting the weight values ​​of connections in each layer of the neural network, making the model output more closely match the actual diagnostic results. Updating feature importance coefficients involves redistributing the contribution of different data features to fault diagnosis, highlighting effective features and suppressing ineffective features. Optimizing the warning threshold range involves adjusting the trigger thresholds for each warning level based on the actual fault occurrence. For example, if a 30% fault probability in reality indicates a serious risk, the trigger threshold for Level 3 is lowered to ensure that the warning level matches the actual risk. Through continuous iteration, the model's diagnostic capabilities are dynamically improved.

[0112] See Figure 3 Here is an application architecture for the method of this invention:

[0113] The sensing layer, centered on a distributed sensor network, constructs a comprehensive data acquisition system encompassing electrical performance, gas state, and thermal environment. It employs a 128-channel synchronous acquisition hardware architecture, achieving multi-channel latency-free synchronization through parallel signal acquisition circuits at the hardware level. The sampling rate reaches 10kHz (i.e., a single-cell voltage is acquired every 0.1 milliseconds), with voltage measurement accuracy controlled within ±0.1mV. This system can dynamically capture millisecond-level fluctuations in single-cell voltage, providing core electrical performance data support for subsequent assessment of single-cell consistency and faulty cell location.

[0114] The multi-band electrochemical impedance spectroscopy (EIS) system covers the full frequency range of 0.1Hz-10kHz. It obtains the impedance response characteristics of membrane electrodes through frequency scanning excitation, and completes a full spectrum scan every 30 seconds to generate a complete impedance spectrum. The system can extract key electrochemical parameters such as high-frequency impedance, charge transfer resistance, and mass transfer impedance from the impedance spectrum, which are directly related to the state of core components such as proton exchange membrane integrity and catalyst performance.

[0115] The laser oxygen concentration analyzer is based on tunable semiconductor laser absorption spectroscopy (TDLAS) technology. It utilizes specific absorption lines of oxygen molecules to achieve high-sensitivity detection with a detection limit as low as 0.5 ppb (i.e., it can identify 0.5 units of oxygen per billion volume units) and a response time of only 100 ms. The equipment is installed in the anode exhaust gas circulation pipeline to monitor the oxygen concentration on the anode side in real time. Under normal operating conditions, the anode contains only hydrogen. If an abnormal increase in oxygen concentration occurs, it can be directly determined that oxygen from the cathode has permeated into the anode (hydrogen-oxygen crosstalk), which is a direct monitoring indicator of hydrogen-oxygen crosstalk faults.

[0116] The distributed temperature and pressure sensing network adopts a modular layout design, with 16 temperature monitoring points (using PT1000 platinum resistance sensors, with a temperature measurement accuracy of ±0.1℃) and 8 pressure monitoring points (using MEMS piezoresistive sensors, with a pressure measurement accuracy of ±0.2kPa) deployed in key areas such as the fuel cell stack inlet, middle reaction zone, and outlet, with an overall sampling rate of 100Hz. By acquiring data from multiple points, a three-dimensional temperature and pressure field distribution inside the fuel cell stack is constructed, which can promptly detect "hot spots" (sudden local temperature rises) and pressure fluctuations caused by local reaction anomalies due to membrane damage, providing spatial dimension data support for fault location.

[0117] The transport layer is responsible for efficient data aggregation, real-time preprocessing, and bidirectional interaction. It addresses the spatiotemporal alignment of multi-source heterogeneous data through an architecture that combines edge-local processing with cloud collaboration. Based on the NVIDIA Jetson AGX Orin edge computing platform, it integrates a GPU core providing 200 TOPS of computing power (optimized for deep learning inference and real-time data processing). Nodes are deployed near the fuel cell stack, enabling local preprocessing of raw sensor data (such as 100ms sliding window denoising for CVM data and outlier removal from temperature and pressure data). This reduces bandwidth consumption for transmitting raw data to the cloud while achieving low-latency data processing (preprocessing latency <10ms), meeting the timeliness requirements of real-time diagnostics.

[0118] The real-time data acquisition system uses the NI PXIe-8840 high-performance controller as its core and integrates an FPGA (Field Programmable Gate Array) coprocessor. The FPGA realizes parallel acquisition and synchronous control of multi-channel data through hardware logic circuits, and can accurately synchronize the acquisition clocks of different types of devices such as CVM, EIS, and temperature and pressure sensors, ensuring that the timestamp deviation of multi-source data is <1ms. The system supports multiple industrial bus interfaces and is compatible with the data output formats of different sensors, realizing "one-stop" data aggregation.

[0119] The cloud-edge collaborative architecture is built on the AWS IoT Greengrass edge computing framework. Edge nodes are connected to the cloud via a high-speed network with a stable bandwidth of >100Mbps. This architecture supports bidirectional data flow between the edge and the cloud: edge nodes upload pre-processed feature data (rather than raw data) to the cloud for global optimization of the digital twin model and storage of historical data; the cloud can issue control commands to edge nodes, and at the same time realize data collaboration among multiple edge nodes. In addition, the architecture has the function of resuming transmission after network failure, ensuring that data is not lost when the network fluctuates and ensuring system continuity.

[0120] The digital twin layer, driven by both physical mechanism modeling and data-driven optimization, constructs a digital mapping model highly consistent with the physical fuel cell stack, enabling fault mechanism analysis and evolution prediction. The multiphysics coupling model, built on the COMSOL Multiphysics real-time solver, integrates four core physical field governing equations: electrochemical reaction, proton conduction, gas diffusion, and heat transfer. The electrochemical reaction equations use the Butler-Volmer equation to describe the hydrogen oxidation and oxygen reduction reaction rates on the electrode surface, and correlate parameters such as catalyst activity and reaction temperature. The proton conduction equation is based on the Nernst-Planck equation to describe the proton transport process in the exchange membrane, and dynamically adjusts the proton diffusion coefficient by combining the membrane hydration state. The gas diffusion equation describes the transport law of hydrogen, oxygen, and water vapor in the porous electrode and gas diffusion layer through a multi-component diffusion model. The heat transfer equation is based on Fourier's law to construct a heat conduction model, taking into account heat sources such as electrochemical reaction heat and Joule heat, and dynamically predicting the temperature distribution of the fuel cell stack. The model achieves accurate replication of the microscopic processes inside the fuel cell stack (such as gas permeation paths and local reaction changes after membrane damage) by solving the coupled equations in real time, providing physical support for the analysis of failure mechanisms.

[0121] The data-driven AI model employs the PyTorch deep learning framework to construct a Physical Information Neural Network (PINN), embedding first-principles constraints for fuel cells (such as the laws of energy conservation and charge conservation) into the traditional neural network. The model input consists of multi-source feature data collected by the sensing layer (voltage fluctuation characteristics, impedance parameters, oxygen concentration, and temperature and pressure data), and the output is the probability of failure and the remaining useful life (RUL). Physical constraints ensure that the model's predictions conform to the actual operating mechanism (e.g., avoiding contradictory results such as "predicting membrane rupture despite a sudden temperature drop"), while also improving the model's generalization ability under small sample sizes and complex operating conditions. The adaptive learning engine constructs a dynamic optimization module based on an online parameter update mechanism. The engine receives real-time feedback from two sources: maintenance results data, such as the deviation between the confirmed fault type and location after maintenance and the model's prediction; and the deviation between real-time sensor data and the model's predicted values. Through gradient descent algorithm, the digital twin model parameters and AI model weights are dynamically adjusted, achieving a parameter update every second. This ensures continuous optimization of the model as the fuel cell stack performance degrades, maintaining high prediction accuracy over the long term.

[0122] At the application level, it provides maintenance personnel and system administrators with a full-process function for fault warning, precise location, and maintenance decision-making, transforming the analysis results of the digital twin layer into executable maintenance solutions. The predictive diagnostic center, based on the multi-physics coupling model and AI model analysis results of the digital twin layer, achieves advanced warning of hydrogen-oxygen crosstalk faults. By monitoring membrane integrity-related parameters and combining fault evolution patterns, it can issue warnings 30 minutes before the actual occurrence of a hydrogen-oxygen crosstalk fault, with a maximum warning window of 24 hours. Simultaneously, the system can cross-validate fault location using single-cell voltage data and temperature-pressure field distribution data, pinpointing the fault location accuracy to a specific single cell, avoiding the limitations of traditional technologies that can only locate faults to the stack module. The precise location engine uses a spatiotemporal graph convolutional network (ST-GCN) to construct a fault location model. Input data includes fault feature evolution sequences in the time dimension and stack structure data in the spatial dimension. The model extracts fault evolution patterns through temporal convolution and correlates the state relationships of adjacent cells through spatial convolution, enabling precise location of the fault source (such as the specific number of the cell with membrane damage) and prediction of the fault evolution trajectory (such as the speed of membrane damage expansion, potential...). The system provides maintenance personnel with dual information: "fault location + development trend," based on fault prediction results (fault type, severity) and remaining useful life (RUL) assessment. The maintenance decision support system generates a full-process maintenance plan based on fault prediction results (fault type, severity) and remaining useful life (RUL). Life prediction combines the current performance status of the stack and historical degradation data to output remaining useful life (e.g., "current RUL is 200 hours") and life ranges at different confidence levels. Plan optimization recommends targeted maintenance measures based on fault type (e.g., membrane damage, catalyst poisoning) and stack operating scenario (e.g., vehicle / energy storage) (e.g., "replace the membrane electrode of the faulty battery number 187" or "reduce the stack load to 70% to delay fault development"). Resource planning automatically calculates the required spare parts (e.g., membrane electrode model, quantity) and estimates the maintenance time (e.g., "replacing a single battery requires 1.5 hours"), generating a maintenance work order. Simultaneously, it links with the spare parts management system to ensure sufficient spare parts inventory, reducing maintenance costs.

[0123] The application scenarios of this invention are as follows:

[0124] Scenario 1: Commercial fuel cell buses / heavy trucks in the field of on-board fuel cells:

[0125] In commercial fuel cell vehicles such as urban buses and long-distance freight trucks, the fuel cell stack, as the core power source, operates under high load and variable working conditions for a long time. The proton exchange membrane is prone to micro-damage due to vibration and temperature fluctuations, leading to the risk of hydrogen-oxygen cross-leakage. The present invention can monitor the changes in the internal temperature field and pressure field of the fuel cell stack in real time through a distributed temperature and pressure sensing network, and capture the millisecond-level voltage fluctuations of a single cell in combination with an ultra-high-speed CVM system, giving an early warning 30 minutes to 24 hours before the occurrence of a hydrogen-oxygen cross-leakage fault; at the same time, a precise positioning engine can lock the faulty single cell, and the maintenance decision support system can generate a maintenance plan in combination with the vehicle operation plan, avoiding breakdowns caused by sudden faults during vehicle driving, and ensuring operation safety and efficiency.

[0126] Scenario 2: Fuel cell passenger vehicles in the field of on-vehicle fuel cells:

[0127] Passenger vehicles have higher requirements for system compactness and reliability, and the operating environment is complex (such as high-temperature exposure and low-temperature cold). The sensing layer of the present invention can be integrated into the vehicle fuel cell stack compartment. The laser oxygen analyzer monitors the oxygen concentration in the anode exhaust gas in real time, and the multi-band EIS system regularly scans the electrochemical state of the membrane electrode to timely detect potential problems such as insufficient membrane hydration during low-temperature startup and membrane aging during high-temperature operation; the transmission layer is linked with the vehicle control system through the CANFD bus. If a Level 4 emergency warning is triggered, it can automatically start the emergency shutdown procedure and activate the explosion suppression system to protect the safety of the driver and passengers; at the same time, the cloud-edge collaborative architecture can upload the vehicle diagnosis data to the cloud platform of the vehicle manufacturer to achieve fleet-level operation and maintenance management and optimize the allocation of maintenance resources.

[0128] Scenario 3: Megawatt-level grid-side energy storage power stations in the field of stationary hydrogen energy storage power stations:

[0129] In a megawatt-level hydrogen energy storage power station supporting wind and solar power generation, multiple fuel cell stacks operate in parallel and need to output electric energy stably for a long time. A hydrogen-oxygen cross-leakage fault may cause the entire power station to shut down. The present invention can adopt a regional distributed architecture, with one edge computing node configured for every five fuel cell stacks to converge multi-source data in real time; the digital twin layer simulates the operating state of the fuel cell stack cluster through a multi-physics field coupling model, and shares the diagnostic model among multiple power stations in combination with federated learning technology, which not only ensures data privacy but also improves the overall diagnostic accuracy; when a certain fuel cell stack triggers a Level 3 danger warning, the maintenance decision support system can dispatch standby fuel cell stacks to operate, and at the same time generate a maintenance work order, reducing the unplanned shutdown time by 85% and ensuring the stable peak shaving and frequency modulation ability of the power station for the power grid.

[0130] Scenario 4: Distributed user-side energy storage power stations in the field of stationary hydrogen energy storage power stations:

[0131] In user-side energy storage scenarios such as industrial parks and data centers, fuel cell stacks are mostly used as backup or peak-shaving power sources, resulting in large fluctuations in operating load. The predictive diagnostic center of this invention can identify faults such as membrane damage and catalyst poisoning in advance, adjusting maintenance schedules based on user electricity demand (e.g., peak data center usage). Simultaneously, the maintenance decision support system can optimize spare parts inventory, reducing spare parts costs by 45%, and by predicting remaining service life, avoid resource waste caused by "over-maintenance," ensuring reliable power supply to the energy storage power station during grid outages.

[0132] Example 2

[0133] See Figure 4 Embodiment 2 of the present invention also provides a fuel cell hydrogen-oxygen crosstalk predictive diagnostic system for implementing the fuel cell hydrogen-oxygen crosstalk predictive diagnostic method of the above embodiments, comprising:

[0134] Initialization module 100 is used to load fuel cell stack design parameters to build and initialize a digital twin model, calibrate and benchmark the sensors that collect data to establish a stack health status baseline, pre-train an AI model for fault analysis and load a historical fault case library.

[0135] The real-time monitoring module 200 is used to synchronously collect data from multiple sources of sensors and perform time-dimensional alignment processing, perform quality assessment on the collected data to remove abnormal measurement point data, and update the state parameters of the digital twin model in real time based on the processed valid data, so that the digital twin model is consistent with the state of the fuel cell stack.

[0136] The predictive diagnostic module 300 is used to periodically perform fusion processing on multi-source data acquired by real-time monitoring, update the probability prediction results of hydrogen-oxygen inter-connection faults based on the characteristics of the fused data, and periodically reassess the remaining service life of the fuel cell stack in order to dynamically control the fault risk and life status.

[0137] The early warning decision module 400 is used to determine the corresponding early warning level based on the probability prediction results and the remaining service life of the fuel cell stack obtained in the prediction and diagnosis stage, generate maintenance suggestions and handling measures based on the early warning level, and issue instructions to the fuel cell system through the control interface to execute equipment adjustment operations.

[0138] The feedback learning module 500 is used to collect the implementation results of maintenance measures and fault confirmation information during the early warning decision-making stage. Based on the implementation results of the maintenance measures and the fault confirmation information, the AI ​​model parameters and data feature weights are updated to optimize the early warning threshold and diagnostic rules, thereby achieving iterative improvement of the diagnostic capabilities of the AI ​​model.

[0139] In this embodiment, the initialization module 100 uses intrinsic orthogonal decomposition (POD), real-time parameter identification, and surrogate model construction techniques to achieve model order reduction and real-time processing.

[0140] 500 dominant modes were extracted through intrinsic orthogonal decomposition, increasing the computational degrees of freedom of the model from 10. 6 Dropped to 500;

[0141] Real-time parameter identification uses an unscented Kalman filter algorithm to update model parameters, with a parameter update period of 1 second, so that the model parameters match the actual state of the fuel cell stack.

[0142] The surrogate model is constructed based on the Kriging method to build a real-time response surface as a surrogate model. The Kriging method describes the spatial correlation of data by constructing a variogram function, thereby achieving accurate prediction of data at unmeasured points.

[0143] The digital twin model integrates multiphysics governing equations, including electrochemical reaction equations, proton conduction equations, gas diffusion equations, and energy conservation equations. Specified parameters from these multiphysics governing equations are used as model input variables. Specifically, the specified parameter for the electrochemical reaction equation is the exchange current density. Transmission coefficient limiting current The proton conduction equation specifies the water diffusion coefficient as a parameter. Electroosmotic drag coefficient The gas diffusion equation specifies the binary diffusion coefficient as the parameter. Component mole fraction The energy conservation equation specifies the thermal conductivity as the parameter. Heat of reaction .

[0144] In this embodiment, the real-time monitoring module 200 includes multi-source sensor data such as electrical performance data, gas concentration data, and thermal data; the electrical performance data is collected by an ultra-high-speed CVM system and a multi-band EIS system, the gas concentration data is collected by a laser oxygen analyzer, and the thermal data is collected by a distributed temperature and pressure sensing network.

[0145] In the real-time monitoring module 200, the data quality assessment includes data integrity check, data consistency check and data outlier detection. The data integrity check confirms that there are no missing data. The data consistency check verifies the logical matching between data from different sensors. The data outlier detection uses the 3σ criterion to identify and remove data that exceeds the normal range.

[0146] In this embodiment, the multi-source data fusion in the prediction and diagnosis module 300 adopts a multi-level fusion strategy, including the DS evidence theory fusion algorithm, the attention mechanism neural network, and the spatiotemporal alignment algorithm.

[0147] The implementation methods of the DS evidence theory fusion algorithm include:

[0148] The support function is calculated based on the distance between data feature values ​​and rank feature values, and a basic probability value is assigned to each sensor data.

[0149] Dempster's combination rule is used to process multi-source evidence. The combination formula is as follows:

[0150]

[0151] In the formula, Assign values ​​to the basic probability of proposition A after the evidence is synthesized. , The basic probability assignments for propositions B and C are given by the two sources of evidence, respectively.

[0152] Based on the synthesized trust function and likelihood function, a quantitative assessment of the failure hypothesis is achieved;

[0153] In the spatiotemporal alignment algorithm, time registration uses sliding window averaging and linear interpolation to unify the time reference, and spatial registration is based on the three-dimensional mesh model of the electric stack, and the spatial continuity mapping of discrete measurement points is achieved through radial basis functions.

[0154] In this embodiment, the prediction and diagnosis module 300 employs a physical information neural network for prediction and diagnosis, and the loss function of the physical information neural network is:

[0155]

[0156]

[0157]

[0158]

[0159] In the formula, For the data fitting term, the mean squared error is used for calculation. These are the model's predicted values. These are actual measured values; For physical constraint terms, The model's predicted value for the physical quantity. Equations describing the physical processes of fuel cells include the Butler-Volmer equation and Fick's law; The term represents a monotonicity constraint, and SOH represents the stack health state. For time, It is a linear rectified function; , These are the weighting coefficients.

[0160] In this embodiment, the early warning decision module 400 has four levels of early warning, including Level 1 Attention Level, Level 2 Warning Level, Level 3 Danger Level and Level 4 Emergency Level.

[0161] The Level 1 trigger condition is a single abnormal data feature, a hydrogen-oxygen crosstalk fault probability of <10%, and a remaining lifespan of the fuel cell stack >500h. The system response is to log, increase the monitoring frequency to twice the original frequency, and perform trend tracking. The expected processing time is >24 hours.

[0162] Level 2 triggering conditions are multiple data feature anomalies, a failure probability of 10%-30%, and a remaining lifespan of 200-500 hours for the fuel cell stack. The system response is to issue an early warning, appropriately reduce the rated load, and prepare for maintenance. The expected processing time is 8-24 hours.

[0163] Level 3 triggering conditions are: the digital twin model predicts an anomaly, the probability of failure is 30%-60%, and the remaining service life of the fuel cell stack is 50-200 hours. The system response is to issue an audible and visual alarm, significantly reduce the rating, and schedule emergency maintenance. The expected processing time is 1-8 hours.

[0164] Level 4 trigger conditions are: multiple data sources jointly confirm the anomaly, the probability of failure is >60%, and the remaining service life of the fuel cell stack is <50 hours. The system response is to start the emergency shutdown procedure, activate the explosion suppression system, and require immediate on-site handling. The expected handling time is <1 hour.

[0165] In one possible embodiment, the real-time monitoring module 200 employs differentiated processing methods for different types of data: CVM voltage inspection data is averaged using a 100ms sliding window; temperature and pressure data are mapped to a unified timestamp via linear interpolation; and EIS electrochemical impedance spectroscopy data undergoes feature extraction before fusion at the decision layer. The extracted EIS features include high-frequency impedance. Charge transfer resistance Mass transfer impedance These correspond to the gas diffusion state, membrane hydration state, and catalyst activity, respectively.

[0166] In one possible embodiment, the prediction and diagnosis module 300 employs an LSTM sequence prediction model and a Cox proportional hazards model. The LSTM sequence prediction model includes an LSTM layer, an attention mechanism layer, and a regressor. The LSTM layer encodes the historical performance data sequence of the fuel cell stack to capture long-term data dependencies. The LSTM layer contains an input gate, a forget gate, and an output gate, which are used to control data input, memory update, and output, respectively. The attention mechanism layer calculates the weight of each time point in the historical sequence to predict key time point data for faults. The regressor is composed of a fully connected layer and outputs the remaining service life of the fuel cell stack and the corresponding confidence interval.

[0167] The Cox proportional hazards model provides RUL prediction results at different confidence levels by calculating the time distribution of failure probability.

[0168] In one possible embodiment, the online parameter updates of the feedback learning module 500 have no fixed cycle and are triggered in real time as maintenance results and fault confirmation information are collected. The updated content includes AI model weights, feature importance coefficients, and early warning threshold ranges.

[0169] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0170] Example 3

[0171] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a fuel cell hydrogen-oxygen inter-traffic prediction diagnostic method. The program code includes instructions for executing the fuel cell hydrogen-oxygen inter-traffic prediction diagnostic method of Embodiment 1 or any possible implementation thereof.

[0172] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0173] Example 4

[0174] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0175] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the fuel cell hydrogen-oxygen inter-tandem predictive diagnostic method of Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0176] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0177] 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. 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 the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0178] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0179] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A predictive diagnostic method for hydrogen-oxygen crosstalk in a fuel cell, characterized in that, Includes the following steps: (1) Initialization: Load the fuel cell stack design parameters to build and initialize the digital twin model, calibrate and benchmark the sensors for data acquisition to establish the stack health status baseline, pre-train the AI ​​model for fault analysis and load the historical fault case library; (2) Real-time monitoring: Synchronously collect multi-source sensor data and perform time dimension alignment processing, perform quality assessment on the collected data to remove abnormal measurement point data, and update the state parameters of the digital twin model in real time based on the processed valid data so that the digital twin model is consistent with the state of the fuel cell stack. (3) Predictive diagnosis: Regularly perform fusion processing on the multi-source data obtained from real-time monitoring, update the probability prediction results of hydrogen-oxygen inter-connection faults based on the characteristics of the fused data, and regularly reassess the remaining service life of the stack in order to dynamically control the fault risk and life status. (4) Early warning decision: Based on the probability prediction results and the remaining service life of the fuel cell stack obtained in the prediction and diagnosis stage, the corresponding early warning level is determined. Maintenance suggestions and handling measures are generated according to the early warning level, and instructions are sent to the fuel cell system through the control interface to execute equipment adjustment operations. (5) Feedback learning: Collect the implementation results of maintenance measures and fault confirmation information in the early warning decision-making stage, update the AI ​​model parameters and data feature weights based on the implementation results of the maintenance measures and the fault confirmation information, optimize the early warning threshold and diagnostic rules, and realize the iterative improvement of the diagnostic capabilities of the AI ​​model.

2. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In the process of constructing the digital twin model, intrinsic orthogonal decomposition (POD), real-time parameter identification, and surrogate model construction techniques are used to achieve model order reduction and real-time operation. 500 dominant modes were extracted through intrinsic orthogonal decomposition, increasing the computational degrees of freedom of the model from 10. 6 Dropped to 500; Real-time parameter identification uses an unscented Kalman filter algorithm to update model parameters, with a parameter update period of 1 second, so that the model parameters match the actual state of the fuel cell stack. The surrogate model is constructed based on the Kriging method to build a real-time response surface as a surrogate model. The Kriging method describes the spatial correlation of data by constructing a variogram function, thereby achieving accurate prediction of data at unmeasured points. The digital twin model integrates multiphysics governing equations, including electrochemical reaction equations, proton conduction equations, gas diffusion equations, and energy conservation equations. Specified parameters from these multiphysics governing equations are used as model input variables. Specifically, the specified parameter for the electrochemical reaction equation is the exchange current density. Transmission coefficient limiting current The proton conduction equation specifies the water diffusion coefficient as a parameter. Electroosmotic drag coefficient The gas diffusion equation specifies the binary diffusion coefficient as the parameter. Component mole fraction The energy conservation equation specifies the thermal conductivity as the parameter. Heat of reaction .

3. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In step (2), the multi-source sensor data includes electrical performance data, gas concentration data and thermal data; the electrical performance data is collected by the ultra-high-speed CVM system and the multi-band EIS system, the gas concentration data is collected by the laser oxygen analyzer, and the thermal data is collected by the distributed temperature and pressure sensing network. Data quality assessment includes data integrity checks, data consistency checks, and outlier detection. Data integrity checks confirm that there are no missing data, data consistency checks verify the logical matching between data from different sensors, and outlier detection uses the 3σ criterion to identify and remove data that exceeds the normal range.

4. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In step (3), the multi-source data fusion adopts a multi-level fusion strategy, including the DS evidence theory fusion algorithm, the attention mechanism neural network, and the spatiotemporal alignment algorithm; The implementation methods of the DS evidence theory fusion algorithm include: The support function is calculated based on the distance between data feature values ​​and rank feature values, and a basic probability value is assigned to each sensor data. Dempster's combination rule is used to process multi-source evidence. The combination formula is as follows: ； In the formula, Assign values ​​to the basic probability of proposition A after the evidence is synthesized. , The basic probability assignments for propositions B and C are given by the two sources of evidence, respectively. Based on the synthesized trust function and likelihood function, a quantitative assessment of the failure hypothesis is achieved; In the spatiotemporal alignment algorithm, time registration uses sliding window averaging and linear interpolation to unify the time reference, and spatial registration is based on the three-dimensional mesh model of the electric stack, and the spatial continuity mapping of discrete measurement points is achieved through radial basis functions.

5. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In step (3), the prediction and diagnosis employs a physical information neural network, and the loss function of the physical information neural network is: ； ； ； ； In the formula, For the data fitting term, the mean squared error is used for calculation. These are the model's predicted values. These are actual measured values; For physical constraint terms, The model's predicted value for the physical quantity. Equations describing the physical processes of fuel cells include the Butler-Volmer equation and Fick's law; The term represents a monotonicity constraint, and SOH represents the stack health state. For time, It is a linear rectified function; , These are the weighting coefficients.

6. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In step (4), the warning level is divided into four levels, including Level 1 Attention Level, Level 2 Warning Level, Level 3 Danger Level and Level 4 Emergency Level; The Level 1 trigger condition is a single abnormal data feature, a hydrogen-oxygen crosstalk fault probability of <10%, and a remaining lifespan of the fuel cell stack >500h. The system response is to log, increase the monitoring frequency to twice the original frequency, and perform trend tracking. The expected processing time is >24 hours. Level 2 triggering conditions are multiple data feature anomalies, a failure probability of 10%-30%, and a remaining lifespan of 200-500 hours for the fuel cell stack. The system response is to issue an early warning, appropriately reduce the rated load, and prepare for maintenance. The expected processing time is 8-24 hours. Level 3 triggering conditions are: the digital twin model predicts an anomaly, the probability of failure is 30%-60%, and the remaining service life of the fuel cell stack is 50-200 hours. The system response is to issue an audible and visual alarm, significantly reduce the rating, and schedule emergency maintenance. The expected processing time is 1-8 hours. Level 4 trigger conditions are: multiple data sources jointly confirm the anomaly, the probability of failure is >60%, and the remaining service life of the fuel cell stack is <50 hours. The system response is to start the emergency shutdown procedure, activate the explosion suppression system, and require immediate on-site handling. The expected handling time is <1 hour.

7. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In the time-dimensional alignment process, different processing methods are adopted for different types of data: CVM voltage inspection data is averaged using a 100ms sliding window; temperature and pressure data are mapped to a unified timestamp through linear interpolation; and EIS electrochemical impedance spectroscopy data are first feature-extracted and then fused at the decision layer. The extracted EIS features include high-frequency impedance. Charge transfer resistance Mass transfer impedance These correspond to the gas diffusion state, membrane hydration state, and catalyst activity, respectively.

8. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, During the predictive diagnosis process, an LSTM-based sequence prediction model and a Cox proportional hazards model are used. The LSTM sequence prediction model includes an LSTM layer, an attention mechanism layer, and a regressor. The LSTM layer encodes the historical performance data sequence of the fuel cell stack to capture long-term data dependencies. The LSTM layer contains an input gate, a forget gate, and an output gate, which are used to control data input, memory update, and output, respectively. The attention mechanism layer calculates the weight of each time point in the historical sequence to predict key time point data for faults. The regressor consists of a fully connected layer and outputs the remaining service life of the fuel cell stack and the corresponding confidence interval. The Cox proportional hazards model provides RUL prediction results at different confidence levels by calculating the time distribution of failure probability.

9. The predictive diagnostic method for hydrogen-oxygen crosstalk in fuel cells according to claim 1, characterized in that, In step (5), the online parameter updates for feedback learning have no fixed cycle and are triggered in real time as maintenance results and fault confirmation information are collected. The updated content includes AI model weights, feature importance coefficients, and warning threshold ranges.

10. A fuel cell hydrogen-oxygen crosstalk predictive diagnostic system, used to implement the fuel cell hydrogen-oxygen crosstalk predictive diagnostic method according to any one of claims 1-9, characterized in that, include: The initialization module is used to load fuel cell stack design parameters to build and initialize a digital twin model, calibrate and benchmark the sensors that collect data to establish a baseline for the health status of the stack, pre-train an AI model for fault analysis, and load a historical fault case library. The real-time monitoring module is used to synchronously collect data from multiple sources of sensors and perform time-dimensional alignment processing. It performs quality assessment on the collected data to remove abnormal measurement point data and updates the state parameters of the digital twin model in real time based on the processed valid data, so that the digital twin model is consistent with the state of the fuel cell stack. The predictive diagnostic module is used to periodically perform fusion processing on multi-source data acquired by real-time monitoring, update the probability prediction results of hydrogen-oxygen inter-connection faults based on the characteristics of the fused data, and periodically reassess the remaining service life of the fuel cell stack in order to dynamically control the fault risk and life status. The early warning decision module is used to determine the corresponding early warning level based on the probability prediction results and the remaining service life of the fuel cell stack obtained in the prediction and diagnosis stage, generate maintenance suggestions and handling measures based on the early warning level, and issue instructions to the fuel cell system through the control interface to execute equipment adjustment operations. The feedback learning module is used to collect the implementation results of maintenance measures and fault confirmation information during the early warning decision-making stage. Based on the implementation results of the maintenance measures and the fault confirmation information, the AI ​​model parameters and data feature weights are updated to optimize the early warning threshold and diagnostic rules, thereby achieving iterative improvement of the diagnostic capabilities of the AI ​​model.

Images