Modeling method for fuel cell sensor screening considering failure degradation effect

Abstract

The application discloses a kind of fuel cell sensor screening modeling methods considering the influence of failure degradation.It is modeled to include fuel cell sensor optimization screening and fuel cell simulation, simulation model is used to calculate the sensitivity of model output parameter, sensor optimization selection is carried out by coupling failure mode weighting method through sensitivity analysis, simulation model and machine learning algorithm are combined to realize the screening of sensor involved in vehicle fuel cell, suitable for fuel cell performance prediction and fault diagnosis application.Through effective sensor optimization screening, the performance prediction result and state classification result of fuel cell are greatly improved.The application can not only filter the sensors that cannot effectively represent the system performance and state, thereby improving the accuracy of prediction and diagnosis, but also improve the computing efficiency, thereby solving the difficulty that high-dimensional complex information is difficult to meet the online diagnosis requirements.The entire optimization screening process can be automatically realized by using the prepared program.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of fuel cells, and particularly relates to a modeling method for optimizing and screening sensors in a fuel cell system. BACKGROUND

[0002] Fuel cells show great potential in the fields of transportation and energy storage due to their high power density and fast response. Due to changes in load and operating environment, component wear and tear, and other factors, many faults will inevitably occur in fuel cells during use. If these faults are not identified and repaired in a timely manner, they will accelerate the performance degradation of the equipment, cause the fuel cell equipment to fail, and even cause safety problems. In order to ensure the safe and stable operation of the fuel cell system, it is of great significance and value to develop a complete fault and performance monitoring system before large-scale commercialization.

[0003] In order to obtain more accurate fuel cell performance prediction and fault diagnosis results, it is necessary to model the historical working state data of a large number of sensors and approximate the mapping mechanism and deep information contained in the system data through pattern recognition, so as to achieve the purpose of fault detection and performance prediction. The balance between calculation efficiency and accuracy is also an important factor that needs to be considered in the diagnosis method. The use of multiple sensors provides more distinguishable information for different system faults. However, the analysis of high-dimensional data sets is complex and time-consuming, and it is difficult to implement in embedded systems with limited computing power.

[0004] The present application proposes a sensor screening modeling method suitable for fuel cell performance prediction and fault diagnosis applications, thereby filtering out sensors that are not sensitive to changes in fuel cell performance. Based on the fuel cell model, the effects of various common failure modes on the fuel cell are considered, the effects of different failure modes on system performance degradation are coupled, and the best sensor set is selected to solve the difficulty of high-dimensional complex information not meeting the requirements of online diagnosis. SUMMARY

[0005] The purpose of the present application is to propose a fuel cell sensor optimization screening modeling method considering the effects of fault degradation. From the perspective of the effects of fuel cell faults on degradation, the method of coupling failure mode weighting through sensitivity analysis is used for sensor optimization selection to solve the difficulty of high-dimensional complex information not meeting the requirements of online diagnosis.

[0006] The following describes the principles of the steps of the present application. The fuel cell sensor screening modeling method considering the effects of degradation and failure includes a fuel cell sensor optimization screening model, a fuel cell simulation model, and a machine learning algorithm. The fuel cell simulation model is used to calculate the sensitivity of the model output parameters, wherein the health parameters involved in the simulation model, the current density, the active area, and the liquid water content are used as inputs; and the set of sensor parameters predicted by the simulation model is used as output. The functional expression of the fuel cell simulation model is:

[0007] α = F(I, A act ,s) (1)

[0008] α represents the set of sensor parameters, wherein the sensor parameters include V, T a ,T c ,RH c ,q a,in ,q a,out ,q c,in ,q c,out . The calculation expression of the sensitivity set of the fuel cell output parameters to the three health parameters of the current density, the active area, and the liquid water content is:

[0009]

[0010]

[0011]

[0012] The sensor parameters have different sensitivities to different health parameters, and a weighted selection method of the effects of three health parameter failures on the degradation of fuel cell performance is used, and the calculation formula is as follows:

[0013]

[0014] In the formula, DF represents the set of sensor weighted degradation sensitivities; D I represents the performance degradation rate caused by the current density failure mode; represents the performance degradation rate caused by the active area failure mode; D s represents the performance degradation rate caused by the liquid water content failure mode.

[0015] The set of sensor weighted degradation sensitivities includes:

[0016]

[0017] According to the descending order of the values from large to small, the parameters included in the set of sensor weighted degradation sensitivities are sorted, and the functional relationship is:

[0018]

[0019] In the formula, sort represents a descending order function of a set; delta represents a sensor weighted degradation sensitivity set after descending order; and the elements in delta are subsequently grouped:

[0020] G i =(delta1,...,delta i i=1,2,...,8 (8)

[0021] Using a machine learning algorithm, the sensor set contained in each group is taken as input, the i group with the highest accuracy is found, and the obtained G i is the best sensor set for screening, thereby completing the process of filtering redundant sensors and sensor optimization screening.

[0022] Further, the available machine learning algorithm classification models include but are not limited to adaptive neural fuzzy systems, convolutional neural networks, artificial neural networks, and long short-term memory artificial neural networks.

[0023] The characteristics and beneficial effects of the present application are as follows: the proposed fuel cell sensor screening method considers the degradation influence of fuel cell failure, uses the sensitivity analysis coupled with the weighted failure mode method to perform sensor optimization selection, and uses the combination of the fuel cell simulation model and the machine learning algorithm to realize the screening of the sensors involved in the vehicle-mounted fuel cell, which is suitable for fuel cell performance prediction and fault diagnosis applications. The present application not only filters sensors that cannot effectively represent the performance and state of the system, thereby improving the accuracy of prediction and diagnosis, but also improves the computational efficiency, thereby solving the difficulty of high-dimensional complex information that cannot meet the requirements of online diagnosis. The entire sensor screening method can be automatically realized by using the prepared program. In the traditional sensor screening research process, researchers need to perform experiments or simulation operations on all possible combinations of sensors, and the final screening sensor process is completed according to a large number of attempts, which requires a large amount of manpower and experimental funds. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is the prediction result of using the adaptive neural fuzzy system after sensor screening by the present application.

[0025] Figure 2 is the prediction result of using the adaptive neural fuzzy system without sensor screening.

[0026] Figure 3 is the classification result of using the convolutional neural network after sensor screening by the present application.

[0027] Figure 4 is the classification result of using the convolutional neural network without sensor screening.

[0028] Figure 5 is the training prediction time of the adaptive neuro-fuzzy system for the screened and un-screened examples. DETAILED DESCRIPTION

[0029] The method and modeling calculation process of the present application are further described below by specific examples. It should be noted that the examples are descriptive rather than limiting, and do not limit the protection scope of the present application.

[0030] The fuel cell sensor screening modeling process in the present example includes a fuel cell sensor optimization screening model, a fuel cell simulation model, and a machine learning algorithm. The fuel cell simulation model is non-limiting. By analyzing the sensitivity of sensor parameters to the health parameters involved in the simulation model, the coupling failure mode weighting method is used for sensor optimization selection.

[0031] The specific process of adaptive matching in the present example is as follows:

[0032] The sensitivity of the model output parameters is calculated using the fuel cell simulation model, where the health parameters involved in the simulation model, current density, activation area, and liquid water content, are used as inputs; and a represents the set of sensor parameters predicted by the simulation model and is used as output. The functional expression of the fuel cell simulation model is:

[0033] a = F (I, A act ,s) (1)

[0034] The calculation expression of the sensitivity set of the fuel cell output parameters to the three health parameters, current density, activation area, and liquid water content, is:

[0035]

[0036]

[0037]

[0038] The failure of the health parameters inside the fuel cell will cause the degradation of the fuel cell performance. The sensor parameters have different sensitivities to different health parameters. A weighted selection method considering the influence of the failure of the three health parameters on the degradation of the fuel cell performance is used to select the sensor parameters with higher sensitivity to the degradation caused by the failure mode. The calculation formula is as follows:

[0039]

[0040] The sensor weighted degradation sensitivity set includes:

[0041]

[0042] The weighted degradation sensitivity of each sensor is calculated: DF V = 17.79; The parameters contained in the set of sensor weighted degradation sensitivity are sorted in descending order according to their numerical values from large to small, and the functional relationship is:

[0043]

[0044] Through sorting, the set δ is obtained, Then, the elements in δ are grouped:

[0045] G i = (δ1,..., δ i i = 1, 2,..., 8 (8)

[0046] The grouping result is: G1 = (V); G2 = (V, q c,out );

[0047] G3 = (V, q c,out , q c,in ); G4 = (V, q c,out , q c,in , RH c ); G5 = (V, q c,out , q c,in , RH c , q a,out );

[0048] G6 = (V, q c,out , q c,in , RH c , q a,out , T c ); G7 = (V, q c,out , q c,in , RH c , q a,out , T c , T a );

[0049] G8 = (V, q c,out , q c,in , RH c , q a,out , T c , T a , q c,in ).

[0050] The sensor parameters included in the above 8 groups are respectively taken as inputs, and the adaptive neuro-fuzzy system is used for training and prediction, and the group with the highest accuracy is G6, and the root mean square error is 0.0044, and the prediction results of the sensor set included in G6 are taken as inputs, as shown in Figure 1 , so G6 is the best sensor set for screening.

[0051] In order to verify the universality of the sensor screening method proposed in the application, the adaptive neuro-fuzzy system is replaced by a convolutional neural network for classification test, and the group with the highest classification accuracy is also G6. The classification results are shown in FIG. 3, and the classification accuracy is 99.2%, so G6 is the best sensor set for screening. Thus, the process of filtering redundant sensors and sensor optimization screening is completed.

[0052] Figure 1 and Figure 2 are the prediction results of the adaptive neuro-fuzzy system after screening and the prediction results of the adaptive neuro-fuzzy system without screening, respectively. It is found from the prediction results that the prediction results after optimization screening are significantly better than those without optimization screening. Through calculation, the prediction root mean square error after optimization screening is 0.0044, and the prediction root mean square error without optimization screening is 0.0191, which has a very significant improvement, and can effectively improve the fuel cell performance prediction accuracy. Therefore, the optimal sensor combination obtained by using the method has high practical value.

[0053] Figure 3 and Figure 4 show the classification results of the convolutional neural network after screening and the classification results of the convolutional neural network without screening, respectively. Figure 3 and Figure 4 The horizontal coordinate in and represents the label of the data sample classified by the convolutional neural network, and the vertical coordinate represents the actual classification label of the data sample. The number in the frame represents the number of data whose actual classification is the vertical coordinate label and whose predicted classification is the horizontal coordinate label. It is found from the classification results that the classification results after optimization screening are improved by 3.6% compared with those without optimization screening, which has a very significant improvement. Therefore, the sensor screening method proposed in the application is not only suitable for fuel cell performance prediction but also suitable for fault diagnosis classification application.

[0054] Figure 5 show the training and prediction time of the adaptive neuro-fuzzy system after screening and without screening. It is found from the comparison results that the proposed sensor screening method not only improves the prediction accuracy but also greatly improves the calculation efficiency.

[0055] Through comparison, it can be seen that through effective sensor optimization screening, the performance prediction result and the state classification result of the fuel cell are greatly improved, effective filtering of the sensors not sensitive to the performance change of the fuel cell can improve the prediction and classification results and improve the calculation efficiency.

[0056] The fuel cell simulation model used in the embodiment is non-limiting, and the calculation process is as follows:

[0057] The output voltage of the fuel cell is calculated:

[0058] V=E rev -η ohm -η act,a -η act,c (9)

[0059] In the formula, V represents the output voltage of the fuel cell (V);E rev represents the reversible voltage loss (V);η ohm represents the ohmic loss (V);η act,a represents the activation loss of the anode (V);η act,c represents the activation loss of the cathode (V), wherein the calculation expression of the reversible loss is as follows:

[0060]

[0061] In the formula, F represents the Faraday constant (96485 C mol -1 ); represents the hydrogen partial pressure of the anode flow channel (atm), represents the oxygen partial pressure of the cathode flow channel (atm), and the calculation expression of the ohmic loss is as follows:

[0062]

[0063] In the formula, I represents the current density of the fuel cell (A m -2 );d represents the flow channel thickness (m);δ GDL represents the gas diffusion layer thickness (m);δ MPL represents the microporous layer thickness (m);δ CL represents the catalyst layer thickness (m);δ MEM represents the proton exchange membrane thickness (m);σ BP represents the bipolar plate conductivity; represents the gas diffusion layer electronic conductivity; represents the microporous layer electronic conductivity; represents the catalyst layer electronic conductivity; represents the catalyst layer proton conductivity;σ MEM represents the proton exchange membrane conductivity.

[0064] The activation loss is divided into anode activation loss and cathode activation loss, and the calculation expression is as follows:

[0065]

[0066]

[0067] In the formula, a represents a coefficient; represents an anode reference exchange current density (A m -2 ), represents a cathode reference exchange current density (Am -2 ); s represents liquid water content; represents an anode flow channel hydrogen concentration (mol m -3 ), represents an oxygen concentration of a cathode flow channel (mol m -3 ); represents an anode reference hydrogen concentration (mol m -3 ), represents a cathode reference oxygen concentration (mol m -3 ).

[0068] The relative humidity calculation formula of the fuel cell is:

[0069] RH = RTP sat c vap (14)

[0070] In the formula, RH represents relative humidity, R represents a general gas constant, T represents environmental temperature (K), c vap represents water vapor concentration (mol m -3 ), and P sat represents saturated water vapor pressure, and the calculation formula is:

[0071]

[0072] The calculation formula of the fuel cell inlet flow is:

[0073] q in = c in v (16)

[0074] In the formula, q in represents fuel cell inlet flow (mol m -2 s -1 ), c in represents inlet gas concentration (mol m -3 ), and v represents gas flow rate (m s -1 ). The outlet flow is the difference between the inlet flow, diffusion amount, and source term, and the calculation formula is:

[0075]

[0076] where q out is the fuel cell outlet flow rate (mol m -2 s -1 ), is the diffusion flow rate to the porous layer (mol s -1 ), S is the gas source term (mol m -2 s -1 ), A act represents the active area (m 2 ). The gas components inside the fuel cell, including water vapor, hydrogen, oxygen, and nitrogen, are solved for the gas components, the flow channel gas concentration, in the specific solution as follows:

[0077]

[0078] where the subscript i represents the type of gas component, including hydrogen, water vapor, nitrogen in the anode, and oxygen, water vapor, and nitrogen in the cathode; represents the effective diffusion rate of gas component i between the m layer and the n layer (m 2 s -1 ); c i,n represents the gas concentration of gas component i in the n layer (mol m -3 ); c i,m represents the gas concentration of gas component i in the m layer (mol m -3 ), S i,m represents the source term of gas component i in the m layer (mol m -2 s -1 ), thereby solving the gas component concentration.

[0079] The local temperature inside the fuel cell is solved by the energy conservation equation, and the calculation expression is as follows:

[0080]

[0081] where T m represents the temperature of the m layer (K); T n represents the temperature of the m layer (K); pCp m represents the effective volume heat capacity of the m layer (Jm -3 K -1 ); S T represents the heat source term (W m -2 ); represents the effective thermal conductivity between the m layer and the n layer (W m -1 K -1 ).

[0082] The fuel cell simulation model used in this example optimizes sensor selection by analyzing the sensitivity of sensor parameters to health parameters involved in the simulation model and employing a coupled failure mode weighting method.

[0083] The above analysis and calculations demonstrate that the proposed fuel cell sensor screening method, considering the degradation impact of fuel cell faults, utilizes a sensitivity analysis coupled with failure mode weighting for sensor optimization. By combining fuel cell simulation models and machine learning algorithms, it can effectively screen sensors involved in onboard fuel cells, making it suitable for fuel cell performance prediction and fault diagnosis applications. This invention not only filters out sensors that cannot effectively characterize system performance and status, thereby improving the accuracy of prediction and diagnosis, but also improves computational efficiency, thus addressing the challenge of high-dimensional, complex information failing to meet online diagnostic requirements.

Claims

1. A fuel cell sensor screening and modeling method considering the effects of fault degradation, characterized by: Modeling includes a fuel cell sensor optimization and screening model, a fuel cell simulation model, and machine learning algorithms. The fuel cell simulation model is used to calculate the sensitivity of the model's output parameters. The health parameters involved in the simulation model, such as current density, activation area, and liquid water content, are used as inputs. The set of sensor parameters predicted by the simulation model is used as the output. The functional expression of the fuel cell simulation model is: α=F(I,A act (1) In the formula, I represents the fuel cell current density, A act α represents the activated area, s represents the liquid water content; α represents the set of sensor parameters, including V and T. a ,T c ,RH c ,q a,in ,q a,out ,q c,in ,q c,out Where V represents the output voltage, and T... a T represents the anode temperature. c Represents the cathode temperature, RH c Represents cathode humidity, q a,in q represents the flow rate at the anode inlet; a,out q represents the flow rate at the anode outlet. c,in q represents the flow rate at the cathode inlet; c,out The expression for calculating the sensitivity set of fuel cell output parameters to three health parameters—current density, activated area, and liquid water content—representing the flow rate at the cathode outlet, is as follows: In the formula, S I S represents the set of sensitivity of fuel cell sensor parameters to current density; Aact S represents the set of sensitivity of fuel cell sensor parameters to the activated area; s The table shows the sensitivity set of fuel cell sensor parameters to liquid water content; I1 represents the current density value before the change; I2 represents the current density value after the change; F(I1,A) act F(I²,A) represents the simulation model output corresponding to the current density before the change; act ,s) represents the simulation model output corresponding to the changed current density; A act1 A represents the value of the activated area before the change; act2 F(I,A) represents the value of the activated area after the change. act1 F(I,A) represents the simulation model output corresponding to the activated area before the change; act2 ,s) represents the simulation model output corresponding to the changed activated area; s1 represents the liquid water content before the change; s2 represents the liquid water content after the change; F(I,A act ,s1) represents the simulation model output corresponding to the liquid water content before the change; F(I,A act ,s2) represents the simulation model output corresponding to the changed liquid water content. Sensor parameters have different sensitivities to different health parameters. A weighted selection method is used to assess the impact of three health parameter failures on fuel cell performance degradation. The calculation formula is as follows: In the formula, DF represents the set of weighted degradation sensitivities of the sensor; D I The performance degradation rate caused by current density failure modes; D represents the performance degradation rate caused by the active area failure mode; s The performance degradation rate caused by the liquid water content failure mode. Sensor weighted degradation sensitivity set include: The formula includes the weighted degradation sensitivity of eight sensor parameters, DF V Weighted degradation sensitivity of voltage sensors; Weighted degradation sensitivity of the anode temperature sensor; Weighted degradation sensitivity of the cathode temperature sensor; Weighted degradation sensitivity of the cathode humidity sensor; Weighted degradation sensitivity of the anode inlet flow sensor; Weighted degradation sensitivity of the anode outlet flow sensor; Weighted degradation sensitivity of the cathode inlet flow sensor; The weighted degradation sensitivity of the anode inlet flow sensor is determined by sorting the parameters in the set of weighted degradation sensitivities from largest to smallest. The functional relationship is as follows: In the formula, δ represents the set of weighted degraded sensitivities of the sensors after descending order, and sort represents the descending order function of the set; then, the elements in δ are grouped: G i =(δ1,...,δ i ) i=1,2,...,8 (8) In the formula, i represents the value of the group; G i This represents the set of sensors contained in the i-th group; δ1 represents the sensor with the largest weighted degradation sensitivity value in the δ set, δ i This represents the sensor with the i-th largest weighted degradation sensitivity value in the δ set. Using machine learning algorithms, the set of sensors contained in each group is taken as input, and the group i with the highest accuracy is found, resulting in G. i This process involves selecting the optimal set of sensors to filter out redundant sensors and optimize the sensor selection process.

2. The fuel cell sensor screening and modeling method considering the impact of fault degradation according to claim 1, characterized in that: The machine learning algorithms include: adaptive neural fuzzy systems, convolutional neural networks, artificial neural networks, and long short-term memory artificial neural networks.

Images