Fault Diagnosis and Prediction System for Hydrogen Energy Cogeneration Unit Based on Deep Learning

Abstract

The present disclosure discloses a fault diagnosis and prediction system for a hydrogen energy combined heat and power unit based on deep learning, which relates to the technical field of fault diagnosis. The system includes: a data acquisition unit, a data analysis and feature extraction unit, and a fault prediction unit; the data acquisition unit is used to collect the operation data of the hydrogen energy combined heat and power unit in real time; the data analysis and feature extraction unit is used to perform electrochemical performance analysis based on polarization curves according to the operation data, establish a polarization model of the hydrogen energy combined heat and power unit; calculate the electrical efficiency and thermal efficiency of the hydrogen energy combined heat and power unit, and construct a fault feature vector of the hydrogen energy combined heat and power unit; the fault prediction unit is used to judge the probability of various types of faults occurring in the next time step. The present invention can not only adaptively capture the dynamic changes during the operation of the unit, but also can early warn potential faults, thereby greatly improving the safety, stability and operation efficiency of the system.

Description

Technical Field

[0001] The present disclosure relates to the technical field of fault diagnosis, and particularly to a fault diagnosis and prediction system for a hydrogen energy co-generation unit based on deep learning. Background Art

[0002] As a clean and efficient energy carrier, hydrogen energy has occupied an important position in the global energy structure transformation in recent years. The hydrogen energy co-generation unit is based on fuel cell technology and can provide both electric energy and heat energy. While improving the energy utilization efficiency, it reduces carbon emissions and is widely used in distributed energy systems, industrial waste heat recovery, household energy supply and other scenarios. As the core component of the hydrogen energy co-generation unit, the operation stability and fault diagnosis ability of the fuel cell directly affect the overall performance of the unit. However, the operation of the fuel cell is affected by various factors, including electrochemical reactions, gas transport, catalyst activity, thermal management systems, and environmental changes. Therefore, how to achieve efficient and intelligent fuel cell fault diagnosis and prediction has become one of the key technologies to ensure the stability and lifespan of the hydrogen energy co-generation unit.

[0003] Currently, the fault diagnosis and prediction methods for fuel cells mainly include methods based on physical models, methods based on signal processing, and methods based on machine learning. Among them, the methods based on physical models mainly rely on the electrochemical equations and kinetic characteristics of fuel cells, such as polarization curve models, equivalent circuit models, etc. For example, the Nernst equation and the Tafel equation can be used to describe the open-circuit voltage and activation polarization loss of fuel cells, Ohm's law can be used to characterize the ohmic polarization of fuel cells, and Fick's law and the Butler-Volmer equation can be used to calculate the transport characteristics of the gas diffusion layer. Such methods can accurately describe the operation characteristics of fuel cells, but their main problems are high computational complexity and strong dependence on model parameters. In practical applications, due to material aging, pollutant accumulation, and environmental factor changes in fuel cells, the parameters in physical models often need to be continuously adjusted, resulting in poor real-time performance and adaptability of fault diagnosis. Summary of the Invention

[0004] The object of the present invention is to provide a fault diagnosis and prediction system for a hydrogen energy combined heat and power unit based on deep learning. Aiming at the problems of diverse fault types, complex fault mechanisms, and difficulty in timely prediction during the operation and maintenance of the hydrogen energy combined heat and power unit, through the integration of multi-dimensional feature construction, deep learning modeling, and time series analysis, an efficient, accurate, and self-adaptive fault diagnosis and prediction system is proposed. It overcomes the deficiencies of traditional technologies in feature selection, timeliness of fault identification, and model scalability, enabling the unit to obtain reliable operation and maintenance guarantees at lower costs and higher efficiency in practical applications. In addition, the system can also issue risk warnings before major faults occur in the unit, transforming the operation and maintenance strategy from passive response to active protection.

[0005] The technical solution of the present invention is implemented as follows:

[0006] A fault diagnosis and prediction system for a hydrogen energy combined heat and power unit based on deep learning, the system includes: a data acquisition unit, a data analysis and feature extraction unit, and a fault prediction unit; the data acquisition unit is used to collect the operation data of the hydrogen energy combined heat and power unit in real time; the data analysis and feature extraction unit is used to perform electrochemical performance analysis based on polarization curves according to the operation data, establish a polarization model of the hydrogen energy combined heat and power unit; calculate the electrical efficiency and thermal efficiency of the hydrogen energy combined heat and power unit according to the polarization model, and then use a deep learning model to construct a fault feature vector of the hydrogen energy combined heat and power unit; the fault prediction unit is used to judge the probability of various types of faults occurring in the next time step according to the fault feature vector and the feature template vectors of various types of set faults.

[0007] Further, the types of the faults include: membrane dehydration fault, membrane flooding fault, catalyst poisoning fault, gas diffusion layer blockage fault, and thermal management system fault.

[0008] Further, the operation data of the hydrogen energy combined heat and power unit collected in real time by the data acquisition unit includes: single cell voltage in V 、load current in A 、stack temperature in K 、hydrogen pressure in Pa 、oxygen pressure in Pa 、hydrogen flow rate in mol / s 、oxygen stoichiometric ratio 、coolant inlet temperature in K 、coolant outlet temperature in K 、fuel cell operating temperature in K 、water vapor partial pressure in Pa 、ohmic resistance in Ω and the recovered heat in watts ; wherein, is the time.

[0009] Furthermore, the polarization model is represented by the following formula:

[0010] ;

[0011] wherein, is the standard electrode potential in volts (V); is the gas constant, with a value of 8.314 J / (mol·K); is the Faraday constant, with a value of 96485 C / mol; is the transfer coefficient, with a value ranging from 0.3 to 0.7; is the exchange current density in amperes per square centimeter (A / cm²); is the first concentration polarization parameter in volts (V), with a value ranging from 0.02 to 0.08; is the second concentration polarization parameter in square centimeters per ampere (cm² / A), with a value ranging from 10 to 30.

[0012] Furthermore, the exchange current density is calculated using the following formula:

[0013] ;

[0014] wherein, is the influence exponent of hydrogen partial pressure on the exchange current density, with a value ranging from 0.7 to 1.2; is the influence exponent of oxygen partial pressure on the exchange current density, with a value ranging from 0.3 to 0.5; is the reference pressure, with a value of 1 atm or Pa; is the activation energy in joules per mole (J / mol); is the pre-exponential factor of the exchange current density, which reflects the magnitude of the exchange current density under ideal conditions without an activation energy barrier, in amperes per square centimeter (A / cm²), with a value ranging from to .

[0015] Furthermore, the transfer coefficient is calculated using the following formula:

[0016] ;

[0017] wherein, is the reorganization energy, representing the energy adjustment at the interface between the catalyst and the electrolyte during charge transfer, in electron volts (eV), with a value ranging from 0.2 to 0.6, where 1 eV = 96.5 kJ / mol; is the activation free energy, defined as the minimum Gibbs free energy required for the reaction to proceed, with the unit of eV and the value range from 0.1 to 0.4; is the electron transfer coefficient, related to the charge transfer rate, with the value range from 0.5 to 0.7; is the surface coverage; the surface coverage is calculated using the following formula:

[0018] ;

[0019] where, is the hydrogen adsorption equilibrium constant, , where is the adsorption free energy of hydrogen on the catalyst surface, with the unit of eV and the value range from -0.3 to -0.1.

[0020] Furthermore, the electrical efficiency and the thermal efficiency are calculated using the following formulas respectively:

[0021] ;

[0022] ;

[0023] where, is the calorific value of hydrogen.

[0024] Furthermore, the process of constructing the fault feature vector of the hydrogen energy cogeneration unit using the deep learning model specifically includes: first constructing the following initial fault vector:

[0025] ;

[0026] where, is the reference voltage; is the reference electrical efficiency; is the reference thermal efficiency; then using the single-layer neural network in deep learning to extract and process to obtain the fault feature vector .

[0027] Furthermore, the following formula is used to calculate the probability of each type of fault occurring at each time step:

[0028] ;

[0029] where, represents the probability of fault occurring at time ; represents the time step; With faults The Mahalanobis distance of the feature template vector; Represents the fault feature vector With faults The square of the Euclidean distance between the feature template vectors; Is the set historical time window; Is the integral variable; And Are both integer subscript indices.

[0030] The fault diagnosis and prediction system of the hydrogen energy cogeneration unit based on deep learning of the present invention has the following beneficial effects:

[0031] At the level of fault feature construction, the present invention introduces multiple indicators closely related to the electrochemical and thermal management processes of the unit, and normalizes and differentiates these indicators, thereby realizing the comprehensive characterization of the full operating conditions of the unit. In this process, not only the dynamic characteristics of core processes such as charge transfer and gas diffusion in the fuel cell stack are considered, but also the regulating effect of the cooling system on the temperature distribution is fully taken into account. This multi-factor integration makes the fault features more representative and distinguishable.

[0032] The present invention introduces the feature extraction mechanism in the deep learning model, and performs non-linear mapping on the constructed multi-dimensional fault vector through a single-layer neural network, so that the differences of different fault modes are retained in the high-dimensional space. Compared with the traditional feature engineering method, the present invention pays more attention to the integration of data-driven and principles, relatively simplifies the setting process of artificial rules, and makes the system more extensible when facing complex or unknown fault types. Through such a feature extraction strategy, the system can utilize the multi-dimensional sensor data continuously obtained during actual operation for self-learning and self-updating, avoiding diagnostic errors caused by environmental changes or operating condition switches.

[0033] In terms of fault discrimination and prediction, the present invention can accurately evaluate the similarity between each known fault mode and the current operating state in the multi-dimensional feature space by setting feature templates of multiple known fault modes and combining the distance measurement method of feature vectors. Different from the traditional pure rule comparison method, the present invention not only considers the matching degree between the current fault feature and the template, but also introduces a correction term of time evolution. Such an approach can effectively cope with the progressive characteristics of fuel cell faults. Whether it is catalyst deactivation, membrane state abnormality, gas diffusion layer blockage, or potential faults in the thermal management link, early signals can be captured through continuous monitoring and trend analysis of the time series. While considering historical data, the present invention takes into account short-term fluctuations and long-term evolution, providing a more reliable basis for the fault prediction of the unit, thereby helping the operation and maintenance personnel to take intervention measures before the large-scale outbreak of faults. Brief Description of the Drawings

[0034] Figure 1 This is a schematic diagram of the system structure of the fault diagnosis and prediction system for a hydrogen energy co-generation unit based on deep learning provided by an embodiment of the present invention. Detailed implementation manners

[0035] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present disclosure clearer and more understandable, the present disclosure will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure and are not used to limit the present disclosure.

[0036] Embodiment 1, refer to Figure 1 : A fault diagnosis and prediction system for a hydrogen energy co-generation unit based on deep learning, the system includes: a data acquisition unit, a data analysis and feature extraction unit, and a fault prediction unit; the data acquisition unit is used to collect the operation data of the hydrogen energy co-generation unit in real time; the data analysis and feature extraction unit is used to perform electrochemical performance analysis based on the polarization curve according to the operation data, establish a polarization model of the hydrogen energy co-generation unit; according to the polarization model, calculate the electrical efficiency and thermal efficiency of the hydrogen energy co-generation unit, and then use a deep learning model to construct a fault feature vector of the hydrogen energy co-generation unit; the fault prediction unit is used to judge the probability of various types of faults occurring in the next time step according to the fault feature vector and the feature template vectors of various types of faults set.

[0037] Specifically, during the operation of a hydrogen energy co-generation unit, the electrochemical reaction process of the fuel cell stack is strongly affected by the supply states of hydrogen and oxygen. The temperature distribution, gas diffusion, catalyst activity, membrane hydration state, and load changes within the fuel cell stack all play important roles in the final output voltage, power, and the overall system's thermoelectric conversion efficiency. Therefore, the data acquisition unit must be able to accurately obtain key parameters such as single-cell voltage, load current, fuel cell operating temperature, hydrogen and oxygen pressures, hydrogen flow rate, water vapor partial pressure, and ohmic resistance to construct a polarization model of the fuel cell and further extract the health state characteristics of the unit. The real-time nature and accuracy of these parameters are crucial for subsequent deep learning models because the fault feature vectors relied on for training deep learning models are extracted from this basic data, and any acquisition deviation of a variable will lead to misleading in model training, thereby reducing the prediction accuracy and even causing misdiagnosis. Therefore, in the design of the data acquisition unit, high-precision sensors are required, and through multi-sensor data fusion technology, the noise resistance and reliability of the data are improved. To ensure the integrity and consistency of the data, the data acquisition unit adopts a high-frequency data sampling strategy to capture the dynamic changes within a short period during the system operation. Since faults in hydrogen energy co-generation units often accumulate from minor anomalies, for example, membrane dehydration may start with a slow increase in the fuel cell stack temperature, and catalyst poisoning may manifest as a slight decrease in oxygen utilization rate. Therefore, if the data acquisition frequency is too low, these early abnormal signals may not be captured, resulting in the system issuing an alarm only when a real fault occurs, missing the best intervention opportunity. Therefore, the data acquisition unit of this system can synchronously measure all key variables with a sampling period in milliseconds and construct a time series in combination with historical data, so that subsequent deep learning models can learn the trends and fault development laws in the time dimension. At the same time, to ensure the stability during the data acquisition process, the system also introduces an outlier detection mechanism to filter out mutant data and avoid incorrect input caused by accidental sensor noise or external interference, affecting the accuracy of fault diagnosis. In terms of data transmission, the data acquisition unit is connected to the data analysis unit through a high-speed communication interface to ensure that the data can be transmitted losslessly to subsequent modeling and calculation modules after acquisition. Since hydrogen energy co-generation units usually have a distributed architecture, multiple subsystems such as multiple fuel cell stacks, gas supply systems, and cooling systems need to work together. Therefore, the sensors of the data acquisition unit are distributed at multiple different physical locations and the data is aggregated through wireless or wired networks. During the data synchronization process, to avoid data misalignment problems caused by transmission delays, the system adopts a timing alignment technology to ensure that the data of all variables corresponds to the same operating state at the same timestamp, thereby ensuring that subsequent steps such as polarization curve calculation, electrical efficiency and thermal efficiency calculation, and fault feature extraction can be based on high-quality inputs.

[0038] In the system architecture of the present invention, the first key task of the data analysis and feature extraction unit is to establish a polarization model of the fuel cell. The polarization model is a mathematical expression that describes the relationship between the fuel cell voltage and the load change. It comprehensively considers various factors such as electrochemistry thermodynamics, ohmic losses inside the cell, kinetic characteristics of charge transfer, and gas diffusion, and can reflect the performance changes of the fuel cell under different operating conditions. By analyzing the unit operation data, extracting variables such as single cell voltage, load current, fuel cell temperature, gas pressure, etc., and combining key parameters such as commutation potential, activation polarization, ohmic polarization, and concentration polarization calculated based on electrochemistry principles, this unit can construct a complete polarization curve model to provide data support for subsequent feature extraction. The change of the polarization curve can intuitively reflect the health state of the fuel cell. For example, when the catalyst is poisoned, the activation polarization loss will increase significantly, resulting in a decrease in the cell voltage, and when membrane dehydration or flooding occurs, the ohmic polarization or concentration polarization loss will show abnormal fluctuations. Therefore, establishing an accurate polarization model can not only quantify the electrochemistry performance of the system but also provide a theoretical basis for the construction of fault characteristics. After constructing the polarization model, the second task of the data analysis and feature extraction unit is to calculate the electrical efficiency and thermal efficiency of the hydrogen energy co-generation unit. Since the energy conversion process of the hydrogen fuel cell involves the coupling of electrical energy and thermal energy, it is difficult to comprehensively characterize the operation state of the system only relying on a single voltage or power variable. This unit calculates the electrical efficiency and thermal efficiency of the unit respectively by combining data such as real-time measured fuel consumption, recovered heat, output power, etc., so as to obtain the overall energy utilization of the system. The electrical efficiency characterizes the ability to convert hydrogen chemical energy into electrical energy, while the thermal efficiency describes the performance of the waste heat recovery system. These variables are not only used to evaluate the energy utilization level of the unit but also important features for fault diagnosis. For example, when the electrical efficiency shows an abnormal decrease, it may mean catalyst poisoning or gas diffusion layer blockage, and abnormal fluctuations in the thermal efficiency may indicate a fault in the cooling system or unstable fuel supply. Therefore, by calculating the electrical efficiency and thermal efficiency, this unit can further explore potential fault factors affecting the unit performance and provide input of fault characteristics for the subsequent deep learning model. After completing the calculation and modeling of the basic data, the core task of the data analysis and feature extraction unit is to construct a feature vector for fault diagnosis. The key to this process lies in how to extract the most representative features from complex, multi-variable time series data so that the deep learning model can accurately identify the fault mode. For this purpose, this unit constructs a fault feature vector containing multiple dimensions based on the calculation results of the polarization model, electrical efficiency, and thermal efficiency, including feature variables such as standardized voltage, efficiency, temperature difference, gas supply state, and derivative information of the voltage-current curve. These variables not only cover the basic parameters of the unit operation but also contain key indicators reflecting the health state of the system.For example, the first-order derivative of voltage to current can be used to characterize the dynamic response characteristics of the battery, while the change of the second-order derivative can reveal the stability of the system under different load conditions. In addition, considering that the operating state of the hydrogen energy cogeneration unit is a time-related process, the data at a single time point cannot fully describe the health trend of the system. Therefore, this unit uses a sliding window method to extract features from historical data to construct a more predictive fault feature vector. By combining time series analysis, this unit can effectively capture the long-term change trend of the system, thereby providing more timely input data for the deep learning model.

[0039] The working principle of the fault prediction unit is based on time-correlation modeling. Its input is the fault feature vector constructed by the data analysis and feature extraction unit, and the output is the probability of occurrence of different fault types. Since the faults of hydrogen energy cogeneration units have complex nonlinear characteristics, and different types of faults often have cross-influences, such as catalyst poisoning may lead to reduced battery activity, which in turn causes blockage of the gas diffusion layer, and abnormal cooling system may cause imbalance in membrane hydration state, which in turn affects the polarization characteristics of the battery. Therefore, it is difficult to accurately predict faults by relying only on a single variable or data at a single moment. The present invention utilizes the pattern recognition ability of the deep learning model, and by training a large amount of operating data, enables the system to automatically learn the evolution laws of different fault modes, and combines the historical trends of time series to predict possible future faults. In the specific implementation process, the fault prediction unit uses a neural network to perform nonlinear mapping on the fault feature vector at the current moment, calculates its matching degree with various fault templates, and further combines the feature change trend within the historical time window to evaluate the probability of occurrence of different faults. This process not only considers the similarity between the current feature vector and the known fault mode, but also introduces the Mahalanobis distance metric to measure the rate of change of the fault feature vector in the time dimension, so as to achieve more refined trend analysis. For example, in the early stage of membrane dehydration failure, the ohmic resistance of the fuel cell may show a slow upward trend, but the voltage drop is small at this time. If it only relies on static data analysis, the system may not be able to determine that the fault has occurred. However, through the time correlation modeling of the fault prediction unit, it can be determined that the risk of membrane dehydration is increasing by identifying the trend of resistance change and its historical evolution pattern, and give early warning, so that the operation and maintenance personnel can take intervention measures before the failure occurs.

[0040] Embodiment 2: The types of faults include: membrane dehydration fault, membrane flooding fault, catalyst poisoning fault, gas diffusion layer blockage fault and thermal management system fault.

[0041] Specifically, membrane dehydration failure is one of the common problems during the operation of fuel cells, mainly occurring under low humidity, high temperature or high load conditions. The proton exchange membrane of a fuel cell relies on the hydrated state to maintain good proton conduction ability. If the membrane dehydrates, the proton transfer rate will decrease significantly, leading to an increase in ohmic polarization loss and further causing a drop in the output voltage. During the construction of the fault feature vector of the system of the present invention, special attention is paid to parameters such as the ohmic resistance, operating temperature, hydrogen flow rate, and water vapor partial pressure of the fuel cell. By learning the change trends of these variables through a deep learning model, the system can identify the early signs of membrane dehydration and issue an alarm before the fault occurs. In contrast to membrane dehydration, membrane flooding failure is caused by an imbalance in water management inside the fuel cell. Excessive water absorption by the membrane or the accumulation of liquid water on the cathode side will cause hindered gas diffusion, thereby affecting the oxygen supply efficiency and ultimately reducing the power output of the battery. Membrane flooding is usually manifested as large voltage fluctuations under low load conditions and a decrease in the oxygen excess coefficient. During the feature extraction process of the present invention, the changes in the oxygen pressure, coolant temperature, and polarization curve of the fuel cell are analyzed emphatically, and combined with the time series analysis method, a feature pattern of membrane flooding failure is established, enabling the system to automatically detect the occurrence of membrane flooding and predict its development trend based on historical data. Catalyst poisoning failure is one of the important factors affecting the lifespan of fuel cells, mainly caused by impurities (such as carbon monoxide, sulfides, etc.) in the fuel gas adsorbing on the catalyst surface, reducing its activity, and thus deteriorating the electrode reaction kinetics. The typical characteristics of catalyst poisoning are a decrease in the exchange current density, an increase in activation polarization loss, and a slow decay of the voltage under a constant load. The present invention constructs the fault features of catalyst poisoning by monitoring the exchange current density, load current, hydrogen and oxygen pressures, and activation energy-related parameters of the fuel cell, and uses a deep learning model to classify and predict them, so as to provide intervention measures before the catalyst completely fails and extend the service life of the system. Gas diffusion layer clogging failure usually occurs when the fuel cell has been operating for a long time or the environmental humidity is high. Due to the accumulation of impurity particles, carbides or excessive moisture in the gas diffusion layer, the supply of reaction gases is restricted, thereby affecting the uniformity of the electrode reaction. This failure is usually manifested as an abnormal steep drop in the polarization curve of the fuel cell in the high load region, abnormal fluctuations in the oxygen stoichiometry ratio, and non-linear changes in the derivative of voltage with respect to current. The present invention enables the system to accurately identify the clogging of the gas diffusion layer and predict its development trend by constructing a fault feature vector containing the derivative information of the polarization curve and combining the pattern recognition ability of the deep learning model, thereby providing accurate diagnostic information for maintenance personnel. Thermal management system failure is one of the key factors affecting the stable operation of hydrogen energy combined heat and power units, mainly manifested as the failure of the cooling system, reduced heat exchange efficiency or abnormal temperature control. When the thermal management system fails, the operating temperature of the fuel cell may rise rapidly or fluctuate excessively, thereby affecting the membrane hydration state, catalyst activity, and the rate of the electrochemical reaction.The fault feature vector of the present invention particularly introduces key parameters such as the temperature difference between the coolant inlet and outlet, the fuel cell temperature gradient, the recovered heat, etc., and combines time series trend analysis, enabling the system to detect the abnormal state of the thermal management system and predict its possible impacts in advance, thereby ensuring that the unit operates within the optimal temperature range.

[0042] Example 3: The operation data of the hydrogen energy co-generation unit collected in real time by the data acquisition unit includes: the single cell voltage in V , the load current in A , the stack temperature in K , the hydrogen pressure in Pa , the oxygen pressure in Pa , the hydrogen flow rate in mol / s , the oxygen stoichiometric ratio , the coolant inlet temperature in K , the coolant outlet temperature in K , the fuel cell operating temperature in K , the water vapor partial pressure in Pa , the ohmic resistance in Ω and the recovered heat in W ; where is time.

[0043] Specifically, the single-cell voltage is the most direct indicator of the health state of a fuel cell, which reflects the driving force of the electrochemical reaction and the impact of polarization losses. Under normal operating conditions, the voltage of the fuel cell shows a certain regularity with the change of the load. However, if faults such as a decrease in catalyst activity, membrane dehydration, or diffusion layer blockage occur, the voltage will exhibit abnormal fluctuations. Therefore, the system needs to accurately collect the single-cell voltage and, in combination with time series analysis methods, determine whether its change trend conforms to the normal operating conditions. At the same time, as the output parameter of the battery, the load current determines the rate of the electrochemical reaction and directly affects the change of the polarization curve. Therefore, the data acquisition unit must record the load current synchronously to ensure that the calculation of the polarization model can accurately reflect the system operating state. The stack temperature is one of the core parameters of the internal thermal management of the fuel cell, which directly affects the rate of the electrochemical reaction, the performance of the catalyst, and the hydration state of the proton exchange membrane. If the temperature is too high, it may cause membrane dehydration and accelerate material aging, while if the temperature is too low, it may lead to a decrease in catalyst activity and a reduction in fuel utilization. Therefore, this system collects the stack temperature in real time to ensure that the thermal management system can maintain the best operating conditions. In addition, the supply states of hydrogen and oxygen are also important factors affecting the performance of the fuel cell. The hydrogen pressure and oxygen pressure determine the gas diffusion rate of the electrode reaction and are closely related to the polarization losses of the fuel cell. When the gas supply is insufficient or the pressure is abnormal, the polarization curve may show obvious changes, thus affecting the stability of the system. Therefore, the data acquisition unit must monitor the hydrogen and oxygen pressures in real time and, in combination with a deep learning model, analyze their correlations with other variables to determine whether there are potential gas supply faults. The hydrogen flow rate is one of the core input parameters of the fuel cell, which determines the hydrogen supply rate of the electrochemical reaction and affects the overall efficiency of the fuel cell.

[0044] Under normal operating conditions, the hydrogen flow rate should maintain a stable corresponding relationship with the load current. However, fluctuations in the flow rate may occur in the case of gas diffusion layer blockage or abnormal hydrogen supply. Therefore, this system collects the hydrogen flow rate in real time and calculates the fuel utilization rate by combining this variable with the oxygen stoichiometric ratio, so as to ensure that the fuel cell's gas supply system is in the best state. In addition, the thermal management state of the fuel cell directly affects the system's stability. The inlet and outlet temperatures of the coolant can reflect the efficiency of heat transfer and the health of the heat dissipation system. When the coolant temperature is abnormal, it may indicate problems such as blockage in the thermal management system, insufficient coolant flow rate, or degraded heat exchanger performance. Therefore, the data acquisition unit must accurately monitor the inlet and outlet temperatures of the coolant to ensure the normal operation of the unit's thermal management system. The water vapor partial pressure is one of the important parameters affecting the hydration state of the proton exchange membrane. It determines the conductivity of the membrane and the ohmic polarization loss of the fuel cell. When the water vapor partial pressure is too low, membrane dehydration may occur, leading to an increase in the battery's resistance and a subsequent decrease in the output voltage. When the water vapor partial pressure is too high, it may cause membrane flooding failure, resulting in blocked gas diffusion. Therefore, this system collects the water vapor partial pressure in real time and calculates the hydration state of the membrane by combining the stack temperature and the coolant temperature, and predicts its change trend. The ohmic resistance is a characteristic parameter of the internal resistance of the fuel cell. It is not only affected by the membrane hydration state but is also related to the catalyst layer, electrode structure, and current collector material. When the ohmic resistance increases abnormally, it usually means membrane damage, material aging, or internal contamination of the battery. Therefore, this system monitors the ohmic resistance in real time through the data acquisition unit and calculates its impact on the output power by combining the polarization model to determine whether the system is in a normal operating state. The recovered heat is a key parameter for measuring the waste heat utilization efficiency of the fuel cell. It reflects whether the unit's thermal management system is operating normally. When the recovered heat is below the normal level, it may indicate degraded heat exchanger performance, blocked cooling system, or mismatched operating parameters. Therefore, the data acquisition unit needs to measure the recovered heat in real time and optimize the unit's energy management strategy by combining the calculation models of electrical efficiency and thermal efficiency.

[0045] Example 4: The polarization model is represented by the following formula:

[0046] ;

[0047] Where, is the standard electrode potential, with the unit of V; is the gas constant, with a value of 8.314 J / (mol·K); is the Faraday constant, with a value of 96485 C / mol; is the transfer coefficient, with a value range of 0.3 to 0.7; is the exchange current density, with the unit of A / cm²; is the first concentration polarization parameter, with the unit of V and the value range from 0.02 to 0.08; is the second concentration polarization parameter, with the unit of cm² / A and the value range from 10 to 30.

[0048] Specifically, the standard electrode potential represents the theoretical open-circuit voltage of the fuel cell under standard conditions, which determines the maximum voltage output of the fuel cell. When the fuel and oxidant exist in the standard state and there is no internal loss, the voltage of the fuel cell can reach , but in actual operation, due to the existence of various losses, the working voltage of the fuel cell is always lower than this theoretical value. Therefore, the subsequent various losses are used to correct the actual voltage so that the model can accurately describe the real working conditions. The first loss of the polarization model comes from the influence of gas pressure. This term is the correction term of the electrochemical equilibrium potential derived from the Nernst equation, which describes the change of the output potential of the fuel cell under different gas pressure conditions. is the gas constant, represents the temperature of the fuel cell, is the Faraday constant. These parameters together determine the influence of gas pressure on the potential. The denominator part of and represent the partial pressures of hydrogen and oxygen respectively, while the numerator part of is the partial pressure of water vapor. According to the mathematical form of this term, it can be obtained that the higher the partial pressures of hydrogen and oxygen, the higher the output voltage of the fuel cell, while the increase of the water vapor partial pressure will reduce the cell voltage. This is because the partial pressures of hydrogen and oxygen determine the concentration gradient of the electrode reaction, and the increase of water vapor may affect the reaction kinetics. If the fuel supply system fails, such as the decrease of hydrogen pressure or the limitation of oxygen diffusion, the value of this term will change abnormally, resulting in the decrease of the open-circuit potential of the fuel cell. The second term represents the activation polarization loss, which describes the kinetic loss of the fuel cell in the low-current region. The activation polarization is mainly determined by the activity of the catalyst and the electrode reaction rate, where is the transfer coefficient, which reflects the kinetic characteristics of the electrode reaction, is the exchange current density, which represents the equilibrium current density of the forward and reverse reactions on the electrode surface. When the catalyst activity decreases or the gas transport is limited, the exchange current density will decrease, resulting in an increase in the activation polarization loss and further reducing the output voltage of the fuel cell. The logarithmic form of this term indicates that with the increase of the load current , the activation polarization loss gradually increases, especially in the low-current region, where this loss dominates. If the fuel cell is poisoned by a catalyst, such as CO molecules adsorbed on the catalyst surface, making the electrode reaction slower, will be significantly reduced, resulting in an increase in the value of this item, thereby causing a drop in the voltage of the fuel cell.

[0049] The third item is the ohmic polarization loss, which describes the resistive loss inside the fuel cell and is mainly determined by factors such as the ionic conduction resistance of the proton exchange membrane, the resistance of the electrode material, and the contact resistance of the current collector. The mathematical form of this item is a linear relationship, indicating that when the load current increases, the ohmic loss will increase proportionally. If the hydration state of the proton exchange membrane decreases, for example, due to membrane dehydration resulting in a reduction in proton transport ability, then will increase, thus exacerbating the ohmic loss and causing a drop in the battery voltage. Therefore, the change of this item can be used as a key indicator of the membrane hydration state. The system of the present invention can continuously monitor the change of this item and, combined with a deep learning prediction model, can identify the membrane dehydration phenomenon in advance and provide early warning information for maintenance. The last item is the concentration polarization loss, which mainly describes the voltage drop caused by limited gas transport. When the fuel cell operates at a high load state, the gas diffusion layer may not be able to provide sufficient hydrogen and oxygen in time, resulting in an increase in the gas concentration gradient, reducing the electrode reaction rate, and thus causing a sharp drop in the output voltage of the fuel cell. The exponential function form of this item indicates that when the load current increases to a certain extent, the concentration polarization loss will show a non-linear rapid growth, leading to a sharp decline in the performance of the fuel cell. The parameters and are the first concentration polarization parameter and the second concentration polarization parameter respectively, which determine the severity of the concentration polarization. If the gas diffusion layer of the fuel cell becomes blocked, for example, due to long-term operation resulting in pollutant deposition on the electrode surface, causing gas transport obstruction, the value of this item will increase significantly, resulting in a decrease in the voltage of the fuel cell.

[0050] Example 5: The exchange current density is calculated using the following formula:

[0051] ;

[0052] where is the influence exponent of hydrogen partial pressure on the exchange current density, and its value range is from 0.7 to 1.2; is the influence exponent of oxygen partial pressure on the exchange current density, and its value range is from 0.3 to 0.5; is the reference pressure, and its value is 1 atm or Pa; is the activation energy, and its unit is J / mol; is the pre - exponential factor of the exchange current density, which reflects the magnitude of the exchange current density under ideal conditions without an activation energy barrier. Its unit is A / cm², and the value range is to .

[0053] Specifically, the first term is the pre - exponential factor of the exchange current density. It reflects the theoretical maximum exchange current density of the fuel cell under ideal conditions, that is, without an activation energy barrier. This parameter mainly depends on the intrinsic activity of the catalyst, the properties of the electrode material, and the catalyst loading. In practical applications, catalyst aging, contamination, or changes in the catalyst layer structure will cause to decay, thus affecting the overall performance of the fuel cell. Therefore, in the diagnostic system of the present invention, the deep - learning model can identify the early signs of catalyst deactivation or poisoning by monitoring the change trend over a long period and learning the deterioration mode combined with experimental data. The second term reflects the influence of temperature on the electrode reaction rate. This term is derived from the Arrhenius equation, which describes the exponential relationship between the reaction rate and temperature. Among them, is the activation energy, representing the energy barrier that the electrode reaction needs to overcome, is the gas constant, is the operating temperature of the fuel cell. When the temperature of the fuel cell increases, the value of this exponential term increases, indicating that the electrode reaction activity increases and the exchange current density increases; conversely, when the temperature decreases, the reaction rate decreases and the exchange current density decreases, resulting in a reduction in the output power of the fuel cell. Therefore, in the system of the present invention, the change of the temperature - related term can be used to monitor the thermal management state of the fuel cell. For example, when the cooling system fails and causes to rise abnormally, the electrochemical reaction rate of the fuel cell will increase temporarily, but the long - term high temperature will accelerate the degradation of the catalyst and the proton exchange membrane, ultimately leading to a decline in system performance. The deep - learning model can identify thermal management failures in advance and predict possible long - term impacts by analyzing the change trend of this exponential term over time, thereby improving the safety and reliability of the system.

[0054] The third term reflects the influence of hydrogen partial pressure on the exchange current density, where is the hydrogen partial pressure on the cathode side of the fuel cell, is the reference pressure, is the influence exponent of hydrogen on the exchange current density, usually between 0.7 and 1.2. The physical meaning of this term is that the higher the hydrogen concentration, the faster the electrode reaction rate, and thus the larger the exchange current density. However, when the hydrogen supply is insufficient, for example, due to abnormal fuel supply systems, When it decreases, the value of this item decreases, thereby reducing the exchange current density, resulting in an increase in activation polarization loss and a decrease in the voltage of the fuel cell. The system of the present invention can identify abnormalities in the fuel supply system and predict the impact of fuel shortage on the performance of the fuel cell by real-time monitoring of the hydrogen partial pressure and combining deep learning to analyze the change trend of this item. The fourth item describes the influence of oxygen partial pressure on the exchange current density, where is the oxygen partial pressure on the cathode side of the fuel cell, is the influence index of oxygen on the exchange current density, usually between 0.3 and 0.5. Similar to hydrogen, the higher the oxygen concentration, the faster the electrode reaction rate of the fuel cell, so the exchange current density increases; conversely, if the oxygen supply is insufficient, for example, due to a failure of the air supply system or blockage of the gas diffusion layer, decreases, the value of this item decreases, resulting in a decrease in the exchange current density, ultimately affecting the electrochemical performance of the fuel cell. The system of the present invention can identify gas transport problems and predict the likelihood of failure by real-time monitoring of the oxygen partial pressure and combining a deep learning model to analyze the dynamic characteristics of the oxygen supply system. Exchange current density The calculation formula covers key factors such as catalyst activity, temperature effect, hydrogen partial pressure and oxygen partial pressure, and each part affects the overall performance of the fuel cell. The present invention uses deep learning technology to real-time monitor the parameters of this formula and analyze their change trends to accurately identify the health status of the fuel cell. For example, when the catalyst ages, the value will gradually decrease, resulting in a decrease in the exchange current density; when the fuel supply is insufficient, decreases, causing an increase in activation polarization loss; when the cooling system fails, may increase abnormally, increasing the exchange current density in the short term, but may accelerate material degradation in the long term. After being processed by the deep learning model, this information can be used for early fault diagnosis and provide long-term performance prediction, thereby improving the operating reliability of the fuel cell system.

[0055] Example 6: Transfer coefficient It is calculated using the following formula:

[0056] ;

[0057] where, is the reorganization energy, representing the energy adjustment at the interface between the catalyst and the electrolyte during charge transfer, with the unit of eV, and the value range is from 0.2 to 0.6, where 1 eV = 96.5 kJ / mol; is the activation free energy, defined as the minimum Gibbs free energy required for the reaction to proceed, with the unit of eV, and the value range is from 0.1 to 0.4; is the electron transfer coefficient, related to the charge transfer rate, with a value range of 0.5 to 0.7; is the surface coverage; the surface coverage is calculated using the following formula:

[0058] ;

[0059] where, is the hydrogen adsorption equilibrium constant, , where is the adsorption free energy of hydrogen on the catalyst surface, in units of eV, with a value range of -0.3 to -0.1.

[0060] Specifically, in the calculation formula of the transfer coefficient, is the reorganization energy, which represents the energy barrier that needs to be overcome during the charge transfer process at the catalyst and electrolyte interface, in units of eV. In the fuel cell reaction, electrons are transferred from hydrogen molecules to the catalyst surface and reach the cathode through the electrolyte, and the reduction reaction is completed under the action of oxygen. The magnitude of the reorganization energy directly affects the charge transfer rate. If is too large, it means that the charge transfer process is blocked, resulting in a decrease in the electrode reaction kinetics, further increasing the activation polarization loss and reducing the working voltage of the fuel cell. In the system of the present invention, the change in the reorganization energy can be used to identify the change in the catalyst activity. For example, when the catalyst ages or is poisoned, will gradually increase, thus affecting the transfer coefficient and ultimately leading to a decrease in the performance of the fuel cell. This system uses a deep learning model to analyze the time variation trend to identify the early signs of catalyst aging and predict its impact on the fuel cell life. is the activation free energy, which represents the minimum Gibbs free energy required for the electrode reaction, that is, the minimum energy required for electrons to transfer from fuel molecules to the catalyst. The magnitude of the activation free energy determines the kinetic speed of the fuel cell reaction. The larger the value, the higher the energy barrier required for the reaction, resulting in a decrease in the exchange current density and further increasing the activation polarization loss of the fuel cell. In the diagnostic system of the present invention, change can reflect the change in the catalyst performance. For example, when the catalyst surface is contaminated, the activation free energy may increase, resulting in a slower electrode reaction and thus affecting the overall efficiency of the fuel cell. The deep learning model can predict the performance change trend of the fuel cell under different load conditions by monitoring the change of this parameter over time and combining other operating data.

[0061] exponential term It reflects the influence of temperature on the transfer coefficient, which is derived from the Arrhenius equation and describes the exponential relationship between the chemical reaction rate and temperature. Among them, is the gas constant, is the operating temperature of the fuel cell. When the temperature increases, the value of this exponential term increases, indicating that the rate of the electrode reaction increases and the transfer coefficient increases, resulting in a reduction in the activation polarization loss of the fuel cell; conversely, when the temperature decreases, the transfer coefficient becomes smaller, leading to a decline in the performance of the fuel cell. However, too high a temperature may accelerate the aging of the catalyst. Therefore, in the system of the present invention, the change of the temperature-related term can be used to monitor the thermal management state of the fuel cell. If abnormally increases, it may increase the transfer coefficient in the short term, but may cause damage to the electrode material in the long run. The deep learning model of the present invention can identify the dynamic change pattern of this exponential term and, in combination with the time series analysis method, evaluate and predict the state of the thermal management system. The multiplication factor describes the relationship between the electron transfer process and the hydrogen adsorption on the catalyst surface. Among them, is the electron transfer coefficient, which determines the charge transfer rate between the electrode and the electrolyte interface and usually takes values between 0.5 and 0.7. If the electron transfer coefficient is small, it means that the charge transfer is blocked, the exchange current density of the fuel cell decreases, and the activation polarization loss of the system increases. On the other hand, is the surface coverage, which represents the proportion of the catalyst surface adsorbed by hydrogen. The calculation formula of the surface coverage illustrates the influence of hydrogen pressure on the reaction activity of the catalyst surface, where is the hydrogen adsorption equilibrium constant, which is determined by (the adsorption free energy of hydrogen on the catalyst surface). The numerical range of the adsorption free energy is from -0.3 to -0.1 eV, indicating that the adsorption of hydrogen on the catalyst is an exothermic reaction. An increase in temperature will reduce the hydrogen adsorption amount, while an increase in pressure will increase the hydrogen adsorption amount. It can be seen from the expression of the surface coverage that when the hydrogen partial pressure increases, gradually approaches 1, meaning that the catalyst surface is completely covered by hydrogen, which can ensure the efficient operation of the fuel cell. However, if the hydrogen supply is insufficient, resulting in a decrease in , will also decrease, thereby affecting the reaction activity of the catalyst and causing the transfer coefficient to decrease, ultimately affecting the output voltage of the fuel cell. The deep learning system of the present invention can detect abnormalities in the fuel supply system and predict possible performance degradation problems by continuously monitoring the change of and combining the calculation results of .

[0062] Example 7: Electrical Efficiency and Thermal Efficiency are calculated using the following formulas respectively:

[0063] ;

[0064] ;

[0065] wherein, is the calorific value of hydrogen.

[0066] Specifically, the calculation formula of the electrical efficiency reflects the electrical energy conversion ability of the fuel cell. The numerator part represents the actual electrical power output of the fuel cell, where is the operating voltage of a single cell, is the load current, and their product is the electrical power provided by the battery at a certain moment. The denominator part represents the total energy of hydrogen input, where represents the hydrogen flow rate, is the calorific value of hydrogen, that is, the total energy released when each mole of hydrogen burns completely. Since the working principle of the fuel cell is that hydrogen and oxygen undergo an electrochemical reaction to generate electrical energy, this ratio can be used to measure the ability of the fuel cell to convert the input hydrogen energy into electrical energy. From the perspective of the formula structure, the electrical efficiency of the fuel cell is affected by multiple factors. If the operating voltage decreases, for example, due to catalyst poisoning, membrane dehydration, or gas transport limitation, the electrical power output of the fuel cell will decrease, resulting in a decrease in electrical efficiency. Similarly, if the hydrogen flow rate increases but the electrical power does not increase proportionally, it indicates a decrease in fuel utilization efficiency, resulting in a decrease in electrical efficiency. In addition, the load current also affects the electrical efficiency. When the load current increases, the polarization loss may increase, resulting in a decrease in the operating voltage and thus a decrease in electrical efficiency. Therefore, the fault diagnosis system of the present invention can analyze the historical change patterns through deep learning, combined with polarization curves and gas supply data, to identify whether there is performance degradation in the battery and predict the possible fault development trend. For example, if the electrical efficiency continuously decreases under the same operating conditions, it may indicate catalyst aging or abnormal gas supply system. This system trains a large amount of operating data through a deep learning model, enabling it to autonomously learn these patterns and provide early warnings before faults occur.

[0067] The calculation formula of the thermal efficiency reflects the waste heat utilization ability of the fuel cell system. The numerator part represents the heat recovered by the system, and the denominator part Represents the remaining energy in the fuel cell reaction that is not converted into electrical energy, i.e., the part available for waste heat recovery. The physical meaning of this formula is that the fuel cell can not only be used as an electrical energy supply device but also utilize its waste heat to improve the overall energy efficiency of the system. One of the core advantages of a hydrogen energy combined heat and power unit is its ability to efficiently utilize the chemical energy of the fuel, providing not only electricity but also heating or for industrial processes. Therefore, the calculation of thermal efficiency is crucial for the overall system optimization. From the perspective of thermal efficiency calculation, if decreases, that is, the recovered heat decreases, indicating a decline in waste heat utilization efficiency, which may be caused by cooling system failure, heat exchanger blockage, or abnormal temperature control. The system of the present invention can analyze the long-term change trend of thermal efficiency through a deep learning model, combined with data such as coolant temperature and fuel cell temperature gradient, to identify possible problems in the thermal management system. For example, if the thermal efficiency of the fuel cell continuously decreases under the same hydrogen input, it may indicate a reduction in the heat transfer efficiency of the heat exchange system and may even predict a blockage in the coolant circulation. The intelligent diagnosis system of the present invention can, based on this data, discover the early characteristics of faults through pattern recognition methods and provide maintenance suggestions in advance to avoid the system being affected before a serious fault occurs.

[0068] Example 8: The process of constructing the fault feature vector of a hydrogen energy combined heat and power unit using a deep learning model specifically includes: First, construct the following initial fault vector:

[0069] ;

[0070] wherein, is the reference voltage; is the reference electrical efficiency; is the reference thermal efficiency; then use a single-layer neural network in deep learning to extract and process to obtain the fault feature vector .

[0071] Specifically, the initial fault feature vector is composed of seven key variables, each variable having a specific physical meaning and being able to effectively reflect the health state of the fuel cell system. First, represents the voltage normalization value of the fuel cell, where is the reference voltage used to eliminate the voltage scale difference between different fuel cell systems. The physical meaning of this ratio is to measure the deviation of the current working voltage of the fuel cell from the normal operating state. When the catalyst ages, the membrane dehydrates, or gas transport is blocked, the working voltage usually drops below the reference voltage. Therefore, this parameter can be used to detect the performance degradation of the fuel cell. The second variable represents the electrical efficiency normalization value of the fuel cell, where is the reference electrical efficiency. Electrical efficiency is an important indicator to measure the energy conversion ability of a fuel cell. Its decrease may indicate catalyst poisoning, insufficient fuel supply, or increased polarization loss. This normalized value is used to exclude the influence of different operating conditions on electrical efficiency, enabling the system to identify abnormal patterns based on relative change trends. For example, under the same hydrogen input, if the electrical efficiency decreases while the load power remains unchanged, it may mean a decrease in the surface activity of the catalyst or limited gas transport. The third variable represents the normalized value of the thermal efficiency, where is the reference thermal efficiency. Thermal efficiency is used to measure the waste heat utilization of a fuel cell system. Its decrease may be caused by cooling system failure, reduced heat exchange efficiency, or abnormal membrane hydration state. Since the thermal management of a fuel cell directly affects its long-term stability, the change trend of this parameter can reflect the health status of the system. For example, at the same input power, if the thermal efficiency decreases significantly, it may mean a blockage in the cooling system or damage to the heat exchanger.

[0072] The fourth variable reflects the difference between the stack temperature and the coolant temperature and is normalized to the ratio of the load current. The physical meaning of this ratio is to describe the working efficiency of the thermal management system. The larger the temperature difference, the more likely it indicates limited heat dissipation capacity of the fuel cell. If this parameter continues to rise under the same load conditions, it may mean insufficient coolant flow or reduced heat exchange efficiency, resulting in the operating temperature of the fuel cell deviating from the optimal range. The fifth variable represents the ratio of hydrogen pressure to flow rate. This parameter is used to measure the health status of the hydrogen transport system. When the hydrogen pressure is high but the flow rate is low, it may mean that the hydrogen supply is blocked, such as a blockage in the hydrogen supply pipeline or the gas diffusion layer. Conversely, if the flow rate is large but the pressure is low, it may mean leakage or abnormal regulation of the gas supply system. The sixth variable represents the first derivative of the fuel cell voltage with respect to the load current. This parameter is used to characterize the slope of the polarization curve and reflects the dynamic response of the fuel cell under different load conditions. When the polarization curve of the fuel cell becomes steeper, the value of this parameter decreases, indicating that the operating state of the fuel cell is limited, such as a decrease in catalyst performance or limited gas transport. The seventh variable represents the second derivative of the fuel cell voltage with respect to the load current. This parameter is used to characterize the curvature of the polarization curve and reflects the non-linear characteristics of the fuel cell under high load conditions. When the absolute value of this parameter increases, it means that the polarization characteristics of the fuel cell have changed significantly, which may be caused by changes in membrane hydration state, abnormal gas transport, or increased polarization loss.

[0073] The initial fault vector is used as the input of the neural network. The dimension of this vector is (i.e., The number of variables included). The input vector passes through the weight matrix of the hidden layer and the bias vector for linear transformation. The formula is as follows:

[0074] ;

[0075] Where: is the initial fault feature vector, with a dimension of (in the example of the present invention ). is the weight matrix, with a dimension of , where is the number of neurons in the hidden layer. is the bias vector, with a dimension of . In order to enhance the expressive power of the neural network, a non-linear activation function is used to process . The present invention adopts the ReLU activation function, and its calculation method is as follows:

[0076] ;

[0077] The characteristic of the ReLU activation function is to maintain a linear relationship for positive inputs and output 0 for negative inputs, which can enhance the learning ability of the neural network and avoid the problem of gradient disappearance. The activation vector after being processed by the hidden layer is further linearly transformed by the output layer to generate the final fault feature vector:

[0078] ;

[0079] Where: is the weight matrix of the output layer, with a dimension of , where is the dimension of the final feature vector. In the present invention, the dimension remains unchanged and takes the value of 7. is the bias vector of the output layer, with a dimension of . is the final fault feature vector, with a dimension of . After the above calculations, the finally obtained is a -dimensional vector, which extracts the core information in the initial fault feature vector and learns more complex fault patterns through the training of the neural network. This vector can be used for subsequent fault classification, anomaly detection or time series prediction to determine whether the hydrogen energy co-generation unit is in a healthy state or predict the probability of future faults. To improve the prediction performance, the present invention can adopt an MLP structure with regularization, such as in the weight matrix Add an L2 regularization term to the calculation to prevent overfitting. In addition, the number of neurons in the output layer can be adjusted according to different application requirements. For example, it can be set to correspond to five typical faults (membrane dehydration, membrane flooding, catalyst poisoning, gas diffusion layer blockage, and thermal management system failure).

[0080] Example 9: Use the following formula to calculate the probability of each type of fault occurring at each time step:

[0081] ;

[0082] where represents the probability of a fault occurring at time ; represents the time step; represents the fault feature vector and the Mahalanobis distance between the fault feature template vector; represents the fault feature vector and the square of the Euclidean distance between the fault feature template vector; is the set historical time window; is the integration variable; and are both integer subscript indices.

[0083] Specifically, the present invention first uses the fault feature vector constructed by deep learning , which is a high-dimensional representation extracted from multiple operating parameters of the fuel cell, including information such as voltage, current, temperature, and pressure. To evaluate the similarity between the current state and known fault modes, the system calculates and the Mahalanobis distance between each known fault mode . The Mahalanobis distance has more advantages than the Euclidean distance. It not only considers the covariance information between variables but also can normalize features of different scales, enabling the system to accurately measure the similarity between different fault modes even in the face of complex data distributions. Its calculation method is essentially a transformation based on the covariance matrix, which can more precisely distinguish fault modes with similar statistical characteristics in the multi-dimensional space. The smaller the distance, the closer the current state is to the fault The closer the characteristics are, the higher the probability of this fault occurring. To map this distance relationship to a probability distribution, the present invention adopts an exponential normalization method, that is, taking the negative exponent of the Mahalanobis distance and normalizing it, so that the sum of the probabilities of all faults is always 1. This method is similar to Softmax normalization, but on this basis, a negative exponent operation is added, so that the state closer to the known fault mode has a higher probability. This means that if the feature vector at a certain moment is very close to the feature template of a specific fault the probability of this fault occurring will be amplified, while the situation with a lower similarity to other faults will be suppressed. This can ensure that the system can correctly identify the most likely fault type.

[0084] However, relying solely on the data at the current moment for classification is not enough, because the faults of the hydrogen energy cogeneration unit do not occur instantaneously, but evolve over time. Therefore, the present invention introduces a time evolution correction term, which calculates the Euclidean distance change trend between the feature vectors within a time window and the fault template to measure the time stability of the system operation state. Specifically, this term calculates the change rate of the Euclidean distance between the fault feature vector and a specific fault template in the past time period, and accumulates it through integration. If the Euclidean distance has changed greatly in the past period of time, it means that the system state fluctuates strongly and has not yet stabilized towards a certain fault mode, then the system will reduce the prediction probability of this fault; if this distance gradually shrinks over time, it means that the system is converging towards this fault mode, then the probability of this fault will increase. This mechanism effectively enhances the system's ability to capture the trend of fault occurrence, ensures that the system will not generate false alarms due to short-term abnormal fluctuations, and can accurately judge the development trend of faults. In practical applications, the physical meaning of this time correction term can be understood as: if within the time the operating state of the fuel cell is continuously approaching a certain fault mode, for example, the ohmic resistance of the battery gradually increases due to membrane dehydration fault, and this trend persists in multiple time steps, then this correction term will strengthen the probability of the membrane dehydration fault, enabling the system to give an early warning. However, if there are large fluctuations in the fault feature vector in the past period of time, such as instantaneous voltage fluctuations caused by load changes, but then the system returns to the normal operating state, then this correction term will reduce the prediction probability of this fault, thus avoiding misjudgment. This analysis method based on the time change trend makes the prediction model of the present invention more robust than the traditional static diagnosis method and can identify the potential development trend of faults in advance.

[0085] The main manifestations of membrane dehydration failure are increased ohmic resistance, decreased fuel cell voltage, and reduced electrical efficiency. Insufficient membrane hydration state will lead to a decrease in proton conduction ability, which in turn causes greater ohmic polarization loss. ; (the operating voltage drops by 15%); (the electrical efficiency drops by 20%); (the thermal efficiency drops by 10%); (the stack temperature rises); (the hydrogen utilization rate rises); the slope of the polarization curve decreases , and the second derivative tends to be flat ; The manifestation of membrane flooding failure is blocked gas diffusion, drastic changes in the polarization curve, and intensified voltage fluctuations, which usually occur in low-load or low-temperature environments. ; The gas diffusion layer is blocked by liquid water, resulting in (restricted hydrogen transport); the polarization curve drops sharply in the high-load region ; Catalyst poisoning failure will lead to an increase in activation polarization loss, a decrease in exchange current density, and a decrease in catalyst activity. Typically, it is manifested as a more obvious voltage drop in the low current density region. ; The exchange current density decreases, resulting in an increase in activation polarization loss ; The gas diffusion layer blockage failure causes restricted oxygen diffusion, a steep drop in the polarization curve, and drastic voltage fluctuations, which are usually caused by long-term operation or pollutant accumulation. ; The oxygen diffusion ability decreases, resulting in ; the polarization curve drops steeply ; The manifestation of the thermal management system failure is a decrease in cooling efficiency, abnormal stack temperature, and a large voltage drop. ; The coolant temperature deviation is relatively large ; The thermal efficiency is significantly reduced .

[0086] The preferred embodiments of the present disclosure have been described above with reference to the accompanying drawings, and the scope of rights of the present disclosure is not limited thereby. Any modifications, equivalent replacements, and improvements made by those skilled in the art without departing from the scope and essence of the present disclosure shall be within the scope of rights of the present disclosure.

Claims

1. A hydrogen energy cogeneration unit fault diagnosis and prediction system based on deep learning, characterized in that: The system includes: a data acquisition unit, a data analysis and feature extraction unit and a fault prediction unit; the data acquisition unit is used to collect the operating data of the hydrogen energy cogeneration unit in real time; the data analysis and feature extraction unit is used to perform electrochemical performance analysis based on polarization curves according to the operating data, and establish a polarization model of the hydrogen energy cogeneration unit; according to the polarization model, the electrical efficiency and thermal efficiency of the hydrogen energy cogeneration unit are calculated, and then the deep learning model is used to construct the fault feature vector of the hydrogen energy cogeneration unit; the fault prediction unit is used to judge the probability of various types of faults occurring in the next time step according to the fault feature vector and the feature template vectors of various types of faults set; the operating data of the hydrogen energy cogeneration unit collected in real time by the data acquisition unit includes: single cell voltage in V , load current in A , Stack temperature in K , Hydrogen pressure in Pa , oxygen pressure in Pa , Hydrogen flow rate in mol / s , oxygen chemical equivalent ratio , coolant inlet temperature in K , coolant outlet temperature in K , fuel cell operating temperature in K , water vapor partial pressure in Pa , Ohm resistance in Ω and the recovered heat in W ;in, is the time; the polarization model is expressed using the following formula: ； in, is the standard potential, in V; is the gas constant, which is 8.314 J / (mol·K); is the Faraday constant, which is 96485C / mol; is the transfer coefficient, ranging from 0.3 to 0.7; is the exchange current density, in A / cm²; is the first concentration polarization parameter, the unit is V, and the value range is 0.02 to 0.08; is the second concentration polarization parameter, in cm² / A, ranging from 10 to 30; Electrical efficiency and thermal efficiency Use the following formulas for calculation: ； ； in, is the calorific value of hydrogen.

2. The deep learning-based hydrogen energy cogeneration unit fault diagnosis and prediction system according to claim 1, characterized in that: The types of faults include: membrane dehydration fault, membrane flooding fault, catalyst poisoning fault, gas diffusion layer blockage fault and thermal management system fault.

3. The hydrogen energy cogeneration unit fault diagnosis and prediction system based on deep learning as claimed in claim 2, characterized in that: The exchange current density is calculated using the following formula: ； in, is the influence index of hydrogen partial pressure on exchange current density, ranging from 0.7 to 1.2; is the influence index of oxygen partial pressure on exchange current density, ranging from 0.3 to 0.5; is the reference pressure, which is 1 atm or Pa; is the activation energy, in J / mol; The exchange current density prefactor reflects the exchange current density under ideal conditions without an activation energy barrier. The unit is A / cm² and the value range is arrive .

4. The deep learning-based hydrogen energy cogeneration unit fault diagnosis and prediction system according to claim 3, characterized in that: Transfer coefficient Calculated using the following formula: ； in, is the reorganization energy, which represents the energy adjustment of the catalyst and electrolyte interface during charge transfer. The unit is eV and the value range is 0.2 to 0.6, where 1eV=96.5kJ / mol; is the activation free energy, which is defined as the minimum Gibbs free energy required for the reaction to proceed, in eV, ranging from 0.1 to 0.4; is the electron transfer coefficient, which is related to the charge transfer rate and ranges from 0.5 to 0.7; is the surface coverage; surface coverage Calculated using the following formula: ； in, is the hydrogen adsorption equilibrium constant, ,in It is the adsorption free energy of hydrogen on the catalyst surface, measured in eV, and ranges from -0.3 to -0.

1.

5. The deep learning-based hydrogen energy cogeneration unit fault diagnosis and prediction system according to claim 1, characterized in that: The process of constructing the fault feature vector of the hydrogen energy cogeneration unit using the deep learning model specifically includes: first constructing the following initial fault vector: ； in, is the reference voltage; is the reference electrical efficiency; is the reference thermal efficiency; then a single-layer neural network in deep learning is used to extract Processing is performed to obtain the fault feature vector .

6. The hydrogen energy cogeneration unit fault diagnosis and prediction system based on deep learning according to claim 5, characterized in that: The probability of each type of failure occurring at each time step is calculated using the following formula: ； in, express Failures occur all the time probability; represents the time step; Represents the fault feature vector With failure Mahalanobis distance of feature template vector; Represents the fault feature vector With failure The square of the Euclidean distance of the feature template vector; is the set historical time window; is the integration variable; and All are integer subscript indices.

Images