Method and device for predicting service life of solid oxide fuel cell SOFC

Abstract

The invention provides a prediction method and a prediction device for the service life of a solid oxide fuel cell SOFC. The prediction method comprises the following steps: acquiring multi-source state data and obtaining a multi-domain feature set based on the multi-source state data; obtaining a first comprehensive health index sequence based on attention features in the multi-domain feature set; obtaining a sub-sequence based on the first comprehensive health index sequence, and dividing the sub-sequence into a low-frequency trend component and a high-frequency residual component; inputting the low-frequency trend component into a first data model, inputting the high-frequency residual component into a second data model, and respectively obtaining a low-frequency prediction sequence and a high-frequency prediction sequence; adjusting parameters of the first data model and the second data model by using the physical model, and predicting to obtain a physical model prediction sequence; and obtaining a residual service life prediction result of the SOFC based on the data model prediction sequence and the physical model prediction sequence. Therefore, the accuracy of predicting the service life of the SOFC can be improved.

Description

Technical Field

[0001] This application relates to the field of fuel cell health management technology, specifically to a method and apparatus for predicting the lifespan of a solid oxide fuel cell (SOFC). Background Technology

[0002] Solid oxide fuel cells (SOFCs) are efficient and clean energy conversion devices. However, predicting the long-term operational stability and lifespan of SOFCs remains one of the key challenges in their commercial application. Because SOFCs operate at high temperatures for extended periods, they undergo various complex degradation mechanisms, including the evolution of electrode microstructure, strontium segregation in the cathode material, aging of sealing materials, and corrosion of the interconnects, leading to a gradual decline in battery performance over time.

[0003] Currently, SOFC lifetime assessment and prediction methods mainly fall into three categories: empirical models, physical models, and machine learning models. Empirical models typically rely on empirical formulas such as the Arrhenius equation to extrapolate lifetime through accelerated aging tests. These methods heavily depend on large amounts of long-term experimental data, making it difficult to accurately reflect the true degradation characteristics under dynamic operating conditions, resulting in low prediction accuracy and generalizability. Physical models, on the other hand, establish mathematical models describing electrochemical processes such as electrode polarization and concentration polarization to make mechanistic predictions of performance degradation. While these models offer good interpretability, they involve numerous parameters that are difficult to obtain in real time, leading to high computational complexity and making them unsuitable for online lifetime prediction and health management. Machine learning models can achieve rapid performance prediction to some extent, but existing research often relies on single operating parameters such as voltage and current, employing linear regression or simple extrapolation methods. This makes it difficult to comprehensively characterize the battery's aging state and effectively capture the nonlinear and time-varying characteristics of the SOFC degradation process, resulting in insufficient stability and accuracy of the prediction results.

[0004] In summary, the existing technology lacks a method that can predict the remaining useful life (RUL) of SOFC in real time and accurately. Summary of the Invention

[0005] This application addresses the aforementioned technical problems in the existing technology. It aims to provide a method and apparatus for predicting the remaining service life (RUL) of a solid oxide fuel cell (SOFC). This method overcomes the problems of insufficient accuracy, poor real-time performance, and weak interpretability in existing SOFC life prediction techniques. By fusing multi-source heterogeneous data and incorporating physical constraints related to the SOFC attenuation mechanism, it achieves early and accurate prediction of the RUL of the SOFC, providing key technical support for SOFC operation optimization, maintenance planning, and system reliability improvement.

[0006] According to the first aspect of this application, a method for predicting the lifespan of a solid oxide fuel cell (SOFC) is provided. The prediction method includes: acquiring multi-source state data at different times during the operation of the SOFC, and obtaining a multi-domain feature set based on the multi-source state data to characterize changes in the health status of the SOFC; performing a first fusion processing on the features of interest in the multi-domain feature set at different times to obtain a first comprehensive health index sequence; obtaining a preset number of sub-sequences with different center frequencies based on the first comprehensive health index sequence, and dividing the sub-sequences into low-frequency trend components and high-frequency residual components; inputting the low-frequency trend components into a first data model for prediction, and inputting the high-frequency residual components into a second data model for prediction, respectively obtaining a low-frequency prediction sequence and a high-frequency prediction sequence, wherein the low-frequency prediction sequence and the high-frequency prediction sequence are superimposed as a data model prediction sequence; constructing a physical model based on prior physicochemical knowledge during the operation of the SOFC, adjusting the parameters of the first data model and the second data model using the physical model, and predicting a physical model prediction sequence; and obtaining a prediction result of the remaining lifespan of the SOFC based on the data model prediction sequence and the physical model prediction sequence.

[0007] According to a second aspect of this application, a device for predicting the lifespan of a solid oxide fuel cell (SOFC) is provided. The device includes: a data acquisition module configured to acquire multi-source state data at different times during SOFC operation; a feature extraction module configured to obtain a multi-domain feature set characterizing changes in the SOFC's health state based on the multi-source state data; a feature fusion module configured to perform a first fusion process on the features of interest in the multi-domain feature set at different times to obtain a first comprehensive health index sequence; and a signal decomposition module configured to obtain a preset number of subsequences with different center frequencies based on the first comprehensive health index sequence, and to divide the subsequences into low-frequency trend components and high-frequency components. The system includes: a residual component; a dual-stream prediction module configured to input the low-frequency trend component into a first data model for prediction, and input the high-frequency residual component into a second data model for prediction, thereby obtaining a low-frequency prediction sequence and a high-frequency prediction sequence, wherein the low-frequency prediction sequence and the high-frequency prediction sequence are superimposed as a data model prediction sequence; a physical constraint module configured to construct a physical model based on prior physicochemical knowledge during the SOFC operation, adjust the parameters of the first and second data models using the physical model, and predict a physical model prediction sequence; and a prediction module configured to obtain the remaining useful life prediction result of the SOFC based on the data model prediction sequence and the physical model prediction sequence.

[0008] According to a third aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for predicting the lifespan of a solid oxide fuel cell (SOFC) as described in various embodiments of this application.

[0009] According to a fourth aspect of this application, a computer program product is provided, including computer instructions for causing a computer to perform the steps of the method for predicting the lifespan of a solid oxide fuel cell (SOFC) as described in various embodiments of this application.

[0010] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows: The method for predicting the lifespan of solid oxide fuel cells (SOFCs) provided in this application addresses the core pain points of traditional prediction methods, such as single data dimension, weak model generalization, and prediction results deviating from physical laws, through the deep integration of multi-source state data fusion, frequency band prediction, and physical mechanism constraints. The multi-domain feature set obtained from multi-source state data can comprehensively characterize the health status of SOFC from multiple dimensions. The first comprehensive health index sequence can characterize the overall battery performance status of SOFC at each acquisition time. By decomposing the first comprehensive health index sequence, low-frequency trend components and high-frequency residual components are obtained. These components are then input into the first and second data models, respectively, to achieve dual-stream prediction. This fully explores the implicit information in the SOFC degradation process and achieves high-precision modeling of complex aging behavior. The dual-stream architecture prediction can take into account both long-term trends and short-term dynamics, improving the robustness of the entire prediction model. The dual-stream channel design can simultaneously capture the slowly evolving degradation trend and the rapidly fluctuating dynamic response, enabling the entire prediction model to maintain good adaptability and stability under different operating conditions, load fluctuations, and diverse degradation modes.

[0011] By introducing prior physicochemical knowledge of SOFC attenuation mechanisms, such as electrode polarization and ion migration characteristics, an effective fusion of data model-driven and physical model-driven approaches is achieved. The parameters of the first and second data models are adjusted using the physical model, ensuring that the prediction results of the data model not only possess high accuracy but also good physical consistency and interpretability. Thus, the prediction method provided in this application can achieve non-disruptive online lifetime prediction during actual SOFC operation without interrupting the system or conducting destructive experiments, providing reliable technical support and an implementation foundation for predictive maintenance, health management, and operational optimization of SOFC systems.

[0012] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above description and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0013] In drawings that are not necessarily drawn to scale, the same reference numerals may describe similar parts in different views. Similar reference numerals with different letter suffixes may indicate different examples of similar components. The drawings generally illustrate various embodiments by way of example rather than limitation, and are used together with the specification and claims to illustrate the disclosed embodiments. Such embodiments are illustrative and exemplary, and are not intended to be exhaustive or exclusive embodiments of the method, apparatus, or non-transitory computer-readable medium having instructions for implementing the method.

[0014] Figure 1 A flowchart is shown for a method for predicting the lifespan of a solid oxide fuel cell (SOFC) according to an embodiment of this application.

[0015] Figure 2 Another flowchart of a method for predicting the lifespan of a solid oxide fuel cell (SOFC) according to an embodiment of this application is shown.

[0016] Figure 3 A flowchart illustrating the prediction results of the remaining useful life of SOFC based on the data model prediction sequence and the physical model prediction sequence according to an embodiment of this application is shown.

[0017] Figure 4 A schematic diagram of a device for predicting the lifespan of a solid oxide fuel cell (SOFC) according to an embodiment of this application is shown. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of this application will be further described in detail below with reference to the accompanying drawings and specific examples, but these are not intended to limit the scope of this application.

[0019] The terms "first," "second," and similar words used in this application do not indicate any order, quantity, or importance, but are merely used for distinction. The terms "including" or "comprising," etc., used in this application mean that the element preceding the word encompasses the elements listed after the word, and do not exclude the possibility of encompassing other elements. In this application, the arrows shown in the figures for each step are merely examples of the execution order, not limitations. The technical solution of this application is not limited to the execution order described in the embodiments. The steps in the execution order can be combined, broken down, or rearranged, as long as the logical relationship of the executed content is not affected.

[0020] All terms used in this application (including technical or scientific terms) have the same meaning as understood by one of ordinary skill in the art to which this application pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein. Technologies and equipment known to one of ordinary skill in the art may not be discussed in detail, but where appropriate, such technologies and equipment should be considered part of the specification.

[0021] Figure 1 A flowchart of a method for predicting the lifespan of a solid oxide fuel cell (SOFC) according to an embodiment of this application is shown. The specific steps of the prediction method are shown in steps S101-S106. The arrows shown in the figure for each step are merely examples of the execution order and not limitations. The technical solution of this application is not limited to the execution order described in the embodiment. The steps in the execution order can be combined, decomposed, or rearranged, as long as the logical relationship of the execution content is not affected.

[0022] In step S101, multi-source state data at different times during the operation of the SOFC are obtained, and a multi-domain feature set for characterizing the changes in the health status of the SOFC is obtained based on the multi-source state data.

[0023] Specifically, the degradation of SOFC is a gradual, time-dependent, and slow process. Therefore, multi-source state data can be continuously collected at preset time intervals to obtain multi-source state data at each collection time, and the remaining lifetime of SOFC can be predicted based on the multi-source state data obtained at each collection time.

[0024] The preset time interval can be every half hour, every hour, or every day, and there is no limitation on it; it can be set by the user.

[0025] The multi-source state data can be understood as state information data collected from the SOFC operating system and related monitoring links that can comprehensively and dynamically capture the health state evolution law throughout the entire life cycle of a solid oxide fuel cell (SOFC), covering multiple dimensions and reflecting the correlation between SOFC performance and degradation.

[0026] In other words, this multi-source state data can provide complete and dynamic raw data for the health status analysis of SOFCs.

[0027] In a preferred embodiment, the multi-source state data includes the SOFC real-time operating data, electrochemical characteristic data, and dynamic response characteristic data.

[0028] The real-time operating data of the SOFC may include current density, voltage, operating temperature, fuel flow and composition, air flow, etc., without specific limitations. This real-time operating data can be used to reflect the overall operating load and operating environment of the SOFC. Specifically, the sampling frequency used to collect the real-time operating data of the SOFC can be set to 1 Hz.

[0029] The electrochemical characteristic data can be a set of parameters obtained through specific electrochemical testing methods that reflect the efficiency of core electrochemical processes such as ion transport, charge transfer, and mass diffusion within the SOFC, as well as the performance status of key components such as related electrodes and electrolytes.

[0030] For example, electrochemical impedance spectroscopy (EIS) can be used to periodically or online measure key parameters such as ohmic resistance and polarization resistance (including charge transfer resistance and diffusion resistance), with a measurement frequency range of 10 mHz to 1 mHz. Alternatively, polarization curve testing, constant current, and constant voltage step tests can be used to monitor the response signals (such as changes in impedance, current, and voltage) of SOFCs by applying specific electrical signals (such as alternating current of different frequencies, stepped current, or voltage) and inferring the efficiency of the internal electrochemical processes.

[0031] The dynamic response characteristic data can be understood as the instantaneous change curves and derived characteristics of core performance parameters (such as voltage and current) of SOFC over time when operating parameters (such as load, fuel flow, and temperature) suddenly change, obtained through high-frequency sampling. This data is used to reflect the SOFC's adaptability to operating conditions and its dynamic health status.

[0032] The health degradation of SOFCs is not only reflected in a decline in steady-state performance, such as a decrease in steady-state voltage, but also in a decline in dynamic response capability. Healthy SOFCs can adapt quickly and stably to sudden changes in operating conditions, while degraded SOFCs (such as those with blocked electrode structures or increased electrolyte ion transport resistance) exhibit characteristics such as sluggish response, violent fluctuations, and slow recovery. The dynamic response characteristic data can capture the instantaneous performance feedback patterns of SOFCs under sudden changes in operating conditions.

[0033] Among them, high-frequency voltage fluctuation signals or high-frequency current fluctuation signals can be used to reflect dynamic response characteristics. The sampling frequency of high-frequency dynamic signals is not less than 10 Hz. Through high time resolution sampling, the instantaneous changes of voltage, current and other parameters of SOFC under sudden changes in operating conditions are accurately captured, providing reliable raw data to reflect dynamic response characteristics.

[0034] Specifically, data preprocessing can be performed on the multi-source state data. For example, a timestamp-based multi-channel data alignment method can be used to ensure the consistency of data from different sampling frequencies. Interpolation techniques and sliding window algorithms are used for data cleaning and completion, and Kalman filtering is applied to smooth EIS data and voltage / current signals to eliminate random noise and retain key degradation features. Simultaneously, the Isolation Forest algorithm can be used to detect and remove abnormal data points caused by sensor anomalies or operational fluctuations, ensuring the reliability and stability of all data.

[0035] This is provided as an example only and does not constitute a limitation on any specific preprocessing method.

[0036] In this embodiment, the multi-domain feature set is a collection of multi-domain features strongly correlated with the health degradation of SOFCs extracted from the original multi-source state data. This multi-domain feature set can comprehensively reflect the different characteristics of SOFC degradation. Specifically, for the multi-source state data, appropriate signal processing and statistical analysis methods can be used to extract features strongly correlated with SOFC health degradation from multiple dimensions to obtain the multi-domain feature set.

[0037] In some embodiments of this application, the multi-domain feature set includes time-domain features, integral features, frequency-domain features, and time-frequency features. The time-domain features directly analyze raw data that changes continuously over time (such as voltage time series, temperature time series, and high-frequency fluctuation signals) without needing to transform the signal dimension, capturing the statistical attributes and trend changes of the data. For example, the time-domain features can be extracted from multi-source state data using methods such as basic statistical analysis, trend fitting and difference analysis, and sliding window statistics. Further, the time-domain features include statistical characteristics such as the mean, variance, skewness, and kurtosis of voltage, current, and resistance signals.

[0038] The integral feature is a cumulative parameter that reflects the long-term losses of SOFCs, obtained through time integration or cumulative calculation, and is used to reflect the electrochemical cumulative effect. For example, the integral feature can be extracted using time integration algorithms, cumulative difference calculation methods, or sliding window integration methods. For example, a time integration algorithm can be used to perform definite integration on time-domain signals such as power and current to obtain cumulative power generation, cumulative discharge capacity, cumulative energy loss, etc., which are used to reflect long-term energy or charge consumption.

[0039] The frequency domain features require signal transformation to convert time-domain data into frequency-domain data, capturing the amplitude and energy distribution of different frequency components. For example, impedance features at key frequency points in EIS data can be extracted using Fast Fourier Transform (FFT). Furthermore, various extraction methods such as power spectral density analysis, impedance spectrum feature extraction, and spectral peak-valley detection can also be employed to obtain frequency domain features.

[0040] The time-frequency features are used to address the problems of the time domain's inability to distinguish frequency changes and the frequency domain's inability to distinguish time distributions. They combine time-domain and frequency-domain information through local signal transformation. For example, wavelet transform can be used to extract local time-frequency features of voltage and current fluctuation signals to characterize short-time variation patterns of dynamic behavior. Alternatively, short-time Fourier transform can be used to divide the time-domain signal into multiple short windows, and FFT can be performed on each window to obtain the peak position and energy distribution of the time-frequency graph, reflecting the frequency change over time.

[0041] In step S102, the features of interest in the multi-domain feature set at different times are subjected to a first fusion process to obtain a first comprehensive health index sequence.

[0042] Specifically, in a multi-domain feature set (such as 128 initial features encompassing the time domain, frequency domain, time-frequency domain, and integral domain), not all features have the same ability to represent the health status of SOFCs. Some features may have a weak correlation with SOFC degradation, or there may be high redundancy among multiple features. To reduce feature redundancy and improve the generalization ability of the prediction model, the Fast Correlation Filtering (FCBF) algorithm can be used to select features of interest that are strongly correlated with SOFC degradation, thereby removing features with low correlation or high collinearity and providing the optimal input set for subsequent prediction model training.

[0043] For example, the features of interest may include the voltage trend slope and dynamic recovery time (e.g., the time to recover to steady state after a load change) in the time domain; ohmic resistance and polarization resistance in the frequency domain; high-frequency wavelet coefficient energy in the time-frequency domain; and cumulative capacity decay rate and cumulative impedance increment in the integral domain. This is merely an example and does not constitute a specific limitation on the features of interest. For instance, some features may be manually set as features of interest, and then a screening method may be used to select features of interest that are strongly correlated with SOFC degradation from the multi-domain feature set.

[0044] In this embodiment, the first fusion process can use a weighted fusion method to perform multi-layer fusion. For example, based on the weight allocation method of Manhattan distance and Pearson correlation coefficient, the more important features that have a greater impact on the health degradation of SOFCs can be assigned higher weights, while the less important features have a smaller impact can be assigned lower weights. The important features are mapped into a single indicator that can comprehensively reflect the degree of SOFC degradation, thereby realizing high-dimensional compression of multi-source information and quantification of health status.

[0045] The First Comprehensive Health Index (HI) is a quantitative score of the overall performance status of the SOFC. HI=1 indicates optimal SOFC performance, while HI=0 indicates that the SOFC has reached the end of its lifespan. At this point, the battery performance can no longer meet operational requirements, such as the voltage dropping below 80% of the initial voltage value or the impedance exceeding the failure threshold. The HI sequence is time-series data of the quantitative scores of the overall performance status of the SOFC at various times. For example, the HI is 0.95 at time t1, 0.93 at time t2, and 0.92 at time t3.

[0046] In step S103, a preset number of subsequences with different center frequencies are obtained based on the first comprehensive health index sequence, and the subsequences are divided into low-frequency trend components and high-frequency residual components.

[0047] The First Comprehensive Health Index (HI) sequence is a continuous time-series data formed by arranging the HI at different times (e.g., t1=0.95, t2=0.93, t3=0.92, t4=0.91, t5=0.92, t6=0.90, etc.). Its numerical changes are not a smooth decreasing trend, but rather include values ​​exhibiting both long-term slow degradation trends and short-term dynamic fluctuations. For example, a value exhibiting a continuous, monotonically decreasing HI over time represents a long-term slow degradation trend, characterized by low frequency (long variation period, such as days or months, with a center frequency ≤0.005Hz), reflecting irreversible core degradation of SOFCs, such as electrolyte aging and continuous loss of electrode active sites. Alternatively, for example, small fluctuations in the HI value based on its long-term trend represent short-term dynamic fluctuations. For instance, if the HI rises from 0.92 to 0.93 and then falls back to 0.91, its frequency characteristic is high (short cycle, such as in hours or days, with a center frequency ≥ 0.05Hz), reflecting reversible operating disturbances, such as sudden load increases, fuel flow fluctuations, or short-term noise. If the original first comprehensive health index sequence is used directly to predict the lifespan of SOFCs, high-frequency fluctuations will interfere with the judgment of the true long-term degradation trend, leading to increased prediction errors.

[0048] Specifically, an adaptive signal decomposition algorithm (such as Variational Mode Decomposition (VMD) can be used to decompose the first comprehensive health index sequence. A preset number K is determined, and the HI sequence is adaptively decomposed into K narrow-band subsequences with different center frequencies. Then, based on the spectral energy distribution, each subsequence is divided into a low-frequency trend component reflecting the irreversible long-term degradation trend of SOFC and a high-frequency residual component reflecting short-term dynamic fluctuations and capacity regeneration phenomena.

[0049] The preset quantity K can be determined based on the complexity (fluctuation level) of the HI sequence. For example, the more complex the fluctuation, the larger the value of K (more subsequences are needed to cover all frequency components). The center frequency of each subsequence is the frequency value at which the energy of the subsequence is most concentrated. Different center frequencies represent different time scale changes characterized by the subsequence. For example, low-frequency subsequences correspond to long-term changes, while high-frequency subsequences correspond to short-term changes.

[0050] In other words, in some embodiments of this application, an adaptive signal decomposition algorithm is used to decompose the first comprehensive health index sequence to obtain a preset number of subsequences with different center frequencies; power spectrum analysis is performed on each subsequence and each subsequence is divided into low-frequency trend components and high-frequency residual components according to the spectral energy distribution.

[0051] Specifically, subsequences can be classified based on their center frequency and spectral energy distribution. For example, a predetermined number of subsequences can be subjected to power spectrum analysis to calculate the energy proportion of each subsequence at different frequencies, and then the subsequences can be classified according to the concentration range of spectral energy. As an example, if the center frequency of a subsequence is extremely low (e.g., ≤0.005Hz), corresponding to a long-term slow degradation trend, it is classified as a low-frequency trend component; if the center frequency of a subsequence is relatively high (e.g., ≥0.05Hz), corresponding to a short-term dynamic fluctuation type, it is classified as a high-frequency residual component. Moreover, the irreversible long-term slow degradation trend of SOFC is the core factor determining its lifetime, so the corresponding low-frequency trend component energy proportion is usually extremely high (e.g., accounting for 70%~90% of the total energy); while short-term dynamic fluctuations are secondary interference factors, and the corresponding high-frequency residual component energy proportion is low (e.g., only 10%~30%).

[0052] In step S104, the low-frequency trend component is input into the first data model for prediction, and the high-frequency residual component is input into the second data model for prediction, thereby obtaining a low-frequency prediction sequence and a high-frequency prediction sequence, wherein the low-frequency prediction sequence and the high-frequency prediction sequence are superimposed as the data model prediction sequence.

[0053] The low-frequency trend component and the high-frequency residual component differ in their variation patterns, physical meanings, and prediction requirements. For example, the low-frequency trend component exhibits a long-term, slow, monotonous, and irreversible variation pattern, which can be used to capture long-term time-series dependencies. In contrast, the high-frequency trend component exhibits a short-term, rapid variation pattern without a fixed trend, which can be used to fit dynamic fluctuation patterns.

[0054] In this embodiment, the low-frequency trend component is input into a first data model, and the high-frequency trend component is input into a second data model. The first data model focuses on low-frequency trend prediction, while the second data model focuses on high-frequency residual prediction, outputting prediction sequences that accurately match the characteristics of the corresponding components. This dual-stream architecture design simultaneously captures both slowly evolving decay trends and rapidly fluctuating dynamic responses, ensuring the model maintains good adaptability and stability under different operating conditions, load fluctuations, and diverse decay modes. Furthermore, each data model can focus on a single pattern, resulting in faster training and more stable predictions.

[0055] In some embodiments of this application, the first data model is a temporal convolutional network, and the second data model is a long short-term memory network.

[0056] Specifically, a temporal convolutional network is used for low-frequency trend component prediction. By employing causal convolution and dilated convolution structures, it can capture long-term temporal dependencies spanning hundreds or even thousands of time steps without increasing computational cost, thus adapting to the slow and long-term changes in low-frequency trend components. For example, the kernel size of the temporal convolutional network can be set to 3, the number of layers to 5, and the dilation coefficient to a power of 2. Based on the input low-frequency trend component sequence, the temporal convolutional network learns the decay rate pattern of HI over time, such as rapid initial decay, stable mid-term decay, and accelerated late-term decay, and outputs a low-frequency prediction sequence for a future period. This low-frequency prediction sequence can reflect the long-term trend of SOFC core degradation.

[0057] By inputting high-frequency residual components into a Long Short-Term Memory (LSTM) network and leveraging its gating mechanism to effectively learn short-term fluctuations and regeneration behavior, this approach effectively handles the non-stationary and rapidly changing characteristics of high-frequency residual components, avoiding overfitting to meaningless noise. Specifically, the hidden layer size can be set to 128, the number of layers to 2, the Adam optimizer to use, the initial learning rate to be 0.001, the batch size to be 32, and the training epochs to be 500. The LSM network learns the frequency, amplitude, and duration patterns of fluctuations, such as increased fluctuation amplitude during sudden load increases and decreased fluctuation during stable conditions, outputting a high-frequency prediction sequence for the future. This high-frequency prediction sequence reflects short-term fluctuations of reversible perturbations and does not include long-term degradation trends.

[0058] By superimposing low-frequency and high-frequency prediction sequences as prediction sequences for the data model, the complete dynamic trajectory of the future health status of SOFCs can be reconstructed. This ensures the accuracy of long-term degradation trends while restoring the dynamics of short-term fluctuations, providing accurate and realistic health status inputs for subsequent remaining useful life (RUL) calculations.

[0059] In step S105, a physical model is constructed based on the prior physicochemical knowledge of the SOFC operation process. The parameters of the first data model and the second data model are adjusted using the physical model, and the predicted sequence of the physical model is obtained.

[0060] Specifically, the aforementioned physicochemical prior knowledge includes gas transport and electrochemical processes and degradation mechanisms based on SOFC stacks. SOFC power generation relies on the oxidation reaction of fuel gas (such as hydrogen) at the anode, the reduction reaction of oxygen at the cathode, and the migration of ions through the electrolyte. The efficiency of these processes directly determines the performance of the SOFC battery. Its degradation is inevitably accompanied by increased ion migration resistance (increased ohmic resistance), decreased electrode reactivity, and increased polarization resistance. Furthermore, the rate of resistance change is affected by operating conditions such as temperature and fuel composition. For example, increased temperature accelerates ion migration, but excessively high temperatures exacerbate material aging.

[0061] The attenuation mechanism includes rapid initial attenuation and slow long-term attenuation. For example, in the early stage of SOFC operation, the active sites on the electrode surface are rapidly reduced due to oxidation or sintering, resulting in a rapid drop in voltage, but the attenuation rate will gradually slow down. During the long-term operation of SOFC, the electrolyte ionic conductivity decreases, the connector is corroded, and the sealing material ages, resulting in slow and continuous performance attenuation, which is irreversible.

[0062] In this embodiment, the physicochemical prior knowledge during SOFC operation is transformed into quantifiable and computable data formulas and constraint rules to constrain the prediction results of the first and second data models to conform to real physical laws.

[0063] In some embodiments of this application, the physicochemical prior knowledge is converted into quantifiable constraint rules using a double exponential decay function; the output results of the first data model and the second data model must satisfy the output value constraint rules of the double exponential decay function, and / or, the prediction parameters involved in the prediction process of the first data model and the second data model must satisfy the parameter constraints in the double exponential decay function.

[0064] For example, a double exponential decay function can be used to describe the two main decay processes: rapid initial decay (mainly caused by electrode activity loss) and slow long-term decay (mainly caused by material aging and increased ion migration resistance), as shown in the following formula: ,in, t For runtime, Predicted by the physical model t Physical model prediction metrics at any given time; A This refers to the initial contribution coefficient of rapid initial decay (reflecting the number of initial active sites), physical range: Because the initial degradation will not exceed 50% of the healthy value, otherwise the initial battery design would be unqualified. This refers to the rapid initial decay rate constant (reflecting the rate of loss of active sites), physical range: (The initial decay rate is much faster than the long-term decay rate), and The initial decay will not be infinitely fast or infinitely slow; B This refers to the initial contribution coefficient of slow, long-term decay, reflecting the initial effects of material aging, within a physical range: The initial contribution of long-term decay is complementary to that of the initial decay, and the sum does not exceed 1; This refers to the slow, long-term decay rate constant, reflecting the aging rate of the material, within a physical range. The long-term decay rate is much slower than the initial rate, which is consistent with the aging process of materials. C This refers to the residual health value, the lowest health level of the battery before it fails. Because the materials cannot completely fail, the physical range is limited. HI will not actually drop to 0, but will be considered invalid when it is close to a residual value of 0.

[0065] The mathematical physical laws are converted into constraint rules for the data model (including the first data model and the second data model) to ensure that the output of the data model does not violate common sense physics.

[0066] Specifically, the constraint rules can be output value constraint rules and / or parameter constraints. The output value constraint rule can be understood as ensuring that the deviation between the output of the data model and the output of the physical model does not exceed a deviation threshold. For example, the deviation between the low-frequency predicted value obtained by the first data model and the predicted value of the physical model does not exceed 0.05. For instance, assuming the physical model predicts a value of 0.3 at t=5000h, then the low-frequency predicted value output by the first data model must not be lower than 0.25 or higher than 0.35; otherwise, the attenuation mechanism is violated.

[0067] Furthermore, parameter constraints can be applied to the prediction parameters involved in the data model prediction process. One such parameter is the decay rate of the SOFC, which must conform to the double exponential decay function. The physical range. For example, assuming the decay rate far exceeding upper limit If so, the data model prediction can be determined to be abnormal, and it needs to be corrected using a loss function.

[0068] It is understood that the above is for illustrative purposes only and does not constitute a limitation on any specific solution.

[0069] Returning to the embodiments of this application, the physical model prediction sequence obtained based on the physical model prediction can be used as a reference benchmark for the data model. The prediction of the data model depends on the fitting of historical data. If there is noise in the historical data, it may output abnormal results that violate physical laws. In this case, the physical model prediction sequence can be used as a benchmark to judge the abnormal prediction situation.

[0070] For example, based on a physical model using a double exponential decay function, the predicted comprehensive health index HI for every 50 hours over the next 1000 hours is: HI=0.30 (t=0), HI=0.29 (t=50), HI=0.28 (t=100)...HI=0.11 (t=1000), which conforms to the physical law of slow, long-term decay and monotonically decreasing HI. If the first data model is noisy due to a certain period of historical data, it might predict a low-frequency HI of HI=0.32 at t=200 hours, which is clearly higher than HI=0.27 at t=150 hours, indicating a recovery in the health status of the SOFC battery. This contradicts the physical mechanism of irreversible aging of SOFC materials. At this point, by comparing the large deviation between the predicted value HI=0.32 of the first data model and the physical baseline value HI=0.26, it can be directly determined that the predicted value of the first data model is unreasonable and a subsequent correction mechanism needs to be triggered to adjust the operating parameters of the first data model so that the low-frequency predicted value output by the first data model meets the constraint requirements of the predicted value of the physical model.

[0071] In practice, the constraint rules of the physical model can be embedded into the training loss function of the data model. For example, a penalty term for violating the constraint can be added to the training loss function of the data model to force the data model to learn physical laws, thereby ensuring that the predicted value output by the data model conforms to the real physical change law.

[0072] In step S106, the remaining useful life prediction result of the SOFC is obtained based on the data model prediction sequence and the physical model prediction sequence.

[0073] Specifically, physical model prediction sequences, based on general mechanisms, cannot adapt to the individual differences of SOFCs, while data model prediction sequences, learned from historical data, can capture these differences. Predicting the remaining useful life of SOFCs based on both data model and physical model prediction sequences effectively balances dynamic details and individual variations. If predictions are based solely on data model prediction sequences, abnormal situations such as sensor failures may lead to unrealistic predictions of extended or shortened remaining useful life. Physical model prediction sequences, serving as a mechanistic benchmark, can constrain the prediction results, ensuring that changes in remaining useful life conform to irreversible decay laws and improving the accuracy of SOFC remaining useful life predictions. Thus, the obtained remaining useful life prediction results have high reliability and accuracy, providing a core decision-making basis for predictive maintenance, spare parts planning, and operational optimization of SOFCs. Figure 2A flowchart of a specific embodiment is shown. In step S201, multi-source state data during SOFC operation is acquired. Then, step S202 is executed to preprocess the multi-source state data and extract multi-domain features from the preprocessed data to obtain a multi-domain feature set. In step S203, a fast correlation filtering algorithm is used to filter out features of interest that are highly correlated with SOFC degradation. In step S204, a feature fusion algorithm is used to perform a first fusion process on the features of interest to obtain a first comprehensive health index sequence. Then, an adaptive signal decomposition algorithm is used to decompose the first comprehensive health index sequence to obtain low-frequency trend components and high-frequency residual components (as in step S205). In step S206, the low-frequency trend components are input into a temporal convolutional network for prediction to obtain a low-frequency prediction sequence (as in step S207). In step S209, a physical model is constructed based on prior physical and chemical knowledge, and the physical model prediction sequence is output (as in step S210). The outputs of steps S206 and S207 are input to step S208 to determine whether the physical knowledge constraints are met. If the result of step S208 is negative, step S211 is executed to optimize the data model parameters. If the result of step S208 is positive, step S212 is executed to obtain the remaining useful life prediction result of the SOFC based on the data model prediction sequence and the physical model prediction sequence. Thus, the fusion of multi-source state data and multi-domain features significantly improves the accuracy of SOFC health status characterization, laying a high-quality foundation for remaining useful life prediction. Dual-stream prediction using the first and second data models can take into account both long-term trends and short-term fluctuations, thereby improving the robustness of model prediction. Furthermore, it can simultaneously capture the slow-evolving decay trend and the dynamic response of rapid fluctuations, enabling the model to maintain good adaptability and stability under different operating conditions, load fluctuations, and diverse decay modes.

[0074] In other embodiments of this application, the remaining useful life prediction result of the SOFC is obtained based on the data model prediction sequence and the physical model prediction sequence, specifically as follows: Figure 3 As shown. In step S301, the prediction credibility of the data model and the physical model is evaluated to obtain the evaluation result, which serves as a benchmark for dynamically allocating the first and second weights.

[0075] Specifically, the predictive reliability of the data model can be evaluated based on the correlation strength between the feature of interest and the first comprehensive health index sequence, quantified using the Pearson correlation coefficient. For example, if the correlation coefficient between the mean voltage and the first comprehensive health index sequence is greater than 0.8, it indicates that the feature of interest effectively reflects the health status of the SOFC, and the low-frequency trend component obtained from the decomposition of the first comprehensive health index sequence is of high quality, meaning the input quality of the first data model is high, and correspondingly, the predictive reliability based on the first data model is high. However, if the correlation coefficient is small (e.g., less than 0.5), it may reflect poor input quality of the first data model, resulting in low predictive reliability based on the first data model, while the physical model has higher predictive reliability.

[0076] Alternatively, the reliability of the predictions can be assessed based on the prediction deviation between the predicted values ​​output by the data model and the first comprehensive health indicator. For example, the prediction deviations at different times can be obtained by comparing the predicted values ​​at different times from the first and second data models with the first comprehensive health indicator. Assuming that the prediction deviations are all less than 0.03 in the last 100 predictions, it indicates that the data model has good generalization ability and the SOFC operating system is relatively stable, thus the reliability of the data model's predictions is high. However, if the prediction deviation is greater than 0.09, it indicates that there are sudden changes in operating conditions during the SOFC operation, and the data model has poor adaptability. In this case, the reliability of the physical model's predictions is higher, while the reliability of the data model's predictions is lower.

[0077] In other words, the predictive reliability of the data model is evaluated based on the correlation strength between the features of interest and the first comprehensive health index sequence or the deviation fluctuation between the data model prediction sequence and the first comprehensive health index sequence.

[0078] Furthermore, the predictive reliability of the physical model can be evaluated based on the degree of matching between the physical model and the SOFC attenuation mechanism.

[0079] Specifically, based on the physical model's double exponential decay function, if SOFC is in its initial stage ( k 1. Dominant rapid decay), and physical model k 1. Fitting error ≤ 0.01, or, if SOFC is in the later stage ( k 2. Dominant slow decay), and physical model k 2. A fitting error ≤ 0.02 indicates a high degree of agreement between the physical model and the SOFC attenuation mechanism, suggesting higher predictive reliability of the physical simulation. However, a fitting error > 0.05 (e.g., due to abnormal material aging) indicates a lower degree of agreement. k If the prediction is outside the physical range (2), then the reliability of the physical model's prediction is considered low.

[0080] The above is for illustrative purposes only and does not constitute a limitation on any specific solution.

[0081] In step S302, the first weight of the data model prediction sequence and the second weight of the physical model prediction sequence are adjusted based on the evaluation results, so that the weight of the model with higher prediction confidence is higher.

[0082] Specifically, as an example, the first and second weights can be determined first based on the decay stage of the SOFC, and then adjusted accordingly. For instance, assuming the SOFC is in its initial stage of operation, when the SOFC system is stable and the prediction reliability based on the data model is high, the first weight can be set higher than the second weight; for example, the first weight can be set to 0.7 and the second weight to 0.3. However, if the fitting error of the physical model decreases in the initial stage, it can be considered that the prediction reliability of the physical model has improved, and the first weight can be adaptively reduced while the second weight is increased; for example, the first weight of 0.7 can be reduced to 0.6, and the second weight of 0.3 can be increased to 0.4.

[0083] In other words, during the initial stage of SOFC operation or when the input data to the data model is of high quality and the operating conditions are stable, the prediction reliability of the data model is higher, and the first weight can be appropriately increased. When SOFC operation is nearing the end of its lifespan or when the input data quality to the data model is poor, the prediction reliability of the physical model is higher, and the second weight can be appropriately increased to give higher weights to models with higher prediction reliability.

[0084] In step S303, a second fusion process is performed based on the data model prediction sequence and its first weight, the physical model prediction sequence and its second weight, to obtain a second comprehensive health index sequence; in step S304, the remaining lifespan prediction result is determined based on the second comprehensive health index sequence and the preset SOFC lifespan termination threshold.

[0085] Specifically, the data model prediction sequence and the physical model prediction sequence can be input into the attention-weighted fusion layer, and a second fusion process can be performed based on the first and second weights to obtain the second comprehensive health index sequence. This second comprehensive health index sequence includes the second comprehensive health index at each time point.

[0086] One approach is to pre-set a SOFC lifetime termination threshold. For example, a comprehensive health index of 0.2 can be set as the SOFC lifetime termination threshold. Then, the termination time corresponding to when the second comprehensive health index reaches or first falls below 0.2 can be found from the second comprehensive health index sequence. For instance, if the second comprehensive health index is 0.2 at t=800h, then t=800h can be used as the SOFC lifetime termination time. The time difference between the current moment of the SOFC and the SOFC lifetime termination time is then the remaining useful life (RUL) of the SOFC.

[0087] This is merely one possible approach and does not constitute a limitation on any specific solution.

[0088] In some embodiments, the prediction method further includes uncertainty quantification and confidence interval assessment. The probability distribution and confidence interval of the RUL prediction result can be calculated by using quantile regression and bootstrap sampling methods, and the prediction interval at different confidence levels (such as 90% and 95%) can be output to provide a risk assessment basis for subsequent maintenance strategies.

[0089] Quantile regression is a tool for quantifying uncertainty. By modeling the predicted values ​​of the Remaining Usage Limit (RUL) at different quantiles, it directly outputs the probability distribution characteristics of the RUL. Uncertainty quantification can be understood as transforming fuzzy factors into quantifiable probabilistic information through mathematical methods. For example, the probability that the RUL falls within 800±50 hours is 95%. A confidence interval is the probability range that includes the true RUL at a specific confidence level (such as 90% or 95%). For example, a 95% confidence interval [750h, 850h] means that there is a 95% certainty that the true remaining useful life of the SOFC is between 750 and 850 hours.

[0090] SOFC's predicted RUL often contains outliers, and traditional regression is easily affected by outliers. Quantile regression, on the other hand, can more robustly characterize the distribution of RUL through the quantile loss function. At the same time, the true distribution of RUL is often skewed (e.g., the uncertainty of RUL increases and the distribution becomes right-skewed as the end of life approaches). Quantile regression does not need to assume the distribution type and can directly adapt to the characteristics of real data.

[0091] The Bootstrap sampling method is a resampling method that generates a large number of independent RUL prediction results by randomly sampling the training data and model parameters multiple times. Then, by statistically analyzing the dispersion of these results, the uncertainty is quantified, and finally, the confidence interval is determined.

[0092] The ultimate goal of uncertainty quantification and confidence interval assessment is to serve SOFC operation and maintenance decisions. Different confidence levels correspond to different risk preferences and maintenance strategies. For example, the lower limit of the 95% confidence interval can be selected as the maintenance trigger time (e.g., interval [750h, 850h], with spare parts procurement and downtime maintenance planned according to 750h).

[0093] Furthermore, the prediction method provided in this application can be updated online, ensuring the continued effectiveness of the prediction model during long-term operation. For example, a sliding window strategy can be used to update model parameters to adapt to the evolution of SOFC degradation paths. Also, for SOFC systems with different stacks or operating conditions, transfer learning techniques can be used to quickly adapt the pre-trained model to new data distributions, achieving cross-stack transfer and rapid convergence of the model.

[0094] SOFC systems exhibit significant individual and operational variations. Different reactors may show degradation rates differing by up to 20% due to variations in manufacturing processes (such as electrode thickness and electrolyte purity); even the same reactor may exhibit different degradation patterns under different operating conditions. Transfer learning allows the knowledge of pre-trained models to be transferred to new reactors or operating conditions, enabling rapid adaptation with only a small amount of new data. This improves the applicability of the prediction methods provided in the various embodiments of this application.

[0095] In some embodiments of this application, a device for predicting the lifespan of a solid oxide fuel cell (SOFC) is provided, such as... Figure 4 As shown, the prediction device 400 includes a data acquisition module 401, configured to acquire multi-source state data at different times during the operation of the SOFC; a feature extraction module 402, configured to obtain a multi-domain feature set based on the multi-source state data to characterize the changes in the health status of the SOFC; a feature fusion module 403, configured to perform a first fusion process on the features of interest in the multi-domain feature set at different times to obtain a first comprehensive health index sequence; a signal decomposition module 404, configured to obtain a preset number of subsequences with different center frequencies based on the first comprehensive health index sequence, and divide the subsequences into low-frequency trend components and high-frequency residual components; and a dual-stream prediction module 405. The configuration is as follows: the low-frequency trend component is input into a first data model for prediction, and the high-frequency residual component is input into a second data model for prediction, to obtain a low-frequency prediction sequence and a high-frequency prediction sequence, wherein the low-frequency prediction sequence and the high-frequency prediction sequence are superimposed as the data model prediction sequence; the physical constraint module 406 is configured to construct a physical model based on the physicochemical prior knowledge during the operation of the SOFC, adjust the parameters of the first data model and the second data model using the physical model, and predict the physical model prediction sequence; and the prediction module 407 is configured to obtain the remaining useful life prediction result of the SOFC based on the data model prediction sequence and the physical model prediction sequence.

[0096] In this way, the remaining service life of SOFC can be predicted in real time, and the prediction accuracy, reliability and robustness are higher, which can provide key technical support for SOFC operation optimization, maintenance plan formulation and system reliability improvement.

[0097] It should be noted that the steps of the method for predicting the lifespan of solid oxide fuel cells (SOFCs) described in the above embodiments can all be incorporated here, and will not be repeated here.

[0098] In some embodiments of this application, a computer program product is provided, including computer instructions for causing a computer to perform the steps of the method for predicting the lifespan of a solid oxide fuel cell (SOFC) as described in various embodiments of this application.

[0099] The exemplary methods described in this application can be implemented, at least in part, by a machine or computer. In some embodiments, a computer-readable storage medium stores computer program instructions that, when executed by a processor, cause the processor to perform the steps of the method for predicting the lifespan of a solid oxide fuel cell (SOFC) as described in the various embodiments of this application. The steps of the method for predicting the lifespan of a solid oxide fuel cell (SOFC) in the various embodiments described above can all be incorporated herein and implemented by a processor executing computer program instructions, and will not be elaborated further here.

[0100] Implementations of such methods may include software code, such as microcode, assembly language code, high-level language code, etc. Various software programming techniques can be used to create various programs or program modules. For example, program parts or program modules can be designed using or with the aid of Java, Python, C, C++, assembly language, or any known programming language. One or more of such software parts or modules can be integrated into a computer system and / or a computer-readable medium. Such software code may include computer-readable instructions for performing various methods. This software code can form part of a computer program product or a computer program module. Furthermore, in the example, the software code may be tangibly stored on one or more volatile, non-transitory, or non-volatile tangible computer-readable media, for example, during execution or at other times. Examples of such tangible computer-readable media may include, but are not limited to, hard disks, removable disks, removable optical discs (e.g., optical discs and digital video discs), magnetic tape cassettes, memory cards or memory sticks, random access memory (RAM), read-only memory (ROM), etc.

[0101] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on this application that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, which will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0102] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments can be used by those skilled in the art when reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the application. This should not be construed as an intention that a disclosed feature not claimed is necessary for any claim. Rather, the subject matter of the application may be less than all the features of a particular disclosed embodiment. Thus, the claims are incorporated herein by reference as examples or embodiments, wherein each claim is an independent, separate embodiment, and these embodiments are contemplated as being able to be combined with each other in various combinations or arrangements. The scope of this application should be determined by reference to the appended claims and the full scope of their equivalents.

[0103] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method of predicting the service life of a solid oxide fuel cell (SOFC), characterized by, The prediction method comprises: acquiring multi-source state data at different time points in the SOFC operation process, and obtaining a multi-domain feature set for characterizing the SOFC health state change based on the multi-source state data; performing first fusion processing on the concerned features in the multi-domain feature set at different time points to obtain a first comprehensive health index sequence; based on the first comprehensive health index sequence, obtaining a preset number of sub-sequences with different center frequencies, and dividing the sub-sequences into a low-frequency trend component and a high-frequency residual component; inputting the low-frequency trend component into a first data model for prediction, and inputting the high-frequency residual component into a second data model for prediction, to obtain a low-frequency prediction sequence and a high-frequency prediction sequence respectively, wherein the low-frequency prediction sequence and the high-frequency prediction sequence are superimposed as a data model prediction sequence; constructing a physical model based on the physical and chemical prior knowledge in the SOFC operation process, adjusting the parameters of the first data model and the second data model by using the physical model, and predicting to obtain a physical model prediction sequence; based on the data model prediction sequence and the physical model prediction sequence, obtaining the remaining useful life prediction result of the SOFC.

2. The prediction method of claim 1, wherein, Based on the data model prediction sequence and the physical model prediction sequence, the remaining useful life prediction result of the SOFC is obtained, specifically comprising: evaluating the prediction credibility of the data model and the physical model to obtain an evaluation result; based on the evaluation result, adjusting the first weight of the data model prediction sequence and the second weight of the physical model prediction sequence, so that the weight corresponding to the model with high prediction credibility is higher; based on the data model prediction sequence and its first weight, the physical model prediction sequence and its second weight, performing second fusion processing to obtain a second comprehensive health index sequence; based on the second comprehensive health index sequence and the preset SOFC life termination threshold, determining the remaining useful life prediction result.

3. The prediction method of claim 2, wherein, The prediction credibility of the data model is evaluated based on the correlation strength of the concerned features and the first comprehensive health index sequence or the deviation fluctuation of the data model prediction sequence and the first comprehensive health index sequence; The prediction credibility of the physical model is evaluated based on the matching degree of the physical model and the SOFC attenuation mechanism.

4. The prediction method of claim 1, wherein, Based on the physical and chemical prior knowledge in the SOFC operation process, a physical model is constructed, and the parameters of the first data model and the second data model are adjusted by using the physical model, specifically comprising: using a double exponential decay function to convert the physical and chemical prior knowledge into quantifiable constraint rules; the output result of the first data model and the second data model needs to satisfy the output value constraint rule of the double exponential decay function, and / or the prediction parameters involved in the prediction process of the first data model and the second data model need to satisfy the parameter constraint in the double exponential decay function.

5. The prediction method of claim 1, wherein, The multi-source state data comprises real-time operation data, electrochemical characteristic data and dynamic response characteristic data of the SOFC. The multi-domain feature set comprises time domain features, integral features, frequency domain features and time-frequency features.

6. The prediction method of claim 1, wherein, The signal decomposition module is configured to obtain a preset number of subsequences with different center frequencies based on the first comprehensive health index sequence, and divide the subsequences into low-frequency trend components and high-frequency residual components. The signal decomposition module is configured to obtain a preset number of subsequences with different center frequencies based on the first comprehensive health index sequence, and divide the subsequences into low-frequency trend components and high-frequency residual components. The signal decomposition module is configured to obtain a preset number of subsequences with different center frequencies based on the first comprehensive health index sequence, and divide the subsequences into low-frequency trend components and high-frequency residual components.

7. The prediction method of claim 1, wherein, The first data model is a time convolution network, and the second data model is a long short-term memory network.

8. A device for predicting the lifespan of a solid oxide fuel cell (SOFC), characterized in that, The prediction device comprises: The data acquisition module is configured to acquire multi-source state data at different time instants in the SOFC operation process. The feature extraction module is configured to obtain a multi-domain feature set for representing changes in the SOFC health state based on the multi-source state data. The feature fusion module is configured to perform first fusion processing on attention features in the multi-domain feature set at different time instants to obtain a first comprehensive health index sequence. The signal decomposition module is configured to obtain a preset number of subsequences with different center frequencies based on the first comprehensive health index sequence, and divide the subsequences into low-frequency trend components and high-frequency residual components. The double-flow prediction module is configured to input the low-frequency trend components into a first data model for prediction, and input the high-frequency residual components into a second data model for prediction, to obtain a low-frequency prediction sequence and a high-frequency prediction sequence, respectively, wherein the low-frequency prediction sequence and the high-frequency prediction sequence are superimposed as a data model prediction sequence. The physical constraint module is configured to construct a physical model based on physical and chemical prior knowledge in the SOFC operation process, adjust parameters of the first data model and the second data model using the physical model, and obtain a physical model prediction sequence. The prediction module is configured to obtain a remaining useful life prediction result of the SOFC based on the data model prediction sequence and the physical model prediction sequence.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the steps of the SOFC useful life prediction method of any one of claims 1-7.

10. A computer program product, characterised in that, The computer readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the steps of the SOFC useful life prediction method of any one of claims 1-7.

Images