Hydrogen fuel cell stack online health assessment and fault early warning method, system, device and medium

Abstract

The application provides an online health evaluation and fault early warning method, system, device and medium for a hydrogen fuel cell stack. The method comprises: collecting external measurable data representing the operating state of the stack; predicting the operating state of the stack by a mechanism-based multi-physical field coupling digital model; using a state estimation algorithm, taking the digital model as a prediction model and the external measurable data as observation values, iteratively correcting the prior estimation of the internal implicit state of the stack by comparing the deviation between the predicted values and the observation values, and online estimating the internal implicit state of the stack; calculating the difference between the external measurable values predicted by the digital model and the external measurable data to obtain a multi-dimensional residual vector; calculating the multi-dimensional residual vector to obtain a comprehensive health index; and matching the multi-dimensional residual vector with a preset fault feature mode library to identify a specific fault type. The application can quantitatively evaluate the health state and identify early faults before the external parameters change.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for hydrogen fuel cells, specifically to an online health assessment and fault early warning method, system, electronic device, and storage medium for air-cooled hydrogen fuel cell stacks. Background Technology

[0002] Air-cooled hydrogen fuel cells have been widely used in drones, two-wheeled vehicles, and other fields due to their advantages of simple structure, lightweight design, and low cost. However, air-cooled hydrogen fuel cell stacks inevitably experience performance degradation and face various failure risks during long-term operation, such as the drying or flooding of the proton exchange membrane and the decay of catalyst activity.

[0003] The technical problems associated with these degradation and failure processes stem from changes in the internal state of the battery stack, such as the distribution of water content within the membrane, local current density, and internal temperature field. These internal states are typically impossible to measure by directly installing sensors. Most existing battery management systems rely on setting fixed thresholds for externally measurable parameters (such as total voltage and outlet temperature) to trigger alarms. This approach is essentially a "post-mortem alarm" mode; an alarm is only triggered when internal degradation accumulates to a certain extent, causing a significant deterioration in external parameters. By this time, the battery stack may have already suffered irreversible damage, leading to unplanned downtime or even safety risks.

[0004] Therefore, existing technologies lack a method to identify performance degradation trends and provide early warnings of specific fault types in the early stages of a failure, when there are no obvious abnormalities in externally measurable parameters, thus failing to achieve predictive maintenance for air-cooled hydrogen fuel cell stacks. Summary of the Invention

[0005] The purpose of this invention is to provide an online health assessment and fault early warning method, system, electronic device and storage medium for air-cooled hydrogen fuel cell stacks, aiming to solve the technical problem that the prior art can only perform post-event alarms and cannot achieve early fault warning.

[0006] To achieve the above objectives, this invention provides an online health assessment and fault early warning method for air-cooled hydrogen fuel cell stacks, comprising: collecting external measurable data characterizing the stack's operating state; predicting the stack's operating state using a mechanism-based multiphysics coupled digital model; employing a state estimation algorithm, using the multiphysics coupled digital model as the prediction model and the external measurable data as the observations, iteratively correcting the prior estimate of the stack's internal implicit state by comparing the deviation between the predicted and observed values, and estimating the stack's internal implicit state online; calculating the difference between the external measurable values ​​predicted by the multiphysics coupled digital model and the external measurable data to obtain a multidimensional residual vector, where each dimension of the multidimensional residual vector corresponds to the deviation between the predicted and measured values ​​of one type of external measurable data; calculating a comprehensive health index from the multidimensional residual vector, where the comprehensive health index is a normalized distance metric of the distribution of the multidimensional residual vector relative to the residual vector distribution under healthy conditions; and matching the multidimensional residual vector with a preset fault feature pattern library to identify specific fault types.

[0007] Furthermore, the mechanism-based multiphysics coupled digital model includes at least an electrochemical sub-model, a thermal sub-model, and a fluid sub-model. These sub-models form a bidirectional coupling mechanism through feedback correlation between the heat generation rate and the temperature field, as well as the influence of temperature on the electrochemical reaction rate, and are solved iteratively within each solution step. This refined model can more accurately simulate the complex physicochemical processes inside the fuel cell stack, providing a reliable foundation for state estimation and residual analysis.

[0008] Furthermore, matching the multidimensional residual vector with the fault feature pattern library includes: calculating the similarity between the multidimensional residual vector and each pattern in the fault feature pattern library using a cosine similarity algorithm. By calculating the cosine of the angle between the vectors, the similarity between real-time fault features and standard fault patterns can be effectively determined, enabling rapid and accurate identification of fault types.

[0009] Furthermore, the calculation of the multidimensional residual vector to obtain the comprehensive health index is performed using the Mahalanobis distance calculation method, and the specific formula is as follows: H = sqrt((x - μ)^T Σ^(-1) (x - μ)); Where H is the comprehensive health index, x is the currently calculated multidimensional residual vector, μ is the mean vector of the residual vector under healthy conditions, T is the vector transpose, and Σ is the covariance matrix of the residual vector under healthy conditions. Mahalanobis distance considers the correlation between the components of the multidimensional residual vector during calculation. Compared to simple Euclidean distance, it can more accurately measure the "abnormality" of the actual state point deviating from the healthy state distribution, thus more sensitively reflecting the early deterioration trend of the overall health status of the fuel cell stack.

[0010] Furthermore, the externally measurable data includes at least two of the following: single-cell voltage, stack outlet temperature, gas flow rate, gas pressure, and environmental parameters. These data reflect the macroscopic operating state of the stack from multiple dimensions such as electrical, thermal, and fluid aspects, providing rich observational information for state estimation algorithms to more accurately calibrate and estimate the implicit internal state of the stack.

[0011] Furthermore, the method also includes: setting at least two thresholds based on the comprehensive health indicators, and classifying the health status of the fuel cell stack into healthy, concern, and alarm levels. By setting graded thresholds, the health status of the fuel cell stack can be quantitatively managed, providing clear guidance for operation and maintenance decisions.

[0012] Furthermore, the method also includes: when the comprehensive health indicator indicates a deterioration in the state, but the similarity between the multidimensional residual vector and all patterns in the fault feature pattern library is lower than a preset threshold, marking the multidimensional residual vector as an unknown abnormal pattern and recording it. This step enables the system to have the ability to learn on its own and expand its pattern library, enabling it to cope with new and unforeseen fault modes and enhancing the robustness and adaptability of the system.

[0013] This invention also provides an online health assessment and fault early warning system for an air-cooled hydrogen fuel cell stack, comprising: a data acquisition module for acquiring external measurable data characterizing the stack's operating state; a model prediction module for predicting the stack's operating state using a mechanism-based multiphysics coupled digital model; and a state estimation module for employing a state estimation algorithm, using the multiphysics coupled digital model as a prediction model and the external measurable data as observations, iteratively correcting the prior estimate of the stack's internal implicit state by comparing the deviation between the predicted and observed values, and estimating the stack's internal implicit state online. The system includes a difference calculation module for calculating the difference between the externally measurable value predicted by the multiphysics coupled digital model and the externally measurable data, resulting in a multidimensional residual vector. Each dimension of the multidimensional residual vector corresponds to the deviation between the predicted and measured values ​​of the externally measurable data. A health index construction module is used to calculate the multidimensional residual vector to obtain a comprehensive health index, which is a normalized distance metric of the multidimensional residual vector relative to the distribution of the residual vector under healthy conditions. Finally, a fault identification module is used to match the multidimensional residual vector with a preset fault feature pattern library to identify specific fault types.

[0014] The present invention also provides an electronic device, comprising: a processor; and a memory storing a computer program, wherein the computer program, when executed by the processor, implements the method described above.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0016] The present invention, through the above technical solution, has the following beneficial effects: 1. Predictive maintenance is achieved: By constructing a multi-physics coupled digital model and combining it with a state estimation algorithm, this invention can estimate the internal implicit state of the fuel cell stack online. By analyzing the multi-dimensional residual vector between the model prediction and the actual data, it can identify the early trend of performance degradation before the external measurable parameters of the fuel cell stack deteriorate significantly. This realizes the transformation from "post-event alarm" to "predictive maintenance" and provides a time window for preventive intervention.

[0017] 2. Achieved transparent observation of internal state: This invention utilizes a state estimation algorithm to use readily available external measurable data as observation values ​​to perform real-time calibration and estimation of the implicit state inside the model (such as membrane water content, local temperature distribution, etc.). Thus, without increasing the cost of any hardware sensors, it achieves "transparent" observation of the key internal state of the fuel cell stack, providing a basis for understanding the operating status of the fuel cell stack.

[0018] 3. Achieves precise fault location: By matching the real-time calculated multidimensional residual vector with a preset fault feature pattern library, this invention can automatically identify and distinguish different fault types (such as membrane drying, flooding, etc.), providing precise cause location for subsequent maintenance decisions, which helps to shorten fault troubleshooting time and improve operation and maintenance efficiency.

[0019] 4. Possesses self-learning and model iteration capabilities: By marking and recording unknown abnormal patterns, the fault feature pattern library of this invention has the ability to learn and dynamically expand, and can continuously adapt to new and unforeseen fault patterns, thereby improving the robustness of the system and the accuracy of long-term applications. Attached Figure Description

[0020] Figure 1 The flowchart illustrates the online health assessment and fault early warning method provided by this invention.

[0021] Figure 2 A schematic diagram of the coupling relationship of the multiphysics coupling digital model provided by this invention.

[0022] Figure 3 This is a schematic diagram of the structure of the online health assessment and fault early warning system provided by the present invention. Detailed Implementation

[0023] To address the technical problem of existing air-cooled hydrogen fuel cell stack operation and maintenance relying on a "post-event alarm" model, which fails to achieve early warning and predictive maintenance, this invention provides a method, system, electronic device, and storage medium for online health assessment and fault early warning based on digital twins. This solution introduces digital twin technology to construct a virtual model that is synchronized in real time with the physical fuel cell stack, thereby achieving transparent observation of the stack's internal state and accurate prediction of future performance. Ultimately, this transforms the operation and maintenance model from passive "post-event repair" to proactive "predictive maintenance."

[0024] In the technical solution of this invention, for ease of understanding, some core terms are first defined. An air-cooled hydrogen fuel cell stack refers to a hydrogen fuel cell stack that uses air simultaneously as the oxidant in the cathode reaction and as the cooling medium during stack operation. Externally measurable data refers to physical quantities that can be directly collected by existing sensors installed on the physical stack, such as single-cell voltage, stack outlet temperature, gas flow rate, and pressure. These data form the data input basis for the method of this invention. A mechanism-based multiphysics coupling digital model refers to a mathematical model that deeply integrates the physical laws of multiple fields such as electrochemistry, thermodynamics, and fluid mechanics. It is not a simple black-box model, but a mechanism model that can describe the interaction of physicochemical processes inside the stack, used to accurately predict the behavior of the stack under specific operating conditions.

[0025] Furthermore, the internal implicit state refers to key state variables within the fuel cell stack that cannot be directly measured by conventional sensors but have a decisive impact on stack performance and lifespan. Examples include the water content distribution of the proton exchange membrane, the local current density distribution of the catalyst layer, and the fine temperature field distribution within the stack. By estimating these states, this invention can reveal the deep-seated health status of the fuel cell stack. The state estimation algorithm refers to an algorithm that optimally estimates the state (i.e., internal implicit state) of a dynamic system using incomplete or noisy observation data (i.e., externally measurable data), such as extended Kalman filtering or unscented Kalman filtering.

[0026] The multidimensional residual vector is a vector composed of multiple difference components between the model-predicted external measurable values ​​and the sensor-measured values. The norm and direction of this vector jointly characterize the degree and pattern of the physical fuel cell stack deviating from its ideal health state. The comprehensive health index is a single scalar value calculated based on the multidimensional residual vector, used to intuitively and quantitatively assess the degree to which the current overall health state of the fuel cell stack deviates from the ideal state. The fault feature pattern library is a pre-established database that stores characteristic multidimensional residual vector templates corresponding to various typical faults (such as membrane drying and flooding), providing a basis for rapid fault identification. Through the combination of the above technical means, this invention can effectively improve the intelligent operation and maintenance level of air-cooled hydrogen fuel cell stacks.

[0027] In one aspect of the invention, an online health assessment and fault prediction method for air-cooled hydrogen fuel cell stacks is provided. The core of this method lies in establishing a digital "twin" to mirror and predict the behavior of the physical fuel cell stack. Figure 1 A flowchart of the online health assessment and fault early warning method provided by this invention. The following is in conjunction with... Figure 1 This invention provides a detailed description of the online health assessment and fault early warning method for air-cooled hydrogen fuel cell stacks. The method includes: Step S101: Collect external measurable data characterizing the operating status of the fuel cell stack. This data serves as a bridge between the physical and digital worlds, and its accuracy and real-time performance directly affect the reliability of subsequent analyses. Specifically, the collectable data may include voltage monitoring data for each individual cell or battery pack, outlet temperature of each cathode cavity, inlet and outlet temperatures and flow rates of cooling air, inlet and outlet pressures of hydrogen, ambient temperature, and atmospheric pressure.

[0028] Step S102: The operating state of the fuel cell stack is predicted using a mechanism-based multiphysics coupled digital model. This model is the core of the entire system. Based on the collected operating data (such as load current, environmental conditions, etc.) and the internal state at the previous moment, it calculates and outputs the predicted values ​​of the internal state of the fuel cell stack and external measurable parameters at the next moment. Compared with purely data-driven black-box models, this prediction method based on physical laws has better generalization ability and interpretability, and can more accurately reflect the response of the fuel cell stack under different operating conditions.

[0029] Step S103: Employing a state estimation algorithm, the multiphysics coupled digital model is used as the prediction model, and the external measurable data is used as the observations. The prior estimate of the internal implicit state of the fuel cell stack is iteratively corrected by comparing the deviation between the predicted and observed values, thus estimating the internal implicit state of the fuel cell stack online. This process typically includes a prediction step and an update step. In the prediction step, the digital model extrapolates from the state at the previous moment to obtain the prior estimate of the current state. In the update step, the state estimation algorithm (such as Kalman filtering) uses the external measurable data (observations) collected at the current moment to calculate the deviation between the predicted and observed values, and iteratively corrects the prior estimate based on this deviation to obtain a more accurate posterior state estimate. Through this step, "soft measurement" of key internal states such as membrane water content and local temperature is achieved without adding any physical sensors, making the internal state of the fuel cell stack transparent.

[0030] For example, the state estimation algorithm can employ extended Kalman filtering or unscented Kalman filtering, or other filtering algorithms with recursive prediction-correction mechanisms. The extended Kalman filtering achieves state estimation by linearizing the nonlinear model, while the unscented Kalman filtering approximates the probability density distribution of the nonlinear function by constructing sampling points through unscented transformation.

[0031] Step S104: Calculate the difference between the external measurable value predicted by the multiphysics coupled digital model and the external measurable data to obtain a multidimensional residual vector. Each dimension of the multidimensional residual vector corresponds to the deviation between the predicted and measured values ​​of the external measurable data. Each dimension of the multidimensional residual vector corresponds to the deviation of a measurable parameter. For example, if there are N single-cell voltage measurement points and M temperature measurement points, the multidimensional residual vector contains at least N voltage residual components and M temperature residual components, corresponding to the deviations between the predicted and measured values ​​of each single-cell voltage and each temperature measurement point, respectively. The generation of residuals indicates a deviation between the actual behavior of the physical stack and its ideal health model, which is usually an early signal of performance degradation or failure. Therefore, the multidimensional residual vector is a direct input and key basis for subsequent health assessment and fault diagnosis.

[0032] Step S105: Calculate the multidimensional residual vector to obtain a comprehensive health index. The comprehensive health index is a normalized distance metric of the multidimensional residual vector relative to the distribution of the residual vector in a healthy state. Since the multidimensional residual vector contains multiple components, direct observation is complex; therefore, it needs to be transformed into a single, easily understood scalar index. This index quantifies the overall degree to which the physical fuel cell stack deviates from its health baseline by comparing the current multidimensional residual vector with the distribution of the residual vector in a healthy state and calculating their normalized distance. If the comprehensive health index continues to increase, it indicates that the health condition of the fuel cell stack is deteriorating and requires attention.

[0033] Step S106: Match the multidimensional residual vector with a preset fault feature pattern library to identify the specific fault type. Different faults (such as membrane drying, flooding, catalyst degradation, etc.) will cause specific changes in the physicochemical processes inside the fuel cell stack, thus leaving unique "fingerprints" or patterns on the multidimensional residual vector. By comparing the residual vector calculated in real time with various typical fault patterns stored in the library, the most likely fault type can be determined. This step achieves a leap from "detecting anomalies" to "locating the cause of anomalies," providing precise guidance for subsequent maintenance decisions.

[0034] In a preferred embodiment, the mechanism-based multiphysics coupled digital model includes at least an electrochemical sub-model, a thermal sub-model, and a fluid sub-model. These sub-models form a bidirectional coupling mechanism through feedback correlation between the heat generation rate and the temperature field, as well as the influence of temperature on the electrochemical reaction rate, and are solved iteratively within each solution step.

[0035] Figure 2 This is a schematic diagram illustrating the coupling relationships in the multiphysics coupled digital model provided by this invention. (See diagram below.) Figure 2 As shown, the three sub-models are interconnected and together describe the complex physicochemical processes inside the fuel cell stack. Specifically, the electrochemical sub-model 21, based on the Butler-Volmer equation, describes the electrochemical kinetics of the hydrogen oxidation reaction (HOR) at the anode and the oxygen reduction reaction (ORR) at the cathode. It is responsible for calculating the activation overpotential generated by the electrochemical reaction, the ohmic overpotential generated by ion and electron conduction, and the concentration overpotential generated by the limited mass transfer of reactants. These factors together determine the output voltage of the battery.

[0036] Based on the above model, thermal model 22 is responsible for describing the generation and transfer of heat inside the fuel cell stack. The heat sources mainly include the heat of irreversible electrochemical reactions and the ohmic heat generated when current flows through resistors. The calculation of these heat sources depends on the current density and potential loss output by electrochemical sub-model 21. Heat transfer includes heat conduction through solid components such as bipolar plates and membrane electrodes, convective heat dissipation as cooling air flows through the cathode channel, and radiative heat dissipation from the fuel cell surface to the environment. The output of thermal model 22 is the temperature field distribution inside the fuel cell stack.

[0037] Meanwhile, the fluid sub-model 23 is responsible for describing the flow and distribution of reactant gases and cooling air within the fuel cell stack. For air-cooled fuel cell stacks, this model pays particular attention to the flow distribution and pressure drop of cooling air in the parallel cathode channels, as well as the pressure distribution of hydrogen in the anode channels. The flow state of the gases directly affects the supply of reactants and the discharge of product water, which in turn affects the concentration overpotential in the electrochemical sub-model 21.

[0038] These three sub-models achieve bidirectional coupling and joint solution through a chain of heat generation, temperature, and reaction rate. Specifically, the heat generated by the electrochemical sub-model 21 is the input to the thermal sub-model 22; and the temperature distribution calculated by the thermal sub-model 22, in turn, affects the electrochemical reaction rate in the electrochemical sub-model 21 (because the reaction rate constant is a function of temperature). Similarly, the pressure and flow distribution calculated by the fluid sub-model 23 affects the mass transport of the electrochemical reaction, while the consumption of reactants by the electrochemical reaction affects the flow field. Through this tight bidirectional coupling, the model can more realistically simulate the complex behavior of multiphysics interactions within the fuel cell stack, thereby improving the accuracy of predictions.

[0039] Furthermore, in the step of matching the multidimensional residual vector with the fault feature pattern library, a cosine similarity algorithm can be specifically used to calculate the similarity between the multidimensional residual vector and each pattern in the fault feature pattern library. The cosine similarity algorithm measures the directional consistency of two vectors by calculating the cosine of the angle between them in multidimensional space. Its value ranges from [-1, 1], and the closer the value is to 1, the more consistent the directions of the two vectors. In fault diagnosis scenarios, the direction of the multidimensional residual vector is more representative of the fundamental fault pattern than its magnitude (norm). For example, membrane drying may cause the battery voltage residual to be positive at certain locations and negative at others, forming a specific direction vector. Using cosine similarity can effectively capture this pattern, thus being insensitive to the overall magnitude of the residual, thereby improving the robustness of fault identification.

[0040] The pre-establishment method for the fault characteristic mode library is as follows: First, under the known healthy condition of the fuel cell stack, normal operation data for a sufficient duration is collected to establish a health baseline. Then, fuel cell stack operation data under various typical fault conditions are obtained through offline simulation or accelerated aging experiments in the laboratory. For each fault type, at least three sets of data samples with different severity and operating conditions are collected.

[0041] Specifically, membrane drying failure can be simulated by continuously operating the fuel cell stack under high-temperature and low-humidity conditions. High temperature refers to an ambient temperature above the upper limit of the fuel cell stack's suitable operating temperature (e.g., above 30°C), and low humidity refers to a humidity condition where the water vapor content in the ambient air is insufficient to maintain the water balance within the membrane, resulting in an evaporation rate exceeding the formation rate of water within the membrane. Flooding failure can be simulated by continuously operating the fuel cell stack under low-temperature and high-humidity conditions. Low temperature refers to a temperature above freezing point but below room temperature, where the water vapor content in the ambient air is sufficient to cause the evaporation rate of water generated inside the fuel cell stack to be significantly lower than the accumulation rate. In practical applications, the temperature can be set, for example, to a range above 0°C and below 15°C to ensure that the water generated inside the fuel cell stack exists in liquid form and can accumulate. Since the cathode of an air-cooled fuel cell stack has an open structure, the cathode side is directly exposed to the ambient air during operation, and the water content within the membrane is directly affected by the temperature and humidity of the ambient air. The above methods can accurately simulate membrane drying and flooding failures caused by excessively low or high ambient humidity during actual operation.

[0042] After obtaining various types of fault data, the multidimensional residual vector for each fault state is calculated according to steps S101 to S104 above. For each fault type, stratified sampling is performed from all residual vector samples of that fault type, based on operating conditions (load current, environmental conditions), with 3-5 representative samples selected from each stratum. The average value of the selected samples is taken to obtain the feature template vector for that fault type. Simultaneously, the standard deviation of the residual vector is calculated in each stratum of samples for uncertainty assessment in subsequent similarity determination.

[0043] Furthermore, the similarity threshold is set as follows: Based on the normal residual vector samples collected under the above-mentioned healthy state, the cosine similarity distribution between each healthy sample and any healthy template is calculated. Statistical results show that the minimum similarity under the healthy state is approximately 0.75. Considering the balance between false alarm rate and false negative rate, and after multiple sets of cross-validation experiments, when the similarity threshold is set to 0.8, the system can achieve a detection rate of over 95% for known faults, while the false alarm rate is less than 5%. Therefore, in practical applications, the preset similarity threshold can be selected as 0.8.

[0044] To further verify the effectiveness of the similarity threshold setting, leave-one-out cross-validation was used to quantitatively evaluate the diagnostic performance of the fault feature pattern library. Specifically, data samples for each fault type were randomly divided into a training set (70%) and a test set (30%), and the validation was repeated 10 times. The results showed that when the similarity threshold was set to 0.8, the system's average detection rate for membrane drying faults was 97.2% (standard deviation ± 1.8%), for flooding faults it was 95.1% (standard deviation ± 2.3%), and for catalyst degradation faults it was 93.8% (standard deviation ± 3.1%), with an overall average detection rate of 96%. Meanwhile, the average false alarm rate of healthy samples being misclassified as faults was 3.8% (standard deviation ± 0.9%). When the threshold was reduced to 0.7, the detection rate increased to 98.5%, but the false alarm rate increased to 9.6%. When the threshold was increased to 0.9, the false alarm rate decreased to 1.2%, but the detection rate decreased to 88.4%. To balance the detection rate and false alarm rate, 0.8 is the optimal threshold.

[0045] The fault feature pattern library can pre-store templates for various typical faults, such as membrane drying templates and flooding templates. When the system calculates the real-time residual vector R_live, it will calculate its cosine similarity with all templates in the library. The calculation formula is as follows: Similarity(R_live, T_i) = (R_live · T_i) / (||R_live|| ||T_i||); Here, Similarity(R_live, T_i) is the cosine similarity, a function used to measure the similarity between the real-time residual vector and the i-th fault template, where T_i is the i-th fault template. If a calculated similarity value exceeds a preset threshold (e.g., 0.8), the system determines that a corresponding fault has occurred and triggers an alert. This pattern-matching-based diagnostic method, compared to simple threshold judgment, can provide more specific and accurate fault location information.

[0046] Optionally, the calculation of the multidimensional residual vector to obtain the comprehensive health index is performed using the Mahalanobis distance calculation method.

[0047] To transform a high-dimensional vector into an intuitive scalar indicator, the Mahalanobis distance method can be used to calculate the comprehensive health index H. The specific calculation formula is as follows: H = sqrt((x - μ)^T Σ^(-1) (x - μ)); Where x is the currently calculated multidimensional residual vector, μ is the mean vector of the residual vector under healthy conditions (theoretically a zero vector), T is the vector transpose, and Σ is the covariance matrix of the residual vector under healthy conditions. Compared with simple Euclidean distance, Mahalanobis distance effectively considers the correlation between residual components and eliminates the influence of different physical dimensions by introducing the inverse Σ^(-1) of the covariance matrix. For example, voltage residuals and temperature residuals may have an inherent correlation, and Mahalanobis distance can capture this joint change, thus more realistically reflecting the degree of deviation of the overall health status of the system. A larger comprehensive health index value indicates a more severe deviation of the actual fuel cell stack from its health baseline model, i.e., a worse health status.

[0048] The externally measurable data provides the foundation for subsequent state estimation and residual calculation. This data consists of physical quantities that can be directly collected by sensors already installed on the fuel cell stack without incurring additional costs. As a specific implementation, the externally measurable data may include: voltage monitoring data of a single cell or a battery pack composed of several single cells connected in series; the temperature at the outlet of each cathode cavity of the fuel cell stack; the inlet and outlet temperatures and flow rates of cooling air; the inlet and outlet pressures of hydrogen; and at least two of the following parameters: ambient temperature and pressure. Collecting as much externally measurable data as possible provides more observational dimensions for the state estimation algorithm, thereby improving the accuracy and robustness of estimating the implicit state within the fuel cell stack.

[0049] As an effective means of intuitively managing health status, the method may further include setting at least two thresholds based on the comprehensive health index and classifying the health status of the fuel cell stack into different levels. For example, a concern threshold and an alarm threshold can be set. When the calculated comprehensive health index is less than the concern threshold, the fuel cell stack is determined to be in a "healthy" state; when the comprehensive health index is between the concern threshold and the alarm threshold, the fuel cell stack is determined to be in a "concerned" state; when the comprehensive health index is greater than or equal to the alarm threshold, it is determined to be in an "alarm" state.

[0050] This tiered management mechanism provides a window of opportunity for implementing predictive maintenance. When a fuel cell stack enters a "watchful" state, although its external performance (such as total voltage and total power) may not yet show a significant decline, the method of this invention has already detected subtle signs of degradation through residual analysis. At this point, maintenance personnel can intervene in advance to inspect, adjust operating strategies, or arrange maintenance plans, thereby preventing further deterioration of the fault and avoiding unplanned downtime and more serious equipment damage.

[0051] To enhance the system's self-learning and adaptive capabilities, the method may further include a mechanism for handling unknown anomalies. Specifically, when the comprehensive health indicator indicates a deterioration in the state (e.g., the comprehensive health indicator value continues to rise and exceeds a concern threshold), but cosine similarity calculation reveals that the similarity between the multidimensional residual vector and all known patterns in the fault feature pattern library is below a preset threshold, the system will not simply ignore this anomaly. Instead, it will mark the current multidimensional residual vector as an unknown anomaly pattern and record and store it.

[0052] This design enables the system to identify and capture new, unexpected failure modes. These recorded unknown anomalies, after manual analysis and annotation, can be added back to the failure characteristic mode library, continuously expanding and improving the library. In this way, the system possesses the ability for continuous learning and iterative optimization; its diagnostic range and accuracy improve with operating time, better adapting to different fuel cell stacks and constantly changing operating environments.

[0053] The present invention also provides an online health assessment and fault early warning system for air-cooled hydrogen fuel cell stacks, which is a hardware or software-hardware combination implementation of the above method. Figure 3 This is a schematic diagram of the online health assessment and fault early warning system provided by the present invention. Figure 3 As shown, the system includes a data acquisition module 10, a model prediction module 20, a state estimation module 30, a residual calculation module 40, a health indicator construction module 50, and a fault identification module 60.

[0054] Specifically, the data flow path of the system is as follows: First, the data acquisition module 10 is responsible for collecting external measurable data characterizing the stack's operating state from various sensors of the operating air-cooled hydrogen fuel cell stack in real time or periodically. This data is then distributed to two downstream modules: one part is input as observations to the state estimation module 30 for calibrating and updating the model state; the other part is input to the residual calculation module 40 for comparison with the model predictions.

[0055] The model prediction module 20 internally runs a mechanism-based multiphysics coupled digital model, which continuously predicts the future state of the fuel cell stack based on the current operating condition input and the estimated internal state. The output of the model prediction module 20 is also divided into two paths: the predicted internal implicit state of the fuel cell stack is sent to the state estimation module 30 as the state prediction value; the predicted external measurable value is sent to the residual calculation module 40.

[0056] After receiving the model-predicted state and the measured data from the data acquisition module 10, the state estimation module 30 executes a state estimation algorithm. It iteratively corrects the prior estimate of the internal implicit state of the fuel cell stack by comparing the deviation between the predicted and observed values. By fusing the two sets of information, it outputs a corrected and updated estimate of the internal implicit state that is closer to the actual situation. The updated state is then fed back to the model prediction module 20 for the next round of prediction, forming a closed-loop correction mechanism.

[0057] After receiving the model prediction value and the sensor measured value, the residual calculation module 40 calculates the difference between the two to obtain a multidimensional residual vector. Each dimension of the multidimensional residual vector corresponds to the deviation between the predicted value and the measured value of the external measurable data.

[0058] The health indicator construction module 50 calculates a single comprehensive health indicator based on the multidimensional residual vector output by the residual calculation module 40 through specific mathematical transformations. The comprehensive health indicator is a normalized distance metric of the multidimensional residual vector relative to the distribution of the residual vector in a healthy state, reflecting the degree of health deterioration of the fuel cell stack.

[0059] Finally, after receiving the multidimensional residual vector, the fault identification module 60 compares and matches it with a preset fault feature pattern library. Once a match is successful, the specific fault type can be identified, and corresponding early warning information can be output. Through a complete data processing chain, the system realizes continuous monitoring, quantitative assessment, and early and accurate early warning of faults in the stack health status.

[0060] These modules can be independent hardware units or different software processes or threads running on the same processor. They communicate through an internal data bus or network interface to work together to complete all the functions of online health assessment and fault early warning.

[0061] Furthermore, this invention also provides an electronic device, which can be a dedicated battery management system (BMS), an on-board diagnostic unit (OBD), a monitoring computer for a UAV ground station, or a cloud server. The electronic device internally includes a processor and a memory. The memory stores a computer program containing instructions for implementing any of the aforementioned methods. When the processor executes these instructions, it can complete a series of steps, including data acquisition, model prediction, state estimation, residual calculation, health indicator construction, and fault identification, thereby achieving online health assessment and fault early warning for air-cooled hydrogen fuel cell stacks.

[0062] The present invention also provides a computer-readable storage medium on which a computer program is stored. The storage medium can be non-volatile, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), or it can be volatile, such as random access memory (RAM). When the computer program is loaded and executed by a processor, it can implement the methods described in any of the above-mentioned embodiments. This indicates that the technical solution of the present invention can be implemented and deployed on a general-purpose computing platform through software programming.

[0063] To more clearly illustrate the technical solution of the present invention, a detailed description is provided below with reference to specific embodiments. In one embodiment, the method of the present invention is applied to a common online monitoring scenario of an air-cooled hydrogen fuel cell stack. First, external measurable data of the stack is continuously collected through the sensor interface on the stack controller area network (CAN) bus, for example, the voltage of all individual cells and the temperature data of several temperature measurement points are collected every 100 milliseconds.

[0064] Next, a pre-built multiphysics coupled digital model predicts the voltage of each cell and the temperature at the measurement point at the next moment (100 milliseconds later) based on the currently collected load current and ambient temperature information. Simultaneously, a pre-built, "black box" state estimation algorithm receives the predicted state from the digital model and the measured voltage and temperature data from the sensors. After internal calculations, it outputs a set of estimated values ​​about the implicit state inside the stack, such as the average membrane water content of each membrane electrode.

[0065] The system then calculates the differences between the voltage and temperature predicted by the model and the voltage and temperature measured by the sensors, and combines these differences into a multidimensional residual vector. For example, if there are 100 single-cell voltage and 10 temperature measurement points, a 110-dimensional residual vector will be formed. Subsequently, the system performs a certain preset mathematical transformation on this 110-dimensional residual vector, such as calculating its norm or weighted sum, to obtain a single health score representing the current health status.

[0066] Finally, the system compares the 110-dimensional residual vector with a pre-defined fault feature pattern library. This library stores several 110-dimensional vectors, each representing a known fault mode. Using a comparison algorithm, the system identifies the fault mode most similar to the current real-time residual vector and outputs its corresponding fault label, such as "suspected flooding" or "decreased catalyst activity." Through this series of steps, even using a relatively simplified model and algorithm, continuous monitoring of the stack's health status and preliminary judgment of basic fault types are achieved.

[0067] Based on the above embodiments, the application of the present invention in early warning of membrane drying in UAV fuel cell stacks is further described. This scenario involves a 12kW air-cooled hydrogen fuel cell stack equipped with an H100 hydrogen-powered UAV. The system employs a high-precision multiphysics coupled digital model that couples electrochemical sub-model 21, thermal sub-model 22, and fluid sub-model 23. This model can accurately simulate the hydrothermal management process inside the fuel cell stack.

[0068] After approximately 800 hours of cumulative operation, the system detected that the overall health index had slowly but steadily increased from a stable baseline of 0.15 to 0.42 over the last 50 hours. This value had exceeded the preset concern threshold of 0.3 but had not yet reached the alarm threshold of 0.7. Therefore, the system updated the stack status from "healthy" to "concerned". It should be noted that at this time, the total stack voltage only decreased from the initial 58.2V to 57.9V, a decrease of approximately 0.5%, far below the 5% alarm threshold set in traditional BMS systems. This indicates that this method can achieve early warning when there are no significant anomalies in external macroscopic parameters. The time interval from the overall health index first exceeding the concern threshold to the traditional voltage threshold alarm triggering is approximately 2.3 hours (i.e., an early warning lead time of approximately 138 minutes).

[0069] The system further analyzes the multidimensional residual vector that triggers the "attention" state. The analysis shows that the voltage residual components corresponding to cells 8 through 12 of the fuel cell stack are consistently positive, indicating that the actual output voltage of these cells is consistently lower than the voltage predicted by the health model. Simultaneously, the cathode cavity outlet temperature residuals adjacent to these cells exhibit a specific non-uniform distribution. The real-time residual vector R_live is matched with a fault feature pattern library, and the system uses a cosine similarity algorithm for calculation.

[0070] Calculation results showed that the cosine similarity between the real-time residual vector R_live and the pre-stored membrane drying template in the library was 0.89, exceeding the preset similarity threshold of 0.8. Based on this high similarity matching result, the system automatically triggered an "early warning of membrane drying" and pushed detailed diagnostic information to the maintenance personnel. Following the warning, the maintenance personnel promptly implemented online recovery measures, such as appropriately reducing the drone's flight load power and optimizing the heat dissipation conditions of the fuel cell stack (e.g., increasing cooling airflow), to lower the internal temperature of the stack and reduce the rate of moisture evaporation, thereby mitigating the membrane drying trend. Under these interventions, the comprehensive health index successfully dropped to 0.20 within the following 10 hours, indicating that the membrane drying trend was effectively curbed, avoiding potential permanent damage to the proton exchange membrane. Disassembly and testing confirmed that the proton exchange membrane of the fuel cell stack did not show any thinning or pinhole defects, thus verifying the timeliness of the warning and the effectiveness of the intervention measures.

[0071] If the overall health index also rises to 0.42, but the cosine similarity of its residual vector with all known templates in the library (including membrane drying templates, flooding templates, etc.) is below 0.8, the system will trigger an "unknown anomaly mode" alarm. Simultaneously, the system will automatically save the current operating data and the corresponding multidimensional residual vector. This saved "unknown anomaly" data package can be used by technical experts for offline analysis. Once confirmed as a new and valuable precursor to a fault, its features can be extracted and used to create a new template, which will be added to the fault feature pattern library, enabling the system to identify such anomalies in the future.

[0072] This invention is not limited to monitoring individual fuel cell stacks, but can be extended to the health management of a group of stacks. The system of this invention was applied to a photovoltaic hydrogen production-drone inspection project, which deployed five H100 drones carrying five identical 12kW air-cooled hydrogen fuel cell stacks. The system established and ran five independent, personalized digital twin models for each of these five physical stacks. This personalized management can adapt to individual differences in each stack caused by variations in manufacturing tolerances and operational history.

[0073] After running for a period of time, the system conducted a comparative analysis of the evolution trend of the comprehensive health index H of the five fuel cell stacks. The analysis revealed that the average growth rate of the health index H of stack #003 reached 0.008 / hour, significantly higher than the average growth rate of the other four stacks (approximately 0.003 / hour). This abnormally rapid degradation rate led the system to automatically identify stack #003 as a high-risk individual requiring close monitoring.

[0074] To trace the root cause of the rapid degradation of the #003 fuel cell stack, the system further analyzed its recent multidimensional residual vector. Matching with a fault characteristic pattern library revealed a high correlation between the residual vector and the template of the "catalyst degradation" mode. Based on this diagnostic result, the maintenance team retrieved the historical operating data archive of the #003 fuel cell stack, which was also recorded by this system. Data tracing revealed that the stack had repeatedly operated at near-full power (e.g., power > 65°C) for extended periods during summer when ambient temperatures exceeded 35°C. High temperature and high load are typical contributing factors to accelerated catalyst (especially platinum) agglomeration and loss, leading to activity degradation.

[0075] Based on this series of analyses involving "data comparison, risk identification, root cause localization, and historical tracing," the operations and maintenance team developed a targeted maintenance strategy for fuel cell stack #003: on the one hand, proactively reducing its operating power limit by 15% to slow its degradation rate; on the other hand, including it in the priority maintenance plan, preparing for early stack replacement. More importantly, this finding provided a basis for optimizing the operational strategy of the entire drone fleet. The team updated the operating guidelines for all fuel cell stacks under high-temperature weather conditions, limiting the maximum output power. Through this group management and proactive intervention based on digital twins, the average lifespan of the entire fuel cell stack group was extended, and the overall reliability of the fleet was improved.

Claims

1. A method for online health assessment and fault early warning of air-cooled hydrogen fuel cell stacks, characterized in that, include: Collect externally measurable data characterizing the operating state of the fuel cell stack; The operating state of the fuel cell stack is predicted using a mechanism-based multiphysics coupled digital model. A state estimation algorithm is used, with the multiphysics coupled digital model as the prediction model and the external measurable data as the observation value. The prior estimate of the internal hidden state of the electric stack is iteratively corrected by comparing the deviation between the prediction value and the observation value, and the internal hidden state of the electric stack is estimated online. The difference between the external measurable value predicted by the multiphysics coupled digital model and the external measurable data is calculated to obtain a multidimensional residual vector. Each dimension of the multidimensional residual vector corresponds to the deviation between the predicted value and the measured value of the external measurable data. The multidimensional residual vector is calculated to obtain a comprehensive health index, which is a normalized distance metric of the multidimensional residual vector relative to the distribution of the residual vector under the health state. The multidimensional residual vector is matched with a preset fault feature pattern library to identify the specific fault type.

2. The method according to claim 1, characterized in that, The mechanism-based multiphysics coupled digital model includes at least an electrochemical sub-model, a thermal sub-model, and a fluid sub-model. The sub-models form a bidirectional coupling mechanism through the feedback relationship between the heat generation rate and the temperature field, as well as the influence of temperature on the electrochemical reaction rate, and are solved simultaneously and iteratively within each solution step.

3. The method according to claim 1, characterized in that, The step of matching the multidimensional residual vector with the fault feature pattern library includes: The cosine similarity algorithm is used to calculate the similarity between the multidimensional residual vector and each pattern in the fault feature pattern library.

4. The method according to claim 1, characterized in that, The calculation of the multidimensional residual vector to obtain the comprehensive health index uses the Mahalanobis distance calculation method, and the specific formula is as follows: H = sqrt((x - μ)^T Σ^(-1) (x - μ)); Where H is the comprehensive health index, x is the currently calculated multidimensional residual vector, μ is the mean vector of the residual vector under the health state, T is the vector transpose, and Σ is the covariance matrix of the residual vector under the health state.

5. The method according to claim 1, characterized in that, The externally measurable data includes at least two of the following: single cell voltage, stack outlet temperature, gas flow rate, gas pressure, and environmental parameters.

6. The method according to claim 1, characterized in that, The method further includes: At least two thresholds are set based on the comprehensive health indicators, and the health status of the fuel cell stack is divided into healthy, concern, and alarm levels.

7. The method according to claim 6, characterized in that, The method further includes: When the comprehensive health indicator indicates a deterioration in the condition, but the similarity between the multidimensional residual vector and all patterns in the fault feature pattern library is lower than a preset threshold, the multidimensional residual vector is marked as an unknown abnormal pattern and recorded.

8. An online health assessment and fault early warning system for air-cooled hydrogen fuel cell stacks, characterized in that, include: The data acquisition module is used to collect external measurable data that characterizes the operating status of the fuel cell stack; The model prediction module is used to predict the operating state of the fuel cell stack based on a mechanism-based multiphysics coupled digital model. The state estimation module is used to employ a state estimation algorithm, taking the multiphysics coupled digital model as the prediction model and the external measurable data as the observation value, and iteratively correcting the prior estimate of the internal hidden state of the electric stack by comparing the deviation between the prediction value and the observation value, thereby estimating the internal hidden state of the electric stack online. The residual calculation module is used to calculate the difference between the external measurable value predicted by the multiphysics coupled digital model and the external measurable data, and obtain a multidimensional residual vector. Each dimension of the multidimensional residual vector corresponds to the deviation between the predicted value and the measured value of the external measurable data. A health index construction module is used to calculate the multidimensional residual vector to obtain a comprehensive health index, wherein the comprehensive health index is a normalized distance metric of the multidimensional residual vector relative to the distribution of the residual vector under the health state. The fault identification module is used to match the multidimensional residual vector with a preset fault feature pattern library to identify the specific fault type.

9. An electronic device, characterized in that, include: processor; And a memory, on which a computer program is stored, which, when executed by the processor, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images