Life prediction method for PEM water electrolyzer

Abstract

The present application relates to the technical field of water electrolyzer life prediction, and specifically discloses a PEM water electrolyzer life prediction method, which uses artificial intelligence technology based on deep learning to analyze the operating voltage data of the PEM water electrolyzer, and performs empirical mode decomposition and frequency domain feature extraction to mine the frequency domain characteristics of the inherent modal components of each voltage signal, and then, by performing kernel feature aggregation processing on the frequency domain characteristics of the inherent modes of each voltage signal, the inherent core regularity and time series change pattern of the voltage signal are captured, so as to realize the intelligent prediction of the life of the PEM water electrolyzer. In this way, the accuracy and efficiency of the life prediction of the PEM water electrolyzer can be effectively improved, while reducing the high cost and high time consumption brought by the traditional laboratory accelerated aging test method.

Description

Technical Field

[0001] The present application relates to the technical field of water electrolyzer life prediction, and more specifically, to a PEM water electrolyzer life prediction method. Background Art

[0002] With the adjustment of the global energy structure and the development of renewable energy technology, the status of hydrogen energy in the future energy system is becoming increasingly prominent. As one of the efficient and clean ways to produce hydrogen energy, proton exchange membrane (PEM) water electrolysis technology has received widespread attention. PEM water electrolyzers produce hydrogen and oxygen by electrolyzing water. Its core components, such as proton exchange membranes and catalyst layers, will be affected by factors such as chemical corrosion and physical wear during long-term operation, resulting in performance degradation, which ultimately affects the efficiency and life of the entire water electrolysis system.

[0003] At present, the life assessment of PEM water electrolyzers mainly relies on the manufacturer's empirical data or accelerated aging tests under laboratory conditions. These methods are time-consuming, costly, and difficult to accurately reflect the aging status of equipment under actual working conditions. In addition, traditional life prediction models are often based on statistical principles. Although they can reveal the performance degradation trend of equipment to a certain extent, their prediction accuracy is relatively limited due to the lack of in-depth analysis and understanding of equipment operation data.

[0004] Therefore, an optimized PEM water electrolyzer life prediction method is expected. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a method for predicting the life of a PEM water electrolyzer, which uses artificial intelligence technology based on deep learning to perform data analysis on the operating voltage data of the PEM water electrolyzer, and performs empirical mode decomposition and frequency domain feature extraction on it to mine the frequency domain characteristics of the inherent modal components of each voltage signal, and then, by performing kernel feature aggregation processing on the frequency domain characteristics of the inherent modes of each voltage signal, the inherent core regularity and time series change pattern of the voltage signal are captured, so as to realize intelligent prediction of the life of the PEM water electrolyzer. In this way, the accuracy and efficiency of the life prediction of the PEM water electrolyzer can be effectively improved, while reducing the high cost and high time consumption brought by the traditional laboratory accelerated aging test method.

[0006] Accordingly, according to one aspect of the present application, a method for predicting the life of a PEM water electrolyzer is provided, comprising:

[0007] Acquire operation data of the PEM water electrolyzer, wherein the operation data is a time queue of voltage data;

[0008] Performing frequency domain analysis based on empirical mode decomposition on the operating data to obtain a set of frequency domain semantic encoding feature vectors of voltage intrinsic mode;

[0009] Performing kernel aggregation guided by characteristic energy potential distribution on the set of the voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a kernel representation vector of the operating data time series;

[0010] Based on the operating data time series kernel representation vector, the service life of the PEM water electrolyzer is determined.

[0011] Compared with the prior art, the PEM water electrolyzer life prediction method provided by the present application uses artificial intelligence technology based on deep learning to analyze the operating voltage data of the PEM water electrolyzer, and performs empirical mode decomposition and frequency domain feature extraction on it to dig out the frequency domain characteristics of the inherent modal components of each voltage signal, and then, by performing kernel feature aggregation processing on the frequency domain characteristics of the inherent modes of each voltage signal, the inherent core regularity and time series change pattern of the voltage signal are captured, so as to realize the intelligent prediction of the life of the PEM water electrolyzer. In this way, the accuracy and efficiency of the life prediction of the PEM water electrolyzer can be effectively improved, while reducing the high cost and high time consumption brought by the traditional laboratory accelerated aging test method. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 Flow chart of a method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application.

[0014] Figure 2 Schematic diagram of data flow of a method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application.

[0015] Figure 3 Flow chart of step S120 in the method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application.

[0016] Figure 4 Flow chart of step S130 in the method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] Below, the embodiments of the present application will be described in more detail in conjunction with the accompanying drawings, and the above and other purposes, features and advantages of the present application will become more apparent. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0018] Figure 1 Flow chart of a method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the embodiment of the present application, the life prediction method of the PEM water electrolyzer includes the steps of: S110, acquiring the operating data of the PEM water electrolyzer, wherein the operating data is a time queue of voltage data; S120, performing frequency domain analysis based on empirical mode decomposition on the operating data to obtain a set of voltage intrinsic mode frequency domain semantic coding feature vectors; S130, performing kernel clustering guided by characteristic energy potential distribution on the set of voltage intrinsic mode frequency domain semantic coding feature vectors to obtain an operating data time series kernel representation vector; S140, determining the service life of the PEM water electrolyzer based on the operating data time series kernel representation vector.

[0019] In the above-mentioned PEM water electrolyzer life prediction method, the step S110 obtains the operating data of the PEM water electrolyzer, and the operating data is a time queue of voltage data. It should be understood that voltage is an important indicator for measuring the working state and performance of the PEM water electrolyzer, and directly reflects the progress of the chemical reaction inside the equipment and the working state of components such as electrodes and membranes. Specifically, the proton exchange membrane and the catalyst used in the PEM water electrolyzer are the core components of the PEM water electrolyzer. The proton exchange membrane is responsible for transmitting protons and isolating hydrogen and oxygen, and the catalyst accelerates the electrochemical reaction. Over time, the proton exchange membrane and the catalyst usually undergo some physical and chemical changes, such as thinning of the proton exchange membrane, reduced mechanical strength, poisoning or reduced activity of the catalyst surface, etc., which leads to a decrease in proton transmission efficiency and an increase in the internal resistance of the electrolyzer, so that at the same current density, a higher voltage is required to drive the electrolysis reaction. In other words, as the performance of the PEM water electrolyzer gradually decreases, its voltage data during operation will show a specific trend of change. Therefore, in the technical solution of the present application, the operating voltage data of the PEM water electrolyzer is analyzed to reveal its performance degradation and predict the remaining service life of the PEM water electrolyzer.

[0020] In practical applications, accurately obtaining the operating data of the water electrolyzer (i.e., the voltage data) is of great significance for the predicted life of the PEM water electrolyzer. Specifically, first, it is necessary to install voltage sensors at key locations of the water electrolyzer, such as near the anode, cathode, or proton exchange membrane of the electrolyzer, to monitor and record voltage changes during the electrolysis process in real time. Next, the voltage sensor needs to be connected to the data recording device. In practical applications, data recording devices generally include data acquisition cards, data recorders, etc., which are mainly used to store and transmit voltage data collected by sensors. The data recording device should have a high sampling rate and accuracy to ensure the integrity and accuracy of the data. For example, the sampling frequency of the data acquisition card can be set to 100 times per second to capture a wealth of data points in a shorter time, thereby more accurately showing the trend of voltage changes. In addition, the data recording device must also have good stability and reliability to cope with various interferences and abnormal conditions that may be encountered in industrial environments.

[0021] Data transmission is also an important part of the data acquisition process. In practical applications, the collected voltage data can be transmitted to a central data processing system or cloud server via wired or wireless communication. Common wired communication methods include Ethernet and RS-485, while wireless communication methods include Wi-Fi, Bluetooth, and Zigbee. Choosing the right communication method depends on the specific application scenario and requirements. For example, if the water electrolyzer is located in a closed industrial environment, wired communication may be a more stable choice; in situations where flexible sensor layout is required, wireless communication is more advantageous.

[0022] Data preprocessing is a key step to ensure data quality and the accuracy of subsequent analysis. Specifically, first, the collected voltage data needs to be cleaned to remove noise and outliers. Noise and outliers may be caused by a variety of factors, such as sensor failure, electromagnetic interference, poor line contact, etc. Noise can be removed by filtering algorithms, sliding average methods and other technical means, while statistical methods can be used to detect and eliminate outliers. For example, the standard deviation of the voltage data at each time point can be calculated, and data points that exceed a certain range can be regarded as outliers and eliminated.

[0023] At the same time, data synchronization is also an important aspect of data preprocessing. Since different sensors may have different sampling rates and clock deviations, it is necessary to ensure that the data collected by all sensors are synchronized in time. In order to ensure the time synchronization of data, the data of different sensors can be aligned to the same time base through the method of timestamp alignment to ensure that the data of each sensor can be correctly matched in subsequent analysis, thereby more accurately reflecting the overall working status of the electrolyzer.

[0024] Next, the synchronized data is formatted. Data formatting is to convert the cleaned and synchronized data into a unified time series format. Usually, the data can be organized into a vector form, where each position represents the voltage value at a time point. This format is not only conducive to data storage and transmission, but also greatly facilitates subsequent data analysis and modeling.

[0025] In addition, data storage is a key step to ensure data security and accessibility. In practical applications, the preprocessed data can be stored in a local database to facilitate subsequent data access and analysis. Common local databases include MySQL, PostgreSQL, etc. These databases have good performance and stability and can efficiently manage and query large amounts of data. In addition, by storing data in a local database, data retrieval and backup can be easily performed, and it is also easy to integrate with other systems.

[0026] In addition to local storage, data can also be uploaded to the cloud to use the powerful computing power and storage resources of the cloud platform for large-scale data analysis and processing. Common cloud storage services include Alibaba Cloud Object Storage Service (OSS) and Amazon S3. After uploading data to the cloud, you can use the various tools and services provided by the cloud platform, such as big data processing frameworks and machine learning platforms, to perform efficient data analysis and model training. In addition, cloud storage is highly available and scalable, and can dynamically adjust storage space and computing resources according to actual needs.

[0027] Furthermore, data verification is a necessary step to ensure the authenticity and validity of data. In practical applications, the consistency and rationality of data can be checked by comparing data from different time periods. For example, by drawing a time series graph of voltage data, the changing trend and periodic characteristics of the data can be observed to determine whether the data is reasonable. At the same time, the quality of the data can be further verified by calculating the statistical characteristics of the data, such as the mean, standard deviation, etc. If the data is found to be abnormal or unreasonable, the cause should be found in time and appropriate measures should be taken to correct it.

[0028] At the same time, data backup is also an important aspect of data verification. In practical applications, by backing up data regularly, data loss or damage can be prevented and data security and integrity can be ensured. For example, data can be backed up to a local disk, an external storage device, or a cloud storage service. In practical applications, the backup frequency should be reasonably set according to the importance and update frequency of the data. It is usually recommended to back up once a day or a week. In addition, a multi-copy backup strategy can be adopted to back up data to multiple locations to improve data redundancy and reliability.

[0029] In summary, in the PEM water electrolyzer life prediction method, the process of obtaining operating data specifically includes multiple steps such as data acquisition, data preprocessing, data storage and data verification. First, it is necessary to install voltage sensors at key parts of the water electrolyzer, connect data recording equipment, and transmit the collected voltage data to the central data processing system or cloud server through wired or wireless communication. Next, the collected data is cleaned, synchronized and formatted to ensure the quality and consistency of the data. Then, the preprocessed data is stored in a local database or cloud storage service for subsequent data access and analysis. Finally, by comparing data from different time periods and calculating statistical characteristics, the authenticity and validity of the data are verified, and data backup is performed regularly to ensure the security and integrity of the data. In this way, the operating voltage data of the PEM water electrolyzer can be obtained efficiently and accurately, thereby providing reliable data support for subsequent life prediction.

[0030] In the above-mentioned PEM water electrolyzer life prediction method, the step S120 is to perform frequency domain analysis on the operating data based on empirical mode decomposition to obtain a set of voltage intrinsic mode frequency domain semantic coding feature vectors. Figure 3 FIG. 1 is a flow chart of step S120 in the method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application. Figure 3 As shown, the step S120 includes: S121, performing empirical mode decomposition on the operating data to obtain a set of voltage intrinsic modal functions; S122, inputting each voltage intrinsic modal function in the set of voltage intrinsic modal functions into a voltage intrinsic modal frequency domain feature extractor based on an LSTM-RNN hybrid model to obtain a set of voltage intrinsic modal frequency domain semantic encoding feature vectors.

[0031] Specifically, in step S121, the operating data is subjected to empirical mode decomposition to obtain a set of voltage intrinsic mode functions. It should be understood that, considering that traditional methods such as Fourier transform and wavelet transform assume that the signal is linear and stable, in actual applications, the voltage data of the PEM water electrolyzer is usually a complex nonlinear and non-stationary signal, containing multiple frequency components and noise. Therefore, in order to improve the effectiveness of data analysis, the present application further uses an empirical mode decomposition method to process the operating data to decompose the complex voltage data into several intrinsic mode functions (IMFs) to generate a set of voltage intrinsic mode functions. Among them, each IMF represents an inherent vibration mode in the voltage data. For example, the high-frequency part usually contains the noise component and transient changes of the data, while the low-frequency part reflects the long-term trend of the signal more, thereby simplifying the data structure and helping to better capture the inherent regularity of the voltage signal.

[0032] Specifically, in the process of empirical mode decomposition, first, the upper and lower envelopes of the signal are constructed. Specifically, it is necessary to find the local maximum points and local minimum points of the signal. A local maximum point refers to a point that is higher than the adjacent points in its neighborhood, and a local minimum point refers to a point that is lower than the adjacent points in its neighborhood. By traversing the time queue of the voltage data, all local maximum points and local minimum points are found, and then the upper envelope and lower envelope are fitted by interpolation methods respectively. Commonly used interpolation methods include cubic spline interpolation and linear interpolation. Smooth curves are generated by the above interpolation methods, representing the upper envelope and lower envelope of the signal respectively.

[0033] Next, after constructing the upper and lower envelopes, calculate the average of the upper and lower envelopes to obtain the mean envelope, and subtract the mean envelope from the original signal to obtain a new signal. Then, check whether the generated new signal meets the definition of IMF. Specifically, IMF must meet two conditions: first, the number of extreme points and the number of zero crossings in the entire data sequence do not differ by more than 1; second, the local average of the upper and lower envelopes is zero. If the above generated signal does not meet the definition of IMF, use it as a new input signal and repeat the above steps of constructing the upper and lower envelopes, calculating the mean envelope, and extracting the IMF until the generated signal meets the definition of IMF and the first intrinsic mode function (IMF) is obtained.

[0034] Then, the first IMF is subtracted from the original signal to obtain the residual signal, and the residual signal is used as the new input signal to repeat the above empirical mode decomposition process to continue to extract the next IMF. After each iteration, a new IMF and a new residual signal are obtained. The iterative decomposition calculation continues until the residual signal becomes monotonic or meets a certain stopping criterion. Finally, the original signal is decomposed into multiple intrinsic mode functions (IMFs) and a residual term.

[0035] After completing the EMD decomposition, the obtained intrinsic mode function (IMF) needs to be verified to ensure that it does represent the different frequency components in the signal. For example, by plotting the time series and spectrum of IMFs, the time domain and frequency domain characteristics of each intrinsic mode function (IMF) can be observed to determine whether it correctly reflects the natural vibration mode of the voltage signal.

[0036] Specifically, in step S122, each voltage intrinsic modal function in the set of voltage intrinsic modal functions is input into a voltage intrinsic modal frequency domain feature extractor based on a LSTM-RNN hybrid model to obtain a set of voltage intrinsic modal frequency domain semantic encoding feature vectors. That is, in order to further reveal the frequency domain variation characteristics of each voltage intrinsic modal function, so as to deeply understand the intrinsic structure and dynamic behavior of the voltage signal, the present application adopts a LSTM-RNN hybrid model to construct a voltage intrinsic modal frequency domain feature extractor, and performs feature extraction on each voltage intrinsic modal function to dig out the variation trend of different frequency components of the voltage signal over time. Those skilled in the art should be aware that the LSTM-RNN hybrid model combines the quick response capability of the Recurrent Neural Network (RNN) to short-term dependencies and the long-term memory capability of the Long Short-Term Memory (LSTM), and can effectively achieve long-term tracking and memory of the dynamic behavior of voltage signals, and capture key features such as the frequency distribution changes and phase information of each voltage intrinsic mode function, thereby generating a set of voltage intrinsic mode frequency domain semantic encoding feature vectors.

[0037] In the above-mentioned PEM water electrolyzer life prediction method, the step S130 performs kernel aggregation based on the characteristic energy potential distribution guidance on the set of the voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain the operation data time series kernel representation vector. It should be understood that the set of voltage intrinsic modal frequency domain semantic encoding feature vectors respectively represents the time series change characteristics of different frequency components of the voltage signal. Therefore, in order to further extract the deeper intrinsic essential time series change regularity of the voltage signal to improve the accuracy of the PEM water electrolyzer life prediction, the present application proposes a kernel aggregation method based on the characteristic energy potential distribution guidance, which learns the energy distribution of each voltage intrinsic modal frequency domain semantic encoding feature vector to identify and strengthen the key features in the voltage signal, realize the effective aggregation of kernel features, and suppress the interference of noise and unimportant information to achieve accurate capture of the intrinsic core change mode of the voltage signal.

[0038] Figure 4 FIG. 1 is a flow chart of step S130 in the method for predicting the life of a PEM water electrolyzer according to an embodiment of the present application. Figure 4As shown, the step S130 includes: S131, performing static energy potential analysis on each voltage intrinsic modal frequency domain semantic coding feature vector in the set of voltage intrinsic modal frequency domain semantic coding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic static energy potential factors; S132, performing dynamic energy potential analysis on each voltage intrinsic modal frequency domain semantic coding feature vector in the set of voltage intrinsic modal frequency domain semantic coding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors; S133, performing dynamic energy potential analysis on each voltage intrinsic modal frequency domain semantic coding feature vector in the set of voltage intrinsic modal frequency domain semantic coding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors; S134, performing dynamic energy potential analysis on each voltage intrinsic modal frequency domain semantic coding feature vector in the set of voltage intrinsic modal frequency domain semantic coding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors; S135 A set of domain semantic static energy potential factors and a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors are used to determine the superposition state energy potential factors of each voltage intrinsic modal frequency domain semantic coding feature vector in the set of voltage intrinsic modal frequency domain semantic coding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors; S134, based on the set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors, kernel feature clustering is performed on the set of voltage intrinsic modal frequency domain semantic coding feature vectors to obtain the operating data timing kernel representation vector.

[0039] In a specific example of the present application, the step S131 includes: dividing the fourth-order central moment of the voltage intrinsic modal frequency domain semantic encoding feature vector by the fourth power of its feature variance as its static energy potential factor to obtain the voltage intrinsic modal frequency domain semantic static energy potential factor, which is expressed as:

[0040]

[0041]

[0042] in, represents the set of frequency domain semantic encoding feature vectors of the voltage intrinsic modal, , , and Respectively represent the first, second, and third in the set of the voltage intrinsic modal frequency domain semantic encoding feature vectors and The voltage intrinsic mode frequency domain semantic encoding feature vector, is the number of the frequency domain semantic encoding feature vectors of the voltage natural mode, Indicates the The first frequency domain semantic encoding feature vector of the voltage natural mode eigenvalues, and Respectively represent the The characteristic mean and characteristic variance of the frequency domain semantic encoding feature vector of the voltage natural mode, represents the length of the voltage intrinsic modal frequency domain semantic encoding feature vector, Indicates the Here, for each voltage intrinsic modal frequency domain semantic encoding feature vector, by calculating its static energy potential factor, the information density it carries is revealed, and the inherent properties and information importance of each voltage intrinsic modal function when it is independent of other intrinsic modal functions are evaluated.

[0043] In a specific example of the present application, the step S132 includes: calculating the position-wise mean vector of the set of the voltage intrinsic modal frequency domain semantic coding feature vectors to obtain the voltage intrinsic modal frequency domain semantic initial aggregation representation vector; calculating the dynamic energy potential factor of each voltage intrinsic modal frequency domain semantic coding feature vector in the set of the voltage intrinsic modal frequency domain semantic coding feature vector relative to the voltage intrinsic modal frequency domain semantic initial aggregation representation vector to obtain the set of the voltage intrinsic modal frequency domain semantic dynamic energy potential factors. More specifically, calculating the dynamic energy potential factor of the voltage intrinsic modal frequency domain semantic coding feature vector relative to the voltage intrinsic modal frequency domain semantic initial aggregation representation vector includes: calculating the position-wise weighted sum of the voltage intrinsic modal frequency domain semantic coding feature vector and the voltage intrinsic modal frequency domain semantic initial aggregation representation vector to obtain the voltage intrinsic modal frequency domain semantic dynamic energy potential representation vector; calculating the square of the L2 norm of the point-wise sum vector between the voltage intrinsic modal frequency domain semantic dynamic energy potential representation vector and the bias vector to obtain the voltage intrinsic modal frequency domain semantic dynamic energy potential factor, which is expressed by the formula:

[0044]

[0045]

[0046] in, is the initial aggregation representation vector of the voltage intrinsic mode frequency domain semantics, Indicates the The dynamic energy potential factor of the frequency domain semantic encoding feature vector of the voltage natural mode, represents the bias vector, and represents different weight parameters, represents dot product, It means calculating the square of L2 norm of a vector.

[0047] It should be understood that, considering that each voltage intrinsic modal function has a certain contribution and influence on the overall dynamic behavior of the voltage signal, the present application further calculates the position mean of the set of the voltage intrinsic modal frequency domain semantic encoding feature vectors to construct a typical example reflecting the overall timing characteristics of the voltage signal, captures the overall timing change pattern of the voltage signal, and generates the voltage intrinsic modal frequency domain semantic initial aggregation representation vector. Then, by calculating the dynamic energy potential factor of each voltage intrinsic modal frequency domain semantic encoding feature vector relative to the voltage signal information initial aggregation representation vector to capture the relative relationship between the data, measure its interaction and influence with the overall timing change pattern of the voltage signal, so as to have a deeper understanding of how the frequency domain characteristics of each voltage intrinsic modal function affect the overall timing change pattern of the voltage signal.

[0048] In a specific example of the present application, the step S133, respectively uses the square of the static energy potential factor and the dynamic energy potential factor of the voltage intrinsic modal frequency domain semantic encoding feature vector as the exponent, calculates the exponential function value with e as the base to obtain the exponential static energy potential factor and the exponential dynamic energy potential factor, and adds the exponential static energy potential factor and the exponential dynamic energy potential factor to obtain the voltage intrinsic modal frequency domain semantic superposition state energy potential factor, which is expressed by the formula: ,in, Indicates the The superposition state energy potential factor of the frequency domain semantic encoding feature vector of the voltage intrinsic mode.

[0049] That is, in order to comprehensively consider the static and dynamic properties of each voltage inherent modal frequency domain semantic coding feature vector, the static energy potential factor and dynamic energy potential factor of each voltage inherent modal frequency domain semantic coding feature vector are further integrated to form a superposition state energy potential factor. This comprehensive indicator is used to reflect the composite influence of each voltage inherent modal frequency domain semantic coding feature vector in the entire set, and to more comprehensively evaluate its contribution to the core change pattern of the voltage signal.

[0050] In a specific example of the present application, the step S134 includes: after the set of the voltage intrinsic modal frequency domain semantic superposition state energy potential factors is input into the sigmoid function for normalization processing, the set of normalized voltage intrinsic modal frequency domain semantic superposition state energy potential factors is masked and inactivated based on a preset mask threshold to obtain a set of voltage intrinsic modal frequency domain semantic feature energy potential weight coefficients; based on the set of voltage intrinsic modal frequency domain semantic feature energy potential weight coefficients, the position-weighted sum of the set of voltage intrinsic modal frequency domain semantic encoding feature vectors is calculated to obtain the operating data timing kernel representation vector, which is expressed as follows:

[0051]

[0052]

[0053]

[0054]

[0055] in, Indicates A normalized voltage intrinsic modal frequency domain semantic superposition state energy potential factor, is the mask function, Indicates the preset mask threshold, For the The potential weight coefficient of the voltage intrinsic modal frequency domain semantic feature, A kernel representation vector representing the operating data timing.

[0056] That is, based on the gated mask mechanism, each generated superposition state energy potential factor is normalized and weighted, so as to perform weighted aggregation on the set of voltage intrinsic modal frequency domain semantic encoding feature vectors, thereby explicitly marking the intrinsic modal function information that has a greater impact on the core change pattern of the voltage signal, filtering out the feature representation that contributes less to the overall pattern, thereby obtaining the operating data time series kernel representation vector and achieving an in-depth understanding of the overall operating status of the PEM water electrolyzer.

[0057] In the above-mentioned PEM water electrolyzer life prediction method, the step S140 determines the service life of the PEM water electrolyzer based on the operating data time series kernel representation vector. In a specific example of the present application, the step S140 includes: inputting the operating data time series kernel representation vector into a decoder-based life prediction module to obtain a decoded value of the service life of the PEM water electrolyzer. In the technical solution of the present application, the decoder adopts a multi-layer perceptron structure to utilize the powerful nonlinear mapping ability of the multi-layer perceptron, and learns the deep features in the operating data time series kernel representation vector through layer-by-layer transmission and transformation of multiple hidden layers, and maps it to the prediction result of the remaining service life of the PEM water electrolyzer. In this way, the performance attenuation of the PEM water electrolyzer can be effectively detected and the life can be predicted, thereby providing a scientific basis for the maintenance and replacement of the equipment, and further improving the operating efficiency and reliability of the water electrolysis system.

[0058] When each voltage intrinsic modal frequency domain semantic coding feature vector in the set of voltage intrinsic modal frequency domain semantic coding feature vectors respectively represents the semantic coding features of the frequency domain component signal waveform of each voltage intrinsic modal function determined by empirical mode decomposition of the time queue of the voltage data of the PEM water electrolyzer, when it is input into the kernel feature extraction based on the dynamic analysis of the feature force field, the operating data timing kernel representation vector will also have the instance decision missing of the timing kernel feature due to the dynamic difference of the frequency domain correlation feature force field under the local frequency domain distribution difference. Therefore, when the operating data timing kernel representation vector is decoded and regressed through the life prediction module based on the decoder, it will affect the accuracy of the decoding result.

[0059] In a preferred example of the present application, inputting the operation data time series kernel representation vector into a decoder-based life prediction module to obtain a decoded value of the service life of the PEM water electrolyzer includes:

[0060] Calculate the distance between each pair of eigenvalues ​​of the running data time series kernel representation vector, such as the L2 distance, and take the square root of the distance to obtain the running data time series kernel distance representation matrix, that is,

[0061]

[0062] in, represents the operation data timing kernel representation vector, and The operation data timing kernel represents the first and Eigenvalues, The operation data timing kernel represents the first and The distance between eigenvalues, Represents the running data timing kernel distance representation matrix The eigenvalue of the position;

[0063] Obtain the operation data timing kernel autocorrelation matrix of the operation data timing kernel representation vector as a row vector, that is, ,in, represents the operation data timing kernel representation vector, represents the transpose of a vector, represents matrix multiplication, Represents the autocorrelation matrix of the kernel running data timing;

[0064] The operation data timing kernel representation vector is matrix-multiplied with the operation data timing kernel distance representation matrix to obtain the operation data timing kernel first-level mapping vector, that is, ,in, represents the operation data timing kernel representation vector, Represents the running data timing kernel distance representation matrix, represents matrix multiplication, Represents the first-level mapping vector of the running data timing kernel;

[0065] The operation data timing kernel first-level mapping vector is matrix-multiplied with the matrix product of the operation data timing kernel distance representation matrix and the operation data timing kernel self-correlation matrix to obtain the operation data timing kernel multi-level mapping vector ,in, Represents the first-level mapping vector of the running data timing kernel, Represents the running data timing kernel distance representation matrix, represents matrix multiplication, represents the autocorrelation matrix of the kernel running the data timing, Represents the multi-level mapping vector of the running data timing kernel;

[0066] The operation data timing kernel multi-level mapping vector is interpolated with the operation data timing kernel associated eigenvector composed of the eigenvalues ​​of the operation data timing kernel autocorrelation matrix to obtain an optimized operation data timing kernel representation vector, wherein interpolation or zero padding is performed when the eigenvalues ​​are insufficient;

[0067] The optimized operating data time series kernel representation vector is input into a decoder-based life prediction module to obtain a decoded value of the service life of the PEM water electrolyzer.

[0068] Accordingly, in this preferred embodiment, a linear target mapping representation based on the similarity distance representation matrix of the operating data timing kernel representation vector is used to perform a secondary target mapping representation based on a multi-level distribution hierarchy on the complete similarity instantiation of the self-correlation of the operating data timing kernel representation vector, and the associated fusion kernel bias is used to compensate for the negative impact factor of the association mismatch, so as to improve the degree of instance determination of the eigenvalue of the operating data timing kernel representation vector under similarity constraints, that is, the degree of significance of the eigenvalue as an instance for the decoding regression decision, and improve the accuracy of the decoding value of the service life of the PEM water electrolyzer obtained by the operating data timing kernel representation vector through the decoder-based life prediction module.

[0069] In summary, the life prediction method of the PEM water electrolyzer according to the embodiment of the present application is explained, which uses artificial intelligence technology based on deep learning to analyze the operating voltage data of the PEM water electrolyzer, and performs empirical mode decomposition and frequency domain feature extraction on it to dig out the frequency domain characteristics of the inherent modal components of each voltage signal, and then, by performing kernel feature aggregation processing on the frequency domain characteristics of the inherent modes of each voltage signal, the inherent core regularity and time series change pattern of the voltage signal are captured, so as to realize the intelligent prediction of the life of the PEM water electrolyzer. In this way, the accuracy and efficiency of the life prediction of the PEM water electrolyzer can be effectively improved, while reducing the high cost and high time consumption brought by the traditional laboratory accelerated aging test method.

[0070] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to being implemented by adopting the above specific details.

Claims

1. A method for predicting the life of a PEM water electrolyzer, characterized in that: include: Acquire operation data of the PEM water electrolyzer, wherein the operation data is a time queue of voltage data; Performing frequency domain analysis based on empirical mode decomposition on the operating data to obtain a set of frequency domain semantic encoding feature vectors of voltage intrinsic mode; Performing kernel aggregation guided by characteristic energy potential distribution on the set of the voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a kernel representation vector of the operating data time series; Determining the service life of the PEM water electrolyzer based on the operating data time series kernel representation vector; The set of the voltage intrinsic modal frequency domain semantic encoding feature vectors is subjected to kernel aggregation guided by feature energy potential distribution to obtain a kernel representation vector of the operating data time series, including: Performing static energy potential analysis on each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic static energy potential factors; Performing dynamic energy potential analysis on each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors; Based on the set of voltage intrinsic modal frequency domain semantic static energy potential factors and the set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors, determine the superposition state energy potential factor of each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors; Based on the set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors, kernel feature aggregation is performed on the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain the operation data timing kernel representation vector.

2. The PEM water electrolyzer life prediction method according to claim 1, characterized in that: Performing frequency domain analysis based on empirical mode decomposition on the operating data to obtain a set of voltage intrinsic mode frequency domain semantic coding feature vectors, including: Performing empirical mode decomposition on the operating data to obtain a set of voltage intrinsic mode functions; Each voltage intrinsic modal function in the set of voltage intrinsic modal functions is input into a voltage intrinsic modal frequency domain feature extractor based on an LSTM-RNN hybrid model to obtain a set of voltage intrinsic modal frequency domain semantic encoding feature vectors.

3. The method for predicting the life of a PEM water electrolyzer according to claim 2, characterized in that: Performing static energy potential analysis on each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic static energy potential factors, including: The calculation result of dividing the fourth-order central moment of the voltage intrinsic modal frequency domain semantic coding feature vector by the fourth power of its feature variance is used as its static energy potential factor to obtain the voltage intrinsic modal frequency domain semantic static energy potential factor.

4. The method for predicting the life of a PEM water electrolyzer according to claim 3, characterized in that: Performing dynamic energy potential analysis on each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors, including: Calculating the positional mean vector of the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain a voltage intrinsic modal frequency domain semantic initial aggregation representation vector; The dynamic potential factor of each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of the voltage intrinsic modal frequency domain semantic encoding feature vectors relative to the voltage intrinsic modal frequency domain semantic initial aggregation representation vector is calculated to obtain the set of voltage intrinsic modal frequency domain semantic dynamic potential factors.

5. The method for predicting the life of a PEM water electrolyzer according to claim 4, characterized in that: Calculating the dynamic energy potential factor of each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors relative to the voltage intrinsic modal frequency domain semantic initial aggregation representation vector to obtain a set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors, including: Calculating the position-weighted sum of the voltage intrinsic modal frequency domain semantic encoding feature vector and the voltage intrinsic modal frequency domain semantic initial aggregation representation vector to obtain a voltage intrinsic modal frequency domain semantic dynamic energy potential representation vector; The square of the L2 norm of the point-added vector between the voltage intrinsic modal frequency-domain semantic dynamic energy potential representation vector and the bias vector is calculated to obtain the voltage intrinsic modal frequency-domain semantic dynamic energy potential factor.

6. The method for predicting the life of a PEM water electrolyzer according to claim 5, characterized in that: Based on the set of voltage intrinsic modal frequency domain semantic static energy potential factors and the set of voltage intrinsic modal frequency domain semantic dynamic energy potential factors, determining the superposition state energy potential factors of each voltage intrinsic modal frequency domain semantic encoding feature vector in the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain the set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors, including: The squares of the static energy potential factor and the dynamic energy potential factor of the voltage inherent modal frequency domain semantic encoding feature vector are respectively used as exponents, and the exponential function values ​​with base e are calculated to obtain the exponential static energy potential factor and the exponential dynamic energy potential factor, and the exponential static energy potential factor and the exponential dynamic energy potential factor are added and summed to obtain the voltage inherent modal frequency domain semantic superposition state energy potential factor.

7. The method for predicting the life of a PEM water electrolyzer according to claim 6, characterized in that: Based on the set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors, kernel feature aggregation is performed on the set of voltage intrinsic modal frequency domain semantic encoding feature vectors to obtain the operation data time series kernel representation vector, including: After the set of voltage intrinsic modal frequency domain semantic superposition state energy potential factors is input into the sigmoid function for normalization processing, the set of normalized voltage intrinsic modal frequency domain semantic superposition state energy potential factors is subjected to mask inactivation processing based on a preset mask threshold to obtain a set of voltage intrinsic modal frequency domain semantic feature energy potential weight coefficients; Based on the set of potential weight coefficients of the voltage intrinsic modal frequency domain semantic feature energy, the position-weighted sum of the set of voltage intrinsic modal frequency domain semantic encoding feature vectors is calculated to obtain the operating data timing kernel representation vector.

8. The method for predicting the life of a PEM water electrolyzer according to claim 7, characterized in that: Determining the service life of the PEM water electrolyzer based on the operation data time series kernel representation vector includes: The operation data time series kernel representation vector is input into a decoder-based life prediction module to obtain a decoded value of the service life of the PEM water electrolyzer.

Images