A method and device for predicting the remaining service life of PEMFC

Through the combination of UKF filtering and GRU network, the problems of insufficient output voltage index and unconsidered randomness in PEMFC residual service life prediction are solved, and more accurate life prediction is achieved and prediction accuracy is improved.

CN114814589BActive Publication Date: 2024-11-15UNIV OF SCI & TECH BEIJING
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210068376.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-20
Publication Date
2024-11-15
Estimated Expiration
2042-01-20

AI Technical Summary

Technical Problem

The prior art is difficult to accurately predict the remaining service life of the proton exchange membrane fuel cell (PEMFC), and the method based on the output voltage as a degradation index is not sufficient to describe system degradation, and it fails to effectively consider randomness and uncertainty.

Method used

Using the method based on UKF filtering and GRU network, the semi-empirical degradation model is introduced, and the SOH index of health status is selected, and the random degradation model is carried out. The three-source uncertainty is considered in real time, and the degradation state is estimated in real time by introducing a semi-empirical degradation model. The network makes the remaining service life prediction.

Benefits of technology

It improves prediction accuracy, overcomes the limitation of directly using the output voltage as a degradation indicator, and achieves a more reasonable and accurate residual service life prediction, which improves prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114814589B_ABST
    Figure CN114814589B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of fuel cell technology, and in particular to a method and device for predicting the remaining service life of a PEMFC, the method comprising: collecting the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC at different times, and preprocessing the collected data; introducing a semi-empirical degradation model, selecting a health state SOH index that can characterize the degradation of the PEMFC; stochastic degradation modeling of the SOH index to obtain a performance degradation state space model; using a UKF filtering algorithm to estimate the SOH index data online according to the output voltage data; using an EM algorithm to update the parameters in the performance degradation state space model according to the output voltage data; training the initial GRU network model, and predicting the remaining service life according to the trained GRU network model. The present invention can improve the accuracy of life prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of fuel cells, and in particular to a method and a device for predicting the remaining service life of a PEMFC. Background Art

[0002] As a new type of power energy supply device with broad application prospects, proton exchange membrane fuel cell (PEMFC) is widely used in vehicle transportation, portable equipment and cogeneration due to its advantages of high energy density, high energy conversion rate, environmental friendliness and low operating temperature, and has become a research hotspot for alternative energy devices. However, in a long-term operating environment (such as heavy load, start-up and shutdown, etc.), the internal key components of PEMFC, such as catalysts, bipolar plates and membrane electrodes, will gradually age, resulting in the degradation of PEMFC performance. At the same time, problems such as short life and high maintenance cost also restrict the development of PEMFC. RUL (Remaining Useful Life) as the core technology of prediction and health management can predict the remaining life of PEMFC system. According to the prediction results, the corresponding maintenance strategy can be formulated in advance to minimize the maintenance time and reduce the maintenance cost, so as to ensure the healthy and continuous operation of the system.

[0003] At present, the methods for predicting the remaining service life of PEMFC mainly include the method based on mechanism model and the method based on data drive. The method based on mechanism model needs to know the internal degradation mechanism of PEMFC system, but the internal changes are often unpredictable, and the degradation mechanism is not fully understood, so it is difficult to establish an accurate degradation model from the perspective of internal mechanism. Although the method based on data drive does not need to know the degradation mechanism of the battery, it needs to train the network through data, and it is highly dependent on data.

[0004] The patent number CN113608137A and the patent name "A method for predicting the life of a proton exchange membrane fuel cell stack" discloses a method for predicting the life of a proton exchange membrane fuel cell stack, which analyzes the water transmission of key materials such as the proton membrane, catalyst, and gas diffusion layer inside the stack, and uses the water content as a basis to determine the attenuation of the performance of the stack, thereby estimating the life of the stack. The patent number CN111044926A and the patent name "method for predicting the life of a proton exchange membrane fuel cell" discloses a method for predicting the life of a proton exchange membrane fuel cell, which uses the output voltage as a degradation indicator, establishes a linear state equation of the voltage decay model under static action, and uses the particle filter method to estimate the voltage in the learning stage, and predicts the voltage value in the prediction stage, thereby predicting the remaining life. The patent number is CN111339712A, the patent name is "Proton Exchange Membrane Fuel Cell Remaining Life Prediction Method", and the patent number is CN111310387A, the patent name is "A Fuel Cell Life Prediction Method". These two patents also use output voltage as a performance degradation indicator. They respectively construct a prediction model that combines a deep belief network and an extreme learning machine, and a prediction model based on variational mode decomposition and a deep belief network, extract high-quality features of battery degradation from historical output voltages to learn the degradation trend and predict the remaining life.

[0005] The existing methods still have the following shortcomings:

[0006] 1) In view of the difficulty in establishing an accurate degradation mechanism model for PEMFC and the fact that data-driven methods require a large amount of data support, it is necessary to combine the advantages of these two methods and find a hybrid model- and data-driven method to predict the remaining life of the PEMFC system.

[0007] 2) Most existing studies directly use output voltage as a performance degradation indicator to predict RUL. However, existing studies have shown that it is two internal key parameters, maximum current density and total resistance, rather than output voltage, that actually cause system degradation. Therefore, choosing output voltage as a degradation indicator is not enough to accurately describe system degradation.

[0008] 3) Affected by the combined effects of external environment and internal factors, the degradation of the system is random. Three sources of uncertainty, namely time uncertainty, individual differences and measurement uncertainty, are prevalent in the degradation process of the system. However, most of the existing research on PEMFC has not taken this issue into consideration. Summary of the invention

[0009] The embodiment of the present invention provides a method and device for predicting the remaining service life of a PEMFC. The technical solution is as follows:

[0010] In one aspect, a method for predicting the remaining service life of a PEMFC is provided, the method being implemented by an electronic device, the method comprising:

[0011] Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, and pre-process the collected data;

[0012] By introducing a semi-empirical degradation model, the health status SOH index that can characterize the degradation of PEMFC is selected;

[0013] The SOH index is subjected to random degradation modeling to obtain a performance degradation state space model;

[0014] According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index data is estimated online using the UKF filtering algorithm;

[0015] According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the EM algorithm is used to update the parameters in the performance degradation state space model;

[0016] The initial GRU network model is trained, and the remaining service life is predicted based on the trained GRU network model.

[0017] Optionally, collecting the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC at different times and preprocessing the collected data include:

[0018] Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, find the outlier points in the data by using the mean value ± 3 times the standard deviation method, and replace the outlier points with adjacent values;

[0019] After the adjacent values ​​are replaced, the data is divided into hours, and the average value of all values ​​in each hour is taken as the value at the current moment to obtain the discrete time point t k Output voltage data under (k=1,2,...,T)

[0020] Optionally, the semi-empirical degradation model is introduced to select a health status SOH index capable of characterizing the degradation of PEMFC, including:

[0021] Based on the semi-empirical degradation model, for a PEMFC stack containing multiple sub-cells, a method for characterizing the overall output voltage is determined;

[0022] By using the overall output voltage characterization method, the unknown equivalent resistance R and the maximum current density J are determined. max It is the main factor affecting the degradation of PEMFC;

[0023] By introducing the variable α, R and J max The changing trend over time is expressed, and α is selected as the SOH indicator to characterize the health status of the system.

[0024] Optionally, performing random degradation modeling on the SOH indicator to obtain a performance degradation state space model includes:

[0025] For the SOH indicator, a Wiener process model driven by Brownian motion is established;

[0026] Establishing an update mechanism for the random drift coefficient in the Wiener process model of the SOH indicator;

[0027] The SOH index and its random drift coefficient, which can reflect the degradation of PEMFC, are considered as system state variables;

[0028] Based on the semi-empirical degradation model, taking into account the influence of measurement error, the relationship between output voltage measurement and SOH index is established and used as the measurement equation; considering the three-source uncertainty, for the above system state variables, a performance degradation state space model is constructed based on the Wiener process model, the update mechanism of random drift coefficient and the measurement equation.

[0029] Optionally, the online estimation of SOH index data using a UKF filtering algorithm based on the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC includes:

[0030] According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index and its random drift coefficient are estimated online in real time through the UKF algorithm.

[0031] Optionally, the updating of parameters in the performance degradation state space model using an EM algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC includes:

[0032] Set the initial unknown parameters;

[0033] Based on Markov theory, the log-likelihood function of complete data with unknown parameters is calculated;

[0034] The EM algorithm is used to estimate the unknown parameters. Specifically, in the E step, the state variables are regarded as implicit variables and the expectation of the log-likelihood function is calculated; in the M step, the expectation of the log-likelihood function is maximized to obtain the update result of the unknown parameters.

[0035] Optionally, the initial GRU network model is trained, including:

[0036] Using the estimated value of the filtered SOH indicator data as a training sample data set, and preprocessing the training sample data set;

[0037] The preprocessed training sample data set is input into the initial GRU network model, and the parameters of the GRU network model are determined through the loss function to obtain the trained GRU network model.

[0038] Optionally, the remaining useful life prediction is performed based on the trained GRU network model, including:

[0039] The degradation state at the current moment is used as the model input and input into the trained GRU network model to obtain the remaining service life prediction.

[0040] On the other hand, a device for predicting the remaining service life of a PEMFC is provided, and the device is applied to a method for predicting the remaining service life of a PEMFC, and the device comprises:

[0041] A data processing module is used to collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC at different times, and pre-process the collected data;

[0042] A selection module is used to select a SOH index capable of characterizing PEMFC degradation by introducing a semi-empirical degradation model;

[0043] A building module is used to perform random degradation modeling on the SOH indicator to obtain a performance degradation state space model;

[0044] An estimation module, used to estimate SOH index data online using a UKF filtering algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC;

[0045] An updating module, used to update the parameters in the performance degradation state space model using an EM algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC;

[0046] The prediction module is used to train the initial GRU network model and predict the remaining useful life based on the trained GRU network model.

[0047] Optionally, the data processing module is used to:

[0048] Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, find the outlier points in the data by using the mean value ± 3 times the standard deviation method, and replace the outlier points with adjacent values;

[0049] After the adjacent values ​​are replaced, the data is divided into hours, and the average value of all values ​​in each hour is taken as the value at the current moment to obtain the discrete time point t k Output voltage data under (k=1,2,...,T)

[0050] Optionally, the selected module is used to:

[0051] Based on the semi-empirical degradation model, for a PEMFC stack containing multiple sub-cells, a method for characterizing the overall output voltage is determined;

[0052] By characterizing the overall output voltage, the unknown equivalent resistance R and the maximum current density J are determined. max It is the main factor affecting the degradation of PEMFC;

[0053] By introducing the variable α, R and J max The changing trend over time is expressed, and α is selected as the SOH indicator to characterize the health status of the system.

[0054] Optionally, the building block is used to:

[0055] For the SOH indicator, a Wiener process model driven by Brownian motion is established;

[0056] Establishing an update mechanism for the random drift coefficient in the Wiener process model of the SOH indicator;

[0057] The SOH index and its random drift coefficient, which can reflect the degradation of PEMFC, are considered as system state variables;

[0058] Based on the semi-empirical degradation model, the relationship between the output voltage measurement and the SOH index is established and used as the measurement equation taking into account the influence of measurement error. Considering the three-source uncertainty, a performance degradation state space model is constructed for the above system state variables based on the Wiener process model, the update mechanism of the random drift coefficient and the measurement equation.

[0059] Optionally, the estimation module is used to:

[0060] According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index and its random drift coefficient are estimated online in real time through the UKF algorithm.

[0061] Optionally, the update module is used to:

[0062] Set the initial unknown parameters;

[0063] Based on Markov theory, the log-likelihood function of complete data with unknown parameters is calculated;

[0064] The EM algorithm is used to estimate the unknown parameters. Specifically, in the E step, the state variables are regarded as implicit variables and the expectation of the log-likelihood function is calculated; in the M step, the expectation of the log-likelihood function is maximized to obtain the update result of the unknown parameters.

[0065] Optionally, the prediction module is used to:

[0066] Using the estimated value of the filtered SOH indicator data as a training sample data set, and preprocessing the training sample data set;

[0067] The preprocessed training sample data set is input into the initial GRU network model, and the parameters of the GRU network model are determined through the loss function to obtain the trained GRU network model.

[0068] Optionally, the prediction module is used to:

[0069] The degradation state at the current moment is used as the model input and input into the trained GRU network model to obtain the remaining service life prediction.

[0070] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned method for predicting the remaining service life of a PEMFC.

[0071] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned method for predicting the remaining useful life of a PEMFC.

[0072] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0073] In an embodiment of the present invention, a method for predicting the remaining useful life of PEMFC based on UKF filtering and GRU network is proposed, which combines the long-term prediction ability of GRU network and the advantage of Wiener process in modeling random uncertainty to perform RUL prediction. By introducing the voltage semi-empirical degradation model, the SOH index that can characterize the degradation of PEMFC is selected, and the SOH index is randomly degraded and modeled, and three sources of uncertainty are considered in the model. In view of the nonlinear relationship between the degradation state and the measured voltage, the UKF algorithm is used to estimate the degradation state in real time, and the GRU network is combined to perform RUL prediction. It overcomes the limitations of directly using the monitoring voltage as a degradation indicator in existing studies and the problem that the randomness and uncertainty of the degradation process are not considered in the existing RUL prediction method for PEMFC hybrid drive, so that the RUL prediction result is more reasonable and accurate, thereby achieving higher RUL prediction accuracy and improving prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0075] Figure 1 This is a flow chart of a method for predicting the remaining service life of a PEMFC provided by an embodiment of the present invention;

[0076] Figure 2 It is a schematic diagram of a stack real measurement voltage provided by an embodiment of the present invention;

[0077] Figure 3 This is a block diagram of a device for predicting the remaining service life of a PEMFC provided by an embodiment of the present invention;

[0078] Figure 4 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0079] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0080] The embodiment of the present invention provides a method for predicting the remaining service life of a PEMFC, which can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The flowchart of the method for predicting the remaining service life of PEMFC is shown in the figure. The processing flow of the method may include the following steps:

[0081] S101, collecting the current, current density, hydrogen and air inlet and outlet pressures, and output voltage data of the PEMFC at different times, and preprocessing the collected data.

[0082] Optionally, the current, current density, hydrogen and air inlet and outlet pressures, and output voltage data of the PEMFC at different times are collected, and the collected data are preprocessed, including:

[0083] Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, find the outlier points in the data by using the mean value ± 3 times the standard deviation method, and replace the outlier points with adjacent values;

[0084] After the adjacent values ​​are replaced, the data is divided into hours, and the average value of all values ​​in each hour is taken as the value at the current moment to obtain the discrete time point t k Output voltage data under (k=1,2,...,T)

[0085] S102. By introducing a semi-empirical degradation model, a health status SOH indicator capable of characterizing PEMFC degradation is selected.

[0086] Optionally, by introducing a semi-empirical degradation model, a health state SOH index capable of characterizing PEMFC degradation is selected, including:

[0087] According to the semi-empirical degradation model, for a PEMFC battery stack containing multiple sub-unit batteries, a characterization method of the overall output voltage is determined; through the characterization method of the overall output voltage, the unknown equivalent resistance R and the maximum current density J are determined. max is the main factor affecting PEMFC degradation; by introducing the variable α, R and J max The changing trend over time is expressed, and α is selected as the SOH indicator to characterize the health status of the system.

[0088] Among them, the unknown equivalent resistance R is the equivalent membrane resistance R M And the equivalent contact resistance R C The sum value of .

[0089] In a feasible implementation, research shows that the output voltage is not the main factor affecting PEMFC degradation, but two parameters within the system: maximum current density and total resistance. Therefore, using the output voltage as a degradation indicator is not enough to accurately characterize the degradation of the system. To this end, this section introduces a voltage semi-empirical degradation model to select a health state (SOH) indicator that can characterize system degradation.

[0090] According to the semi-empirical degradation model, for a PEMFC stack containing n sub-cells, its overall output voltage can be expressed as:

[0091] V st =n(E Nernst -V act -V ohmic -V con ) (1)

[0092] Among them, V st represents the stack voltage, n represents the number of sub-unit batteries, E Nernst Represents the reversible voltage, V act represents the voltage drop associated with anode and cathode activation, V ohmic Represents the ohmic voltage drop, V con represents the voltage drop caused by the decrease in oxygen and hydrogen concentrations, such as Figure 2 The following is a schematic diagram of the actual measured voltage of the stack. ohmic and V con The empirical calculation formula is:

[0093] V ohmic =i FC (R M +R C ) (2)

[0094]

[0095] Among them, i FC Represents the operating current of the battery, R M and R C Represents the equivalent membrane resistance for proton conduction and the equivalent contact resistance for electron conduction. J max is the maximum current density. B is a constant. J is the actual battery current density.

[0096] The study found that in equations (2) and (3), the parameter R M , R C and J max During the degradation process of PEMFC performance, there will be significant changes relative to the initial value, and R M , R C and J max There is a certain correlation between the trends that change over time. Therefore, the present invention uses the variable α to adjust R and J max The changing trend over time is expressed as shown in formula (4) and formula (5):

[0097] R(k)=R0(1+α(k)) (4)

[0098]

[0099] Among them, R represents the sum of the equivalent film resistance and the equivalent contact resistance, which can be called the total resistance, R0 and Represents R and J respectively max The initial value of α represents the parameters R and J max The degree of change relative to the initial value.

[0100] From equations (4) and (5), we know that α represents the degree of change of system model parameters and can reflect the internal degradation state of the system. Therefore, α is selected as the indicator to characterize the system SOH.

[0101] S103 . Perform random degradation modeling on the SOH indicator to obtain a performance degradation state space model.

[0102] Optionally, random degradation modeling is performed on the SOH indicator to obtain a performance degradation state space model, including:

[0103] For the SOH indicator, a Wiener process model driven by Brownian motion is established;

[0104] Establish an updating mechanism for the random drift coefficient in the Wiener process model of the above SOH index;

[0105] The SOH index and its random drift coefficient, which can reflect the degradation of PEMFC, are considered as system state variables;

[0106] Based on the semi-empirical degradation model, taking into account the influence of measurement error, the relationship between output voltage measurement and SOH index is established and used as the measurement equation; considering the three-source uncertainty, for the above system state variables, a performance degradation state space model is constructed based on the Wiener process model, the update mechanism of random drift coefficient and the measurement equation.

[0107] In a feasible implementation, in existing studies on PEMFC, the SOH index α is generally modeled as a first-order or second-order evolution model. This modeling method ignores the randomness of the degradation process. Therefore, based on the above selection of α as the index to characterize the system SOH, this section will perform random degradation modeling on α while fully considering the randomness and uncertainty of the degradation process, and at the same time incorporate three sources of uncertainty into the model. The three sources of uncertainty specifically refer to: the inherent time-varying uncertainty of the degradation process as time changes, the individual differences shown by the similar degradation paths of different individuals of the same type of equipment but different actual degradation rates of each individual, and the measurement uncertainty shown by the difference between the measured degradation data and the actual performance degradation state. Based on the above description, the model is established as a Wiener process driven by Brownian motion:

[0108] α(t)=α(0)+λt+σB(t) (6)

[0109] Among them, α(t) is the degradation amount at time t, α(0) is the initial degradation state; λ is the drift coefficient, which is used to describe individual differences and satisfies λ~N(μ λ , σ λ 2 ); σ>0 is the diffusion coefficient, {B(t), t>0} is the standard Brownian motion, which describes the time uncertainty of the degradation process and satisfies σB(t)~N(0,σ 2 t).

[0110] Based on the semi-empirical degradation model, the output voltage measurement V is established by considering the influence of measurement error. st Relationship with SOH indicator α:

[0111] V st (t)=h(α(t))+ε(t) (7)

[0112]

[0113] Where ε is the observation noise, which is used to describe the measurement uncertainty and satisfies ε~N(0,γ 2 S104: According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the UKF filtering algorithm is used to estimate the SOH index data online.

[0114] According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index and its random drift coefficient are estimated online in real time through the UKF algorithm.

[0115] In a feasible implementation, it can be seen from equations (7) and (8) that the SOH index α is related to the output voltage V st There is a nonlinear relationship between them, and there is measurement noise in the monitored voltage. If you want to realize state estimation in a nonlinear system, the traditional linear filtering algorithm is no longer applicable, and a nonlinear filtering algorithm needs to be used. The extended Kalman filter is a commonly used nonlinear filtering method, but it needs to solve the Jacobian matrix, and ignores the higher-order terms above the second order during the linearization process, resulting in a large estimation error. The UKF algorithm does not need to calculate the Jacobian matrix, and has a third-order approximation accuracy for Gaussian noise. For this reason, this section will use the UKF algorithm to estimate the degraded state α.

[0116] Further expand the drift coefficient to the drift state and establish the current t k The update mechanism of the moment random parameter λ over time:

[0117] λ k =λ k-1 +v (9)

[0118] Where v~N(0,υ 2 ), initial parameters λ0~N(μ λ ,σ λ 2 ). Therefore, based on the Wiener process model of the SOH index (6), the update mechanism of the random drift coefficient (9), and the relationship between the output voltage and the SOH index (7), it can be obtained that:

[0119]

[0120] where h is given by (8) and the reversible voltage E Nernst and the voltage drop V associated with anode and cathode activation act It can be calculated from the current, hydrogen and air inlet and outlet pressures:

[0121]

[0122]

[0123] Among them, T emp Represents the absolute temperature of the battery working environment (K), i FC Represents the operating current of the battery (A), represents the oxygen concentration in the cathode catalytic interface (mol / cm 3 ),ξ i (i=1,2,3,4) represents the parameter coefficient of each unit model.

[0124] Further degenerate state α k and drift state λ k Considering them together as state variables, the performance degradation state space model considering the three sources of uncertainty can be constructed as:

[0125]

[0126] where x k ∈R 2×1 , A∈R 2×2 ,ω k ∈R 2×1 And ω k ∈N(0,Q k ), specifically, written in vector form

[0127]

[0128] Let x 1:k ={x1,x2,…,x k} represents the degradation state data set, defining the given output voltage observation data y 1:k (For ease of representation, ) at each moment x k The expectation and variance of are:

[0129]

[0130] in, ρ 2 α,k =var(α k |y 1:k ), ρ 2 λ,k =var(λ k |y 1:k ), ρ 2 αλ,k = cov(α k λ k |y 1:k ).

[0131] The mean value of the state variable at time k-1 is known Under the conditions of and variance P, the UKF filtering steps are as follows:

[0132] (1) Calculate Sigma Points

[0133]

[0134] (2) Proportional correction of Sigma point set

[0135]

[0136] Where λ = α 2 (n+ξ)-n is the parameter to be selected, n is the state dimension, α is a constant with a value of [0.0001 1], and β≥0.

[0137] (3) Time update

[0138]

[0139]

[0140]

[0141] (4) Measurement update

[0142]

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149] According to the monitored output voltage data, the UKF algorithm is used to update the state equation and the measurement equation, so as to jointly estimate the state variable x and obtain the estimated value of α

[0150] S105 . According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the EM algorithm is used to update the parameters in the performance degradation state space model.

[0151] Optionally, according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, an EM algorithm is used to update the parameters in the performance degradation state space model, including:

[0152] Set the initial unknown parameters;

[0153] Based on Markov theory, the log-likelihood function of complete data with unknown parameters is calculated;

[0154] The EM algorithm is used to estimate the unknown parameters. Specifically, in the E step, the state variables are regarded as implicit variables and the expectation of the log-likelihood function is calculated; in the M step, the expectation of the log-likelihood function is maximized to obtain the update result of the unknown parameters.

[0155] In a feasible implementation, in the process of using the UKF algorithm for state estimation in the previous part, the process noise and measurement noise are artificially given and cannot represent the impact of random noise in the general sense. Moreover, in the state space model (11), the state variable x is implicit, and the maximum likelihood estimation cannot be used to solve the unknown parameters. Therefore, in order to realize the adaptive estimation and update of model parameters, the EM algorithm is used to deal with the noise parameter estimation problem with implicit variables. Let θ = (σ 2 ,υ 2 ,γ 2 ).

[0156] According to Markov theory, the log-likelihood function of complete data with unknown parameters θ is calculated as:

[0157]

[0158] The specific steps of estimating the unknown parameter θ using the EM algorithm are:

[0159] Step E: Calculate the expectation of the log-likelihood function of (24) under the complete data set:

[0160]

[0161] in Represents the remeasurement data set y 1:k The result of the i-th parameter iteration under , The iterative process starts from the i-th estimate At the beginning, the parameters are updated at the i+1th iteration until convergence and the iteration ends.

[0162] In formula (25), Tr is the trace operation of the matrix, Ψ and Ω are conditional expectations. In order to fully implement the EM algorithm, these conditional expectations are solved according to the unscented Kalman smoothing (URTSS) algorithm.

[0163] Through the URTSS smoothing algorithm, the conditional expectation of Ψ and Ω can be obtained as follows:

[0164]

[0165] Among them, x j|k , P j|k and M j|k They are respectively the smoothed mean, smoothed variance and smoothed covariance results of the latent variable at time j based on the current time k.

[0166] M step: maximize the log-likelihood function

[0167]

[0168] By taking the partial derivative of the expected likelihood function, that is, Get the parameter estimation results at each moment:

[0169]

[0170] S106: Train the initial GRU network model, and predict the remaining useful life based on the trained GRU network model.

[0171] Optionally, the initial GRU network model is trained, including:

[0172] The estimated value of the filtered SOH index data is used as a training sample data set, and the training sample data set is preprocessed;

[0173] The preprocessed training sample data set is input into the initial GRU network model, and the parameters of the GRU network model are determined through the loss function to obtain the trained GRU network model.

[0174] Optionally, the remaining useful life prediction is performed based on the trained GRU network model, including:

[0175] The degradation state at the current moment is used as the model input and input into the trained GRU network model to obtain the remaining service life prediction.

[0176] In a feasible implementation, since the degradation of the system is a gradual accumulation process, it is necessary to use the long-term prediction ability of the GRU recurrent neural network to predict the SOH of the PEMFC system. The above steps realize the estimation of the degradation state α. Since α can reflect the actual degradation state of the system, the estimated value of α after filtering is directly used as a training sample to train the GRU network, so as to achieve the purpose of predicting RUL. The adopted GRU prediction framework is mainly composed of input layer, hidden layer, output layer and network training. The input layer is mainly responsible for dividing and normalizing the data set representing the degradation state α of the system to meet the requirements of the network input. The hidden layer uses the memory module of a single node to build a GRU network, and the output layer is responsible for outputting the prediction results. The specific implementation steps of the prediction framework are:

[0177] (1) Data preprocessing

[0178] Before using the neural network, the data set is first divided. The first 60% of the data is used as the training set, and the last 40% is used as the test set. Secondly, the Min-Max standardization method is used to normalize the data in the data set so that the results are mapped to the interval [0,1].

[0179] (2) Network training

[0180] The normalized training set is segmented. Set the window length to N+1, the step length to 1, and slide in the data set α. Each slide results in N+1 The first N values ​​in the value form the input vector, and the N+1th The value is used as the expected output value to construct the training sample, thus obtaining the training sample set. At the same time, the system model goes through the back propagation process, and then continuously iterates and updates the weights to obtain the minimum value of the loss function Loss (the loss function Loss is represented by the mean square error between the output prediction value and the actual value), until the appropriate weight is found, and finally the parameters of the GRU hidden layer are determined.

[0181] (3) GRU model parameter optimization

[0182] In the back-propagation process, the gradient descent method is used to find the model parameters that minimize the loss function. The Adam (Adaptive Moment Estimation) algorithm is used to update the parameters, and the Dropout method is used to solve the overfitting phenomenon during the training process, making the model more generalizable.

[0183] (4) GRU prediction

[0184] Assuming that the prediction start time is m, when the output values ​​at the current time m and the previous N-1 times are used as the input vector of the GRU, the output prediction value at time m+1 can be obtained. Then, when the output values ​​at the previous N-1 times of time m+1 and the output prediction value at time m+1 are input into the GRU network as new input vectors, the output prediction value at time m+2 can be obtained, and so on, the final prediction sequence can be obtained.

[0185] After the GRU model training is completed, the RUL prediction can be performed. Based on the concept of first arrival time, the current t k The remaining life at the moment L k for

[0186] L k = inf{l k >0:α(t k +l k )≥α th} (29)

[0187] Among them, α(t k ) is the current time t k The degenerate state of α th is the failure threshold of the degradation state.

[0188] In relevant literature, the voltage failure threshold is generally defined as 96% of the rated voltage. Since the present invention uses the state variable α as the SOH indicator of PEMFC, the failure threshold α th It is the α value at the time corresponding to the voltage failure threshold.

[0189] In an embodiment of the present invention, a method for predicting the remaining useful life of PEMFC based on UKF filtering and GRU network is proposed, which combines the long-term prediction ability of GRU network and the advantage of Wiener process in modeling random uncertainty to perform RUL prediction. By introducing the voltage semi-empirical degradation model, the SOH index that can characterize the degradation of PEMFC is selected, and the SOH index is randomly degraded and modeled, and three sources of uncertainty are considered in the model. In view of the nonlinear relationship between the degradation state and the measured voltage, the UKF algorithm is used to estimate the degradation state in real time, and the GRU network is combined to perform RUL prediction. It overcomes the limitations of directly using the monitoring voltage as a degradation indicator in existing studies and the problem that the randomness and uncertainty of the degradation process are not considered in the existing RUL prediction method for PEMFC hybrid drive, so that the RUL prediction result is more reasonable and accurate, thereby achieving higher RUL prediction accuracy and improving prediction accuracy.

[0190] Figure 3 1 is a block diagram of a device for predicting the remaining service life of a PEMFC according to an exemplary embodiment. Figure 3 The device includes a data processing module 310, a selection module 320, a construction module 330, an estimation module 340, an update module 350 and a prediction module 360, wherein:

[0191] The data processing module 310 is used to collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC at different times, and pre-process the collected data;

[0192] A selection module 320 is used to select a SOH index capable of characterizing PEMFC degradation by introducing a semi-empirical degradation model;

[0193] A construction module 330 is used to perform random degradation modeling on the SOH indicator to obtain a performance degradation state space model;

[0194] An estimation module 340 is used to estimate SOH index data online using a UKF filtering algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC;

[0195] An updating module 350 is used to update the parameters in the performance degradation state space model using an EM algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC;

[0196] The prediction module 360 ​​is used to train the initial GRU network model and predict the remaining useful life based on the trained GRU network model.

[0197] Optionally, the data processing module 310 is used to:

[0198] Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, find the outlier points in the data by using the mean value ± 3 times the standard deviation method, and replace the outlier points with adjacent values;

[0199] After the adjacent values ​​are replaced, the data is divided into hours, and the average value of all values ​​in each hour is taken as the value at the current moment to obtain the discrete time point t k Output voltage data under (k=1,2,...,T)

[0200] Optionally, the selection module 320 is used to:

[0201] Based on the semi-empirical degradation model, for a PEMFC stack containing multiple sub-cells, a method for characterizing the overall output voltage is determined;

[0202] By characterizing the overall output voltage, the unknown equivalent resistance R and the maximum current density J are determined. max It is the main factor affecting the degradation of PEMFC;

[0203] By introducing the variable α, R and J max The changing trend over time is expressed, and α is selected as the SOH indicator to characterize the health status of the system.

[0204] Optionally, the construction module 330 is used to:

[0205] For the SOH indicator, a Wiener process model driven by Brownian motion is established;

[0206] Establish an updating mechanism for the random drift coefficient in the Wiener process model of the above SOH index;

[0207] The SOH index and its random drift coefficient, which can reflect the degradation of PEMFC, are regarded as system state variables. Based on the semi-empirical degradation model, the relationship between the output voltage measurement and the SOH index is established and used as the measurement equation taking into account the influence of measurement error. Considering the three-source uncertainty, a performance degradation state space model is constructed for the above system state variables based on the Wiener process model, the update mechanism of the random drift coefficient and the measurement equation.

[0208] Optionally, the estimation module 340 is configured to:

[0209] According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index and its drift coefficient are estimated online in real time through the UKF algorithm.

[0210] Optionally, the updating module 350 is used to:

[0211] Set the initial unknown parameters;

[0212] Based on Markov theory, the log-likelihood function of complete data with unknown parameters is calculated;

[0213] The EM algorithm is used to estimate the unknown parameters. Specifically, in the E step, the state variables are regarded as implicit variables and the expectation of the log-likelihood function is calculated; in the M step, the expectation of the log-likelihood function is maximized to obtain the update result of the unknown parameters.

[0214] Optionally, the prediction module 360 ​​is used to:

[0215] Using the estimated value of the filtered SOH indicator data as a training sample data set, and preprocessing the training sample data set;

[0216] The preprocessed training sample data set is input into the initial GRU network model, and the parameters of the GRU network model are determined through the loss function to obtain the trained GRU network model.

[0217] Optionally, the prediction module 360 ​​is used to:

[0218] The degradation state at the current moment is used as the model input and input into the trained GRU network model to obtain the remaining service life prediction.

[0219] In an embodiment of the present invention, a method for predicting the remaining useful life of PEMFC based on UKF filtering and GRU network is proposed, which combines the long-term prediction ability of GRU network and the advantage of Wiener process in modeling random uncertainty to perform RUL prediction. By introducing the voltage semi-empirical degradation model, the SOH index that can characterize the degradation of PEMFC is selected, and the SOH index is randomly degraded and modeled, and three sources of uncertainty are considered in the model. In view of the nonlinear relationship between the degradation state and the measured voltage, the UKF algorithm is used to estimate the degradation state in real time, and the GRU network is combined to perform RUL prediction. It overcomes the limitations of directly using the monitoring voltage as a degradation indicator in existing studies and the problem that the randomness and uncertainty of the degradation process are not considered in the existing RUL prediction method for PEMFC hybrid drive, so that the RUL prediction result is more reasonable and accurate, thereby achieving higher RUL prediction accuracy and improving prediction accuracy.

[0220] Figure 4It is a structural schematic diagram of an electronic device 400 provided in an embodiment of the present invention. The electronic device 400 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 401 and one or more memories 402, wherein at least one instruction is stored in the memory 402, and the at least one instruction is loaded and executed by the processor 401 to implement the steps of the above-mentioned method for predicting the remaining service life of PEMFC.

[0221] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions, which can be executed by a processor in a terminal to perform the above-mentioned method for predicting the remaining useful life of a PEMFC. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0222] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0223] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of PEMFC, It is characterized in that The method comprises: Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, and pre-process the collected data; By introducing a semi-empirical degradation model, the health status SOH index that can characterize the degradation of PEMFC is selected; The semi-empirical degradation model is introduced to select the health status SOH index that can characterize the degradation of PEMFC, including: Based on the semi-empirical degradation model, for a PEMFC stack containing multiple sub-cells, a method for characterizing the overall output voltage is determined; By characterizing the overall output voltage, the unknown equivalent resistance R and the maximum current density J are determined. max To be the factors that affect the degradation of PEMFC; By introducing the variable α, R and J max The trend of change over time is expressed, and α is selected as the SOH indicator representing the health status of the system; The SOH index is subjected to random degradation modeling to obtain a performance degradation state space model; The random degradation modeling of the SOH indicator to obtain the performance degradation state space model includes: For the SOH indicator, a Wiener process model driven by Brownian motion is established; an update mechanism is established for the random drift coefficient in the Wiener process model of the SOH indicator; The SOH index and its random drift coefficient, which can reflect the degradation of PEMFC, are considered as system state variables; Based on the semi-empirical degradation model, the relationship between the output voltage measurement and the SOH index is established and used as the measurement equation, taking into account the influence of measurement error. Considering the three-source uncertainty, the performance degradation state space model is constructed for the above system state variables based on the Wiener process model, the update mechanism of the random drift coefficient and the measurement equation. According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index data is estimated online using the UKF filtering algorithm; According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the EM algorithm is used to update the parameters in the performance degradation state space model; Train the initial GRU network model, and predict the remaining service life based on the trained GRU network model; The remaining useful life prediction based on the trained GRU network model includes: The degradation state at the current moment is used as the model input and input into the trained GRU network model to obtain the remaining service life prediction.

2. The method according to claim 1, It is characterized in that The collecting of the PEMFC current, current density, hydrogen and air inlet and outlet pressures and output voltage data at different times and preprocessing of the collected data include: Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, find the outlier points in the data by using the mean value ± 3 times the standard deviation method, and replace the outlier points with adjacent values; After the adjacent values ​​are replaced, the data is divided into hours, and the average value of all values ​​in each hour is taken as the value at the current moment to obtain the discrete time point t k Output voltage data under (k=1,2,...,T) 3. The method according to claim 1, It is characterized in that The method of online estimating SOH index data using a UKF filtering algorithm based on the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC includes: According to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC, the SOH index and its random drift coefficient are estimated online in real time through the UKF algorithm.

4. The method according to claim 1, It is characterized in that The method of updating the parameters in the performance degradation state space model using the EM algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC includes: Set the initial unknown parameters; According to Markov theory, the log-likelihood function of complete data with unknown parameters is calculated; the unknown parameters are estimated using the EM algorithm, specifically: in the E step, the state variables are regarded as implicit variables, and the expectation of the log-likelihood function is calculated; in the M step, the expectation of the log-likelihood function is maximized to obtain the updated result of the unknown parameters.

5. The method according to claim 1, It is characterized in that The training of the initial GRU network model includes: Using the estimated value of the filtered SOH indicator data as a training sample data set, and preprocessing the training sample data set; The preprocessed training sample data set is input into the initial GRU network model, and the parameters of the GRU network model are determined through the loss function to obtain the trained GRU network model.

6. A device for predicting the remaining service life of a PEMFC, It is characterized in that The device comprises: A data processing module is used to collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC at different times, and pre-process the collected data; A selection module is used to select a SOH index capable of characterizing PEMFC degradation by introducing a semi-empirical degradation model; Wherein, the selected module is used to: Based on the semi-empirical degradation model, for a PEMFC stack containing multiple sub-cells, a method for characterizing the overall output voltage is determined; By characterizing the overall output voltage, the unknown equivalent resistance R and the maximum current density J are determined. max To be the factors that affect the degradation of PEMFC; By introducing the variable α, R and J max The trend of change over time is expressed, and α is selected as the SOH indicator representing the health status of the system; A building module is used to perform random degradation modeling on the SOH indicator to obtain a performance degradation state space model; Wherein, the building block is used for: For the SOH indicator, a Wiener process model driven by Brownian motion is established; an update mechanism is established for the random drift coefficient in the Wiener process model of the SOH indicator; The SOH index and its random drift coefficient, which can reflect the degradation of PEMFC, are considered as system state variables; Based on the semi-empirical degradation model, the relationship between the output voltage measurement and the SOH index is established and used as the measurement equation, taking into account the influence of measurement error. Considering the three-source uncertainty, the performance degradation state space model is constructed for the above system state variables based on the Wiener process model, the update mechanism of the random drift coefficient and the measurement equation. An estimation module, used to estimate SOH index data online using a UKF filtering algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC; An updating module, used to update the parameters in the performance degradation state space model using an EM algorithm according to the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of the PEMFC; The prediction module is used to train the initial GRU network model and predict the remaining service life based on the trained GRU network model; Wherein, the prediction module is used to: The degradation state at the current moment is used as the model input and input into the trained GRU network model to obtain the remaining service life prediction.

7. The device according to claim 6, It is characterized in that The data processing module is used to: Collect the current, current density, hydrogen and air inlet and outlet pressures and output voltage data of PEMFC at different times, find the outlier points in the data by using the mean value ± 3 times the standard deviation method, and replace the outlier points with adjacent values; After the adjacent values ​​are replaced, the data is divided into hours, and the average value of all values ​​in each hour is taken as the value at the current moment to obtain the discrete time point t k Output voltage data under (k=1,2,...,T)