Key parameter monitoring and early warning method in PEM electrolytic water hydrogen production process

Abstract

This invention relates to the field of electrical performance testing technology for hydrogen production systems, and discloses a method for monitoring and early warning of key parameters in the PEM (Polymer Electrolysis) water electrolysis hydrogen production process. The method includes: injecting a composite excitation signal into the electrolyzer, the amplitude of which is related to the real-time operating current density and average temperature; collecting multi-dimensional operating parameters, performing time-frequency domain joint processing on the voltage response signal to obtain a time-frequency spectrum; extracting real-time harmonic distortion feature vectors and impedance changes based on the time-frequency spectrum, calculating the spectral offset from the reference vector, and determining the fault tendency type and region by combining the frequency band energy ratio; splicing the feature vectors and outputting the fault probability distribution through a lightweight artificial intelligence model, and executing graded early warning. This invention improves the timeliness and accuracy of fault identification through multi-signal excitation and multi-dimensional parameter fusion analysis, achieving precise monitoring and graded early warning of the electrolyzer's operating status, and ensuring the stable and reliable hydrogen production process.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electrical performance test of a hydrogen production system, in particular to a key parameter monitoring and early warning method in a PEM water electrolysis hydrogen production process. BACKGROUND

[0002] In the field of hydrogen production, proton exchange membrane water electrolysis hydrogen production is widely used due to its high reaction efficiency, high product purity, flexible start-stop and other advantages. PEM (Proton Exchange Membrane) as a core component directly affects the operation stability and service life of the electrolytic cell. At present, the electrical parameter measurement and fault early warning in the operation process of the electrolytic cell mainly rely on single-dimensional parameter monitoring or traditional offline detection methods, which have problems such as incomplete parameter coverage, unsystematic feature extraction, and lagging fault identification, and it is difficult to accurately capture the state changes of the electrolytic cell under different working conditions, and it is difficult to timely predict potential fault risks such as electrode corrosion and electrolyte abnormalities, which may affect the continuity and safety of the hydrogen production process. Therefore, there is an urgent need for a method that can integrate multi-dimensional operating parameters, accurately monitor and grade early warning through scientific signal processing and feature analysis, to ensure the stable and reliable operation of the PEM water electrolysis hydrogen production system. SUMMARY

[0003] In order to solve or at least partially solve the above technical problems, the application provides a key parameter monitoring and early warning method in a PEM water electrolysis hydrogen production process, comprising the following steps:

[0004] S1, under the constant current steady-state operation condition of the electrolytic cell, a composite excitation signal is generated and injected into both ends of the electrolytic cell, the composite excitation signal includes a basic sweep signal, a high-frequency probe signal, a medium-frequency probe signal and a low-frequency calibration signal, the amplitude of the composite excitation signal is related to the real-time working current density and average temperature of the electrolytic cell;

[0005] S2, multi-dimensional operating parameters of the electrolytic cell are collected, and voltage response signals in the multi-dimensional operating parameters are processed in time and frequency domains to obtain a time-frequency spectrum;

[0006] S3, based on the time-frequency spectrum, real-time harmonic distortion feature vectors and impedance change amounts of the electrolytic cell are extracted, a spectral shift amount of the real-time harmonic distortion feature vectors and a reference harmonic distortion feature vector under a corresponding working condition is calculated, and a fault tendency type and a fault area of the electrolytic cell are determined in combination with different frequency band energy ratios and the impedance change amounts;

[0007] S4, splice the spectrum offset, the impedance change and the characteristic parameter in the multi-dimensional operation parameter into a feature vector, input the light-weight artificial intelligence model after standardization processing to perform inference, output the probability distribution of various faults, and execute hierarchical early warning and corresponding processing measures according to the probability distribution.

[0008] Optionally, in the S1, the basic sweep signal is a sine wave, the load interval is divided according to the real-time working current density of the electrolytic cell to determine the sweep frequency range, when in the first load interval, the sweep frequency range is the first sweep interval, when in the second load interval, the sweep frequency range is expanded from the first sweep interval to the second sweep interval, when in the third load interval, the sweep frequency range is changed to the third sweep interval, the minimum frequency of the third sweep interval is less than or equal to the maximum frequency of the first sweep interval, the maximum frequency of the third sweep interval is greater than the maximum frequency of the second sweep interval, the first load interval, the second load interval and the third load interval are adjacent and increasing in turn;

[0009] The high-frequency probe signal is a triangular wave, the medium-frequency probe signal is a square wave, and the low-frequency calibration signal is a sine wave.

[0010] The amplitude of the composite excitation signal first increases with the increase of the real-time working current density to set an amplitude base value, and then the amplitude base value is compensated and adjusted according to the deviation of the average temperature from the reference temperature, the greater the temperature deviation, the greater the compensation amplitude.

[0011] Optionally, in the S1, the sweep rate of the basic sweep signal is a fixed value, the frequency of the high-frequency probe signal is higher than the maximum sweep frequency of the basic sweep signal, the frequency of the medium-frequency probe signal is within the sweep frequency range of the basic sweep signal, and the frequency of the low-frequency calibration signal is lower than the minimum sweep frequency of the basic sweep signal.

[0012] The amplitude of the high-frequency probe signal is greater than the amplitude base value and has a fixed proportional relationship with the amplitude base value, the amplitude of the medium-frequency probe signal is greater than the amplitude base value and has another fixed proportional relationship with the amplitude base value, and the amplitude of the low-frequency calibration signal is less than the amplitude base value and is a preset proportion of the amplitude base value.

[0013] The duty cycle of the high-frequency probe signal and the medium-frequency probe signal is a fixed value, and the duty cycle of the medium-frequency probe signal is less than the duty cycle of the high-frequency probe signal.

[0014] Optionally, in the S2, the process of performing time-frequency domain joint processing on the voltage response signal includes:

[0015] According to a frequency range of each signal in the composite excitation signal, a targeted filtering algorithm is used to separate response components corresponding to the basic sweep signal, the high-frequency probe signal, the medium-frequency probe signal and the low-frequency calibration signal one by one;

[0016] Time-frequency domain transformation is performed on each response component respectively, and time-frequency features of each response component are extracted, including instantaneous frequency, instantaneous amplitude and phase information;

[0017] According to an action weight of each response component in fault monitoring, time-frequency features of each response component are weighted and fused to generate the time-frequency spectrum, and the action weight is related to a signal type corresponding to each response component.

[0018] Optionally, in the S3, the process of extracting the real-time harmonic distortion feature vector and calculating the impedance change amount includes:

[0019] Based on the separated response components, an amplitude distortion rate and a phase shift value of a 3rd or higher harmonic are extracted for the high-frequency probe signal response component, an amplitude ratio and a frequency deviation of a fundamental wave and a 2nd harmonic are extracted for the medium-frequency probe signal response component, the amplitude distortion rate, the phase shift value, the amplitude ratio and the frequency deviation are combined into a harmonic feature subset;

[0020] With a phase reference of the low-frequency calibration signal response component as a reference, each parameter in the harmonic feature subset is subjected to temperature drift error correction, and equivalent impedance parameters of the electrolytic cell are calculated based on full-frequency band impedance spectrum data of the basic sweep signal response component, and then the impedance change amount relative to a reference working condition is obtained;

[0021] The impedance change amount is derived to obtain three types of derived parameters, including a real part change rate, an imaginary part change rate and an impedance angle change value, the harmonic feature subset subjected to temperature drift error correction and the three types of derived parameters are spliced in a preset order to form the real-time harmonic distortion feature vector.

[0022] Optionally, in the S3, the process of calculating the spectral offset amount of the real-time harmonic distortion feature vector and a reference harmonic distortion feature vector under a corresponding working condition includes:

[0023] Based on a current real-time working current density of the electrolytic cell, a corresponding reference harmonic distortion feature vector is matched from a preset multi-working-condition reference parameter library;

[0024] The real-time harmonic distortion feature vector and the reference harmonic distortion feature vector are aligned dimension by dimension, and a relative deviation value of each corresponding dimension parameter is calculated;

[0025] According to the contribution degree of each dimension parameter to fault identification, a weight coefficient is preset, the relative deviation values of the dimension parameters are multiplied by the corresponding weight coefficients, and then summed to obtain a comprehensive frequency spectrum offset; the Euclidean distance between the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector is calculated as an auxiliary offset indicator, and the comprehensive frequency spectrum offset and the auxiliary offset indicator jointly constitute the frequency spectrum offset.

[0026] Optionally, in the S3, the process of determining the fault tendency type and the fault region of the electrolytic cell in combination with the different frequency band energy ratios and the impedance change amount comprises:

[0027] Based on the time-frequency spectrum, the energy values of each frequency band are extracted, the energy ratios of any two frequency bands are calculated, and at least two groups of different frequency band energy ratios are obtained;

[0028] A preset fault determination rule library is called, and the fault determination rule library pre-stores the correlation characteristics of different fault tendency types, corresponding fault regions, and frequency spectrum offset-different frequency band energy ratio-impedance change amount;

[0029] The different frequency band energy ratios, the frequency spectrum offset, and the impedance change amount are substituted into the fault determination rule library, the real part change rate and / or the imaginary part change rate of the impedance change amount are used to preliminarily lock the fault category, the different frequency band energy ratios and the auxiliary offset indicators of the frequency spectrum offset are combined to match the specific fault tendency type, and finally the corresponding fault region is located according to the abnormal energy distribution position of each frequency band.

[0030] Optionally, in the S3, after determining the fault tendency type and the fault region of the electrolytic cell, the steps of fault warning verification and monitoring parameter adjustment are further included, specifically comprising:

[0031] Based on the determination result, a preset verification excitation strategy is matched, the parameters of the composite excitation signal are adjusted according to the fault tendency type, if it is a resistance type fault, the amplitude ratio of the low frequency calibration signal is increased, if it is a reactance type fault, the sweep frequency range of the basic sweep frequency signal is expanded, and the adjustment amplitude and the frequency spectrum offset are positively correlated;

[0032] The adjusted composite excitation signal is re-applied to the electrolytic cell, and the steps of S2 to S3 are repeated to obtain a secondary determination result. The fault tendency type, the fault region, and the frequency spectrum offset of the two determinations are compared, if the consistency is higher than a consistency threshold, a fault warning is triggered, and if the consistency is lower than the consistency threshold, it is marked as a suspected interference.

[0033] Optionally, in the S2, the time-frequency domain joint processing of the voltage response signal is performed, and the process of local current density acquisition and signal verification is further included, specifically comprising:

[0034] Collecting local current density data corresponding to the voltage response signal through a plurality of current collection nodes arranged on the surface of the anode and the cathode of the electrolytic cell;

[0035] Calculating the mean and standard deviation of the current density of each node based on the collected local current density data to obtain the local current density distribution characteristics;

[0036] Verifying the relevance of the local current density distribution characteristics and the separated basic sweep signal response component, if the local current density standard deviation exceeds the preset standard deviation range, the amplitude of the voltage response signal in the corresponding region is corrected, and if the local current density standard deviation is within the preset standard deviation range, the voltage response signal is confirmed to be effective.

[0037] Optionally, the constructing process of the harmonic feature subset also includes the steps of extracting and integrating even harmonic parameters, specifically including:

[0038] For the high-frequency probe signal response component, on the basis of extracting 3rd and higher harmonic parameters, the amplitude distortion rate and phase shift value of 2nd and 4th even harmonics are additionally extracted;

[0039] For the medium-frequency probe signal response component, on the basis of extracting the amplitude ratio of the fundamental wave and the 2nd harmonic, the amplitude ratio of the 4th even harmonic and the fundamental wave, and the frequency deviation of the 4th even harmonic are additionally extracted;

[0040] The newly added amplitude distortion rate, phase shift value, amplitude ratio and frequency deviation of the 2nd and 4th even harmonics are merged with the parameters of the original harmonic feature subset to form an extended harmonic feature subset containing even harmonic information;

[0041] The extended harmonic feature subset is subjected to temperature drift error correction, and then spliced with the three types of derived parameters of the impedance change amount to form a real-time harmonic distortion feature vector containing even harmonic features.

[0042] The method provided by the application has the following beneficial effects:

[0043] The present application injects a composite excitation signal containing a basic sweep signal, a high-frequency probe signal, a medium-frequency probe signal and a low-frequency calibration signal when the electrolytic cell is in a constant-current steady-state operation, and the amplitude of the composite excitation signal is related to the real-time working current density and the average temperature of the electrolytic cell, so that the injected composite excitation signal can adapt to different operating states of the electrolytic cell, improving the fit of the signal to the actual working condition and providing a more targeted basis for subsequent parameter collection and analysis.

[0044] The time-frequency domain joint processing of the collected voltage response signal and the generation of the time-frequency spectrum diagram can comprehensively extract the instantaneous frequency, amplitude and phase and other characteristic information in the signal, and fully reflect the running state of the electrolytic cell, avoiding the possible characteristic omission caused by a single signal processing method. Based on the time-frequency spectrum diagram, the related feature vectors and impedance change are extracted, the fault is judged by combining the frequency spectrum offset and the energy ratio of different frequency bands, the abnormal change of the electrolytic cell can be captured from multiple dimensions, the fault-related judgment is more in line with the actual situation, and the effective identification of the fault tendency type and the fault area is realized.

[0045] After the related characteristic parameters are spliced and standardized, the light artificial intelligence model is input, the occurrence of various faults can be reasonably judged and the probability distribution is output, the hierarchical early warning and the corresponding processing measures executed accordingly can timely feedback the running risk of the electrolytic cell, help the staff to take timely response actions, reduce the adverse effects of the fault on the electrolytic water hydrogen production process, and ensure the stable development of the hydrogen production process. BRIEF DESCRIPTION OF DRAWINGS

[0046] Fig. 1 A key parameter monitoring and early warning method in a PEM electrolytic water hydrogen production process is provided for the embodiments of the present application.

[0047] Fig. 2 A time-frequency spectrum diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0049] Referring to Figs. 1-2 The embodiments of the present application provide a key parameter monitoring and early warning method in a PEM electrolytic water hydrogen production process, comprising the following steps:

[0050] S1, under the constant current steady state running condition of the electrolytic cell, a composite excitation signal is generated and injected into both ends of the electrolytic cell, the composite excitation signal includes a basic sweep signal, a high-frequency probe signal, a medium-frequency probe signal and a low-frequency calibration signal, the amplitude of the composite excitation signal is related to the real-time working current density and the average temperature of the electrolytic cell.

[0051] S2, the multi-dimensional running parameters of the electrolytic cell are collected, and the voltage response signal in the multi-dimensional running parameters is processed in time and frequency domain to obtain a time-frequency spectrum diagram.

[0052] S3, based on the time-frequency spectrogram, extracting the real-time harmonic distortion feature vector and impedance change of the electrolytic cell, calculating the frequency spectrum offset of the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector under the corresponding working condition, combining the energy ratio of different frequency bands and the impedance change, determining the fault tendency type and fault area of the electrolytic cell.

[0053] S4, the frequency spectrum offset, impedance change and characteristic parameters in multi-dimensional operating parameters are spliced into a feature vector, which is input into a lightweight artificial intelligence model after standardization processing for inference, and the probability distribution of each type of fault is output, and the corresponding processing measures are executed according to the probability distribution.

[0054] In the process of hydrogen production by electrolytic water, the electrolytic cell as the core equipment, its running state is related to the efficiency and safety of the hydrogen production process, through accurate capture of operating data, timely discovery of potential problems and provision of response basis. Constant current steady state operating condition refers to the electrolytic cell working with stable current to avoid the interference of current fluctuation on the measurement results, and to ensure that the collected parameters can truly reflect the normal operating characteristics of the equipment. The composite excitation signal includes four types of basic sweep signal, high frequency probe signal, medium frequency probe signal and low frequency calibration signal, and different signals undertake different measurement functions.

[0055] In actual implementation, first, let the electrolytic cell enter the constant current steady state operating condition, at this time the equipment running state is stable, there is no obvious fluctuation. Then generate a composite excitation signal, adjust the amplitude of the composite excitation signal according to the current density and average temperature of the electrolytic cell at present, and when it is confirmed that the composite excitation signal matches the equipment running state, inject the composite excitation signal into both ends of the electrolytic cell. Next, collect multi-dimensional operating parameters, and focus on extracting voltage response signal, perform time-frequency domain joint processing on the voltage response signal, and generate time-frequency spectrogram which can clearly present the signal characteristics by integrating the analysis results of time domain and frequency domain. Based on the generated time-frequency spectrogram, further extract the real-time harmonic distortion feature vector and impedance change, the real-time harmonic distortion feature vector can reflect the distortion degree of the harmonic component in the voltage signal, and the impedance change reflects the change of the equivalent impedance of the electrolytic cell.

[0056] Then the reference harmonic distortion feature vector corresponding to the working condition needs to be obtained, which is a reference standard determined in advance by multiple tests under the normal running state of the electrolytic cell. By calculating the frequency spectrum offset between the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector, the difference between the real-time running state and the normal state can be quantified. At the same time, the energy information of different frequency bands is extracted from the time-frequency spectrum, and the energy ratio of different frequency bands is calculated. Combined with the impedance change obtained before, the correlation between these parameters is comprehensively analyzed, and then it is determined whether the electrolytic cell has a failure tendency and the specific area where the failure may occur. Finally, the frequency spectrum offset, the impedance change and the selected characteristic parameters in the multi-dimensional running parameters are integrated and spliced to form a feature vector containing comprehensive information of the equipment running. The feature vector is standardized to eliminate the dimensional differences between different parameters.

[0057] The standardized feature vector is input into a lightweight artificial intelligence model and subjected to inference analysis, and finally the probability distribution of various faults is output. According to the results of the probability distribution, hierarchical early warning and corresponding treatment measures are performed, and corresponding operations are taken for different probability levels of fault risk to timely investigate potential problems and avoid the expansion of faults.

[0058] In some embodiments, in S1, the basic sweep signal is a sine wave, the load interval is divided according to the real-time working current density of the electrolytic cell to determine the sweep frequency range, when in the first load interval, the sweep frequency range is the first sweep frequency interval, when in the second load interval, the sweep frequency range is expanded from the first sweep frequency interval to the second sweep frequency interval, when in the third load interval, the sweep frequency range is changed to the third sweep frequency interval, the minimum frequency of the third sweep frequency interval is less than or equal to the maximum frequency of the first sweep frequency interval, the maximum frequency of the third sweep frequency interval is greater than the maximum frequency of the second sweep frequency interval, the first load interval, the second load interval and the third load interval are adjacent and increasing in turn;

[0059] In the preferred embodiment, the upper limit of the first load interval can be set as the current density at which the electrolytic cell begins to show obvious concentration polarization; the upper limit of the second load interval can be set as the critical current density at which the electrode surface begins to precipitate by-products. These turning points can be determined by standardized polarization curve testing. The high-frequency probe signal is a triangular wave, the medium-frequency probe signal is a square wave, and the low-frequency calibration signal is a sine wave; the amplitude of the composite excitation signal first increases with the increase of the real-time working current density to set the amplitude basic value, and then the amplitude basic value is compensated and adjusted according to the deviation of the average temperature from the reference temperature. The greater the temperature deviation, the greater the compensation amplitude.

[0060] Specifically, the basic sweep signal as the core detection signal, its sweep range needs to be adapted to the real-time working current density of the electrolytic cell, so as to fully capture the impedance change characteristics under different loads. In actual implementation, first, the rated working current density range of the electrolytic cell is obtained, and three continuous and increasing load intervals are divided according to the numerical distribution law of the current density. The division basis is combined with the running characteristics of the electrolytic cell. The current density is close to the lower limit of the rated value under low load, the middle interval of the rated value under medium load, and close to or reaches the upper limit of the rated value under high load. The three intervals do not overlap and cover the entire operating range, avoiding the connection of the sweep range due to the load breakpoint.

[0061] After determining the load interval, the corresponding sweep range is matched for each interval: the first load interval corresponds to the low-load running state, and the impedance change of the electrolytic cell is concentrated in the low-frequency band. Therefore, the first sweep interval is set to cover this low-frequency band to ensure that the impedance basic characteristics under low load can be accurately captured. When the real-time working current density enters the second load interval (medium load), the ion migration rate inside the electrolytic cell increases, and the impedance characteristics extend to the medium-high frequency. At this time, the sweep range is expanded from the first sweep interval to the second sweep interval, which not only retains the monitoring of the low-frequency band, but also adds the coverage of the medium-high frequency band to avoid missing the key electrical characteristic changes under medium load. After entering the third load interval (high load), the electrolytic cell may appear local overheating, uneven distribution of electrolyte concentration, and other situations, and the impedance characteristics show wide frequency domain fluctuations. Therefore, the minimum frequency of the third sweep interval does not exceed the maximum frequency of the first sweep interval, ensuring that the low-frequency basic information is not lost, and the maximum frequency exceeds that of the second sweep interval, realizing the coverage of higher frequency bands and adapting to the complex impedance changes under high load. For example, when the working current density is less than 0.5 A / cm 2 , the corresponding sweep range can be 10 Hz~5kHz, when the working current density is between 0.5~1.5 A / cm 2 , the corresponding sweep range can be 10 Hz~10kHz, and when the working current density is greater than 1.5 A / cm 2 , the corresponding sweep range can be 5kHz~20kHz. If the real-time working current density is at the boundary of two load intervals, a buffer range of ±5% needs to be set. When the current density fluctuates within the buffer range, the current sweep range is maintained to avoid signal instability caused by frequent switching. Only when the current density continuously exceeds the buffer range, the sweep range corresponding to the load interval is switched. The purpose of setting the buffer range of ±5% is to prevent frequent switching caused by small current fluctuations at the boundary of the load interval, ensuring the stability of the system operation.

[0062] The amplitude setting of the composite excitation signal needs to be performed in two steps: the first step is to determine the amplitude base value, and the minimum effective response amplitude of the electrolytic cell under different real-time working current densities is obtained through experiments. The lower the current density, the lower the signal response sensitivity of the electrolytic cell, and the amplitude base value needs to be set to the minimum value that can be effectively captured by the sensor. The higher the current density, the greater the internal electric field strength of the electrolytic cell, and the more obvious the signal response, and the amplitude base value is appropriately increased with the increase of the current density, but does not exceed the safety voltage threshold of the electrolytic cell, so as to avoid damaging the proton exchange membrane due to excessive amplitude. Finally, a corresponding relationship curve of the current density and the amplitude base value is formed, and the amplitude base value is set in real time according to the curve. The second step is to adjust the temperature compensation. First, the reference temperature of the electrolytic cell is determined, which is the best operating temperature calibrated when the equipment is shipped. The real-time average temperature is collected through the temperature sensor of the electrolytic cell, and the deviation of the real-time average temperature from the reference temperature is calculated. If the real-time average temperature is higher than the reference temperature, the internal resistance of the electrolytic cell will decrease, and the signal response amplitude may be too high, so the amplitude base value needs to be appropriately reduced. If the real-time average temperature is lower than the reference temperature, the internal resistance increases, and the signal response amplitude may be too low, so the amplitude base value needs to be increased. The greater the temperature deviation, the greater the compensation amplitude, so as to ensure that the actual effect of the composite excitation signal is consistent under different temperature conditions, and to avoid measurement errors caused by temperature changes.

[0063] During implementation, the changes of the real-time working current density and the average temperature need to be monitored in real time, for example, the amplitude base value and the compensation amplitude are updated every 100 milliseconds, to ensure that the composite excitation signal is always adapted to the operating state of the electrolytic cell; at the same time, the parameters of each load interval switching, sweep range adjustment and amplitude compensation are recorded to form an operation log, which is convenient for subsequent fault tracing and parameter optimization. This scheme can accurately detect the electrical characteristics of the electrolytic cell under different operating conditions, and provides high-quality raw data for subsequent parameter measurement and fault warning.

[0064] In some embodiments, in S1, the sweep rate of the base sweep signal is a fixed value, the frequency of the high-frequency probe signal is higher than the maximum sweep frequency of the base sweep signal, the frequency of the medium-frequency probe signal is within the sweep range of the base sweep signal, and the frequency of the low-frequency calibration signal is lower than the minimum sweep frequency of the base sweep signal.

[0065] The amplitude of the high-frequency probe signal is greater than the amplitude base value and has a fixed proportional relationship with the amplitude base value, the amplitude of the medium-frequency probe signal is greater than the amplitude base value and has another fixed proportional relationship with the amplitude base value, and the amplitude of the low-frequency calibration signal is less than the amplitude base value and is a preset proportion of the amplitude base value.

[0066] The duty cycle of the high-frequency probe signal and the medium-frequency probe signal is a fixed value, and the duty cycle of the medium-frequency probe signal is less than that of the high-frequency probe signal.

[0067] Specifically, the sweep rate of the base sweep signal is set with a fixed value, because the fluctuation of the sweep rate will cause distortion of the time-frequency characteristics of the response signal, increasing the difficulty of subsequent separation and analysis. In practice, the fixed value needs to be determined through pre-test: taking the typical running state of the electrolytic cell under the rated current density as the benchmark, test the response signal quality under different sweep rates respectively, when the rate is too low, the single sweep takes too long time, and cannot capture the dynamic changes of the electrolytic cell in time; when the rate is too high, the signal stays at each frequency point for insufficient time, and the sensor cannot collect stable response data. Finally, the rate that can ensure stable collection of response signal at each frequency point and meet the real-time monitoring requirement is selected as the fixed value, and after the value is determined, it remains unchanged in all load intervals, even if the sweep range of the base sweep signal is adjusted with the switching of the load interval, the sweep rate remains constant, ensuring the comparability of the measurement data under different working conditions.

[0068] The frequencies of each signal need to avoid frequency overlap interference between signals. Before implementation, first determine the sweep range of the base sweep signal corresponding to the current load interval, and then determine the frequencies of the other three types of signals based on this: the frequency of the high-frequency probe signal needs to be higher than the maximum frequency value of the sweep range, the specific value needs to be determined in combination with the micro characteristics of the electrolytic cell electrode, usually selected to be able to penetrate the oxide film on the surface of the electrode but not to cause abnormal polarization of the electrode, to ensure that it can detect the subtle changes of the internal structure of the electrode; the frequency of the medium-frequency probe signal needs to fall within the sweep range of the base sweep signal, and the middle frequency band of the sweep range is preferred, the impedance change of this frequency band can best reflect the core running state of the electrolytic cell, and can form a complementary detection with the base sweep signal; the frequency of the low-frequency calibration signal needs to be lower than the minimum frequency value of the base sweep signal, and the low-frequency signal is used to provide a reference for the phase measurement of other signals due to its stable phase and strong anti-interference ability. When the sweep range of the base sweep signal is switched with the load interval, the frequencies of the high-frequency, medium-frequency and low-frequency probe signals need to be adjusted synchronously, and the above frequency hierarchy relationship is always maintained, for example, when the base sweep range is expanded upwards, the frequency of the high-frequency probe signal is also increased accordingly, to ensure that it is always above the base sweep range.

[0069] The proportional relationship between the amplitude of each signal and the amplitude base value can be determined through multiple sets of comparison tests, and once determined, it remains fixed and is only adjusted proportionally with the change in the amplitude base value. The proportion of the amplitude of the high-frequency probe signal to the amplitude base value needs to be slightly higher than that of the medium-frequency probe signal, because the high-frequency signal decays faster during transmission, and a higher initial amplitude is needed to ensure that it still has identifiable strength when it reaches the sensor; although the amplitude proportion of the medium-frequency probe signal is lower than that of the high-frequency signal, it still needs to be higher than the amplitude base value to ensure that the signal-to-noise ratio of its response signal meets the needs of harmonic feature extraction; the amplitude of the low-frequency calibration signal is set to a preset proportion of the amplitude base value, and the proportion is less than 1, because the main role of the low-frequency signal is calibration rather than active detection, and a lower amplitude can both avoid interference with other signals and ensure the stability of the phase reference.

[0070] The duty cycles of the high-frequency probe signal and the medium-frequency probe signal are both set to fixed values, and the duty cycle of the medium-frequency probe signal is less than that of the high-frequency probe signal. This setting is to balance the detection priority of the two types of signals and the conflict when the signals are superimposed. The determination of the duty cycle needs to be combined with the action period of the signal: the high-frequency probe signal needs to be sent more frequently to capture rapidly changing microscopic features, so a higher duty cycle is set to ensure the number of detections per unit time; the medium-frequency probe signal is mainly used to extract stable harmonic features, and does not need a high sending frequency. A lower duty cycle can both meet the needs and reserve transmission time for the high-frequency probe signal and the low-frequency calibration signal, avoiding waveform distortion caused by the superposition of multiple signals in the time domain. In the implementation process, the sending time of the two types of signals needs to be strictly controlled by the signal timing control module to ensure the stability of the duty cycle. Even when the amplitude of the composite excitation signal is adjusted, the duty cycle remains unchanged to ensure the regularity of signal transmission.

[0071] Referring to Fig. 2 In some embodiments, the process of performing time-frequency domain joint processing on the voltage response signal in S2 includes:

[0072] Based on the frequency range of each signal in the composite excitation signal, a targeted filtering algorithm is used to separate the response components corresponding to the base sweep signal, the high-frequency probe signal, the medium-frequency probe signal, and the low-frequency calibration signal one by one;

[0073] Time-frequency domain transformation is performed on each response component to extract the time-frequency features of each response component, including instantaneous frequency, instantaneous amplitude, and phase information;

[0074] The time-frequency features of each response component are weighted and fused according to the role weight of each response component in fault monitoring to generate a time-frequency spectrum, and the role weight is related to the signal type corresponding to each response component.

[0075] In implementation, first, the frequency range boundary of each type of signal in the composite excitation signal needs to be determined, which needs to be determined based on the signal parameters set in advance. The frequency intervals of the basic sweep signal, the high-frequency probe signal, the medium-frequency probe signal, and the low-frequency calibration signal do not overlap with each other, and each signal has a clear upper and lower limit of the frequency range. Select a targeted filtering algorithm according to the frequency characteristics of each signal, for example, for the high-frequency characteristics of the high-frequency probe signal, a filtering method that can effectively retain high-frequency components and suppress low-frequency interference is used; for the low-frequency characteristics of the low-frequency calibration signal, a filtering method that can filter high-frequency noise is used; for the medium-frequency probe signal in the intermediate frequency band and the basic sweep signal covering a certain frequency band, the filtering parameters are adjusted according to the frequency distribution range, so that the filtering algorithm can match the frequency range of the corresponding signal. In the filtering process, the collected voltage response signal needs to be input into the filtering module, and the cutoff frequency, bandwidth and other parameters of the filtering algorithm are adjusted step by step, so that the output signal component only contains the response information of the corresponding excitation signal. At the same time, the separation effect is verified by spectrum analysis. If it is found that a response component still contains the frequency components of other signals, the filtering parameters need to be adjusted again until the frequency spectrum of each response component has no overlapping, and complete separation is achieved.

[0076] After separating each response component, time-frequency domain transformation processing is performed on each component. Time domain analysis mainly captures the law of signal change with time, records the amplitude change of the response component at different times, obtains the instantaneous amplitude information, and tracks the phase shift of the signal with time to obtain the instantaneous phase data; frequency domain analysis focuses on the frequency distribution characteristics of the signal, converts the time domain signal into a frequency domain spectrum through transformation, extracts the instantaneous frequency information, and determines the frequency composition of the fundamental wave and the harmonic contained in each response component. The reason for adopting joint time-frequency domain processing is that single time domain analysis cannot clearly present the frequency distribution law of the signal, and single frequency domain analysis cannot reflect the dynamic change of the signal with time. The combination of the two can fully capture the complete characteristics of the signal, for example, the response component of the basic sweep signal can be observed through time domain analysis. The amplitude fluctuation during the sweep process can be determined through frequency domain analysis. The response intensity of different frequency points can be combined to completely restore the impedance change characteristics during the sweep process; the response component of the high-frequency probe signal can be accurately captured through joint time-frequency domain processing, and the instantaneous frequency shift and amplitude distortion caused by the micro changes of the electrode.

[0077] After the time-frequency features of each response component are extracted, they are weighted and fused according to their respective roles in fault monitoring. The response component of the basic sweep frequency signal is directly used to extract the impedance change of the electrolytic cell, which is the core basis for fault determination, and therefore it is assigned the highest weight. The response component of the high-frequency probe signal is used to detect changes in the microstructure of the electrode, and the response component of the medium-frequency probe signal is used to analyze harmonic distortion characteristics. Both are important aids in fault type determination and are assigned medium weights. The response component of the low-frequency calibration signal is mainly used to correct errors caused by temperature drift and ensure the accuracy of other characteristic parameters, and is assigned a relatively low weight. During the fusion process, the instantaneous frequency, instantaneous amplitude, and phase information of each response component are first quantized and adjusted according to their corresponding weights. Then, the adjusted feature information is integrated into the same analysis framework. Through a time-frequency domain image generation algorithm, the multi-dimensional features are transformed into an intuitive time-frequency spectrum. The graph shows the feature distribution of each response component at different times and frequencies, retaining the key information of each individual component while also reflecting the correlation between the components. Fig. 2 As shown in the figure, the horizontal axis t represents time in seconds, and the vertical axis f represents frequency in Hz. The region from 10 to 100 Hz represents the low-frequency region, corresponding to the horizontal strip of the low-frequency calibration signal; the region from 500 to 2000 Hz represents the mid-frequency region, corresponding to the horizontal strip of the mid-frequency probe signal; the region from 5000 to 10000 Hz represents the high-frequency region, corresponding to the horizontal strip of the high-frequency probe signal; the diagonally extending strip represents the basic sweep frequency signal, which gradually covers the area from the upper limit of the low-frequency region, around 500 Hz, upwards to around 5000 Hz; the color scales on the right side of the figure correspond to the signal amplitude, with darker colors indicating larger signal amplitudes at that time and frequency.

[0078] During implementation, a real-time verification mechanism must be established: after separation, the absence of interference in each response component is confirmed by spectrum comparison; after time-frequency domain transformation, the integrity of characteristic parameters is checked to ensure no key information is omitted; after weighted fusion, the fusion effect is verified by comparing the time-spectrum diagram with that of the standard signal to confirm whether the fusion effect meets expectations. If any abnormality occurs in any step, the corresponding parameters need to be adjusted retrospectively. For example, if the filter parameters are unreasonable, the cutoff frequency should be re-optimized; if the weight allocation is improper, it should be recalibrated through experiments until the time-spectrum diagram can comprehensively and accurately reflect all the key characteristics of the voltage response signal.

[0079] In some implementations, the process of extracting the real-time harmonic distortion feature vector and calculating the impedance change in S3 includes:

[0080] Based on the separated response components, the amplitude distortion rate and phase shift value of the third and higher harmonics are extracted for the high-frequency probe signal response components, and the amplitude ratio and frequency deviation of the fundamental and second harmonics are extracted for the mid-frequency probe signal response components. The amplitude distortion rate, phase shift value, amplitude ratio and frequency deviation are combined into a harmonic feature subset.

[0081] The temperature drift error of each parameter in the harmonic feature subset is corrected with the phase reference of the low-frequency calibration signal response component as the reference, and the equivalent impedance parameters of the electrolytic cell are calculated based on the full-frequency impedance spectrum data of the basic sweep signal response component, and then the impedance change amount relative to the reference working condition is obtained;

[0082] The impedance change amount is derived to obtain three types of derived parameters: impedance real part change rate, imaginary part change rate, and impedance angle change value. The harmonic feature subset corrected by the temperature drift error and the three types of derived parameters are spliced in a predetermined order to form a real-time harmonic distortion feature vector.

[0083] In implementation, first focus on the high-frequency probe signal response component after separation, which carries key information of electrode microstructure changes, and the distortion degree of 3rd and higher harmonics is related to electrode polarization abnormalities, etc. With the help of spectrum analysis tools, the high-frequency response component is analyzed in detail. First, the fundamental frequency is locked by spectrum peak positioning, and then the frequency interval corresponding to the 3rd, 5th, etc. higher harmonics is identified based on the fundamental frequency, and the amplitude distortion rate of each higher harmonic is calculated, i.e. the ratio of the amplitude of the higher harmonic to the amplitude of the fundamental wave. The larger the ratio, the more obvious the signal distortion. At the same time, the phase information of each higher harmonic is captured, and the phase shift value is obtained by comparing with the phase of the fundamental wave. The positive and negative and size of the phase shift value can reflect the phase disorder degree of the harmonic component. These amplitude distortion rates and phase shift values are classified and recorded as important components of the harmonic feature.

[0084] For the intermediate-frequency probe signal response component, the correlation characteristics of its fundamental wave and 2nd harmonic are important basis for judging the uniformity of electrolyte distribution. First, the frequency and amplitude corresponding to the peak value of the fundamental wave of the intermediate-frequency probe signal are located through the frequency spectrum after time-domain to frequency-domain conversion, and then the peak position of the 2nd harmonic is identified, and the amplitude ratio of the two is calculated. This ratio is in a stable range when the electrolytic cell is running normally, and any fluctuation may indicate an abnormal electrolyte concentration. At the same time, the frequency deviation is obtained by comparing the actual frequency of the 2nd harmonic with the theoretical frequency, i.e. twice the fundamental frequency. The value of the frequency deviation can reflect the frequency drift in the signal transmission process. These two parameters are supplemented to the previously recorded feature data to form a harmonic feature subset.

[0085] Since the electrical characteristics of the electrolytic cell are susceptible to temperature, the parameters in the harmonic feature subset may have temperature drift errors, which need to be corrected based on the low-frequency calibration signal response component. The phase stability of the low-frequency calibration signal is significantly better than that of the medium-frequency probe signal and the high-frequency probe signal, and the phase change is basically not affected by load fluctuations and only has a weak correlation with temperature, which can be used as a reliable phase reference standard. When correcting, the real-time phase of the low-frequency calibration signal response component is extracted first, and the phase drift compensation value is obtained by comparing it with the preset standard phase. Then the compensation value is applied to each phase parameter in the harmonic feature subset, and at the same time, the amplitude parameters are adjusted synchronously according to the correlation between temperature and amplitude drift, to ensure that the corrected feature parameters eliminate the interference of temperature factors.

[0086] The calculation of the impedance change amount needs to take the basic sweep signal response component as the core data source, which covers the electrical signal characteristics of the full frequency band and can fully reflect the equivalent impedance characteristics of the electrolytic cell. Through the impedance analysis model, the equivalent impedance parameters corresponding to different frequency points are calculated based on the instantaneous amplitude and phase information of the basic sweep signal response component. The equivalent impedance parameters include real and imaginary parts, the real part reflects the resistance characteristics of the electrolytic cell, and the imaginary part reflects the reactance characteristics. Then the equivalent impedance parameters of the electrolytic cell under the standard working condition, i.e. normal operating state, are retrieved through multiple tests, and the real-time equivalent impedance parameters are compared with the reference parameters at each frequency point to obtain the impedance changes at each frequency point. By integrating the change data of all frequency points, the overall impedance change amount is finally determined, which can directly reflect the difference between the equivalent impedance of the electrolytic cell and the normal state.

[0087] To further enrich the basis for fault judgment, the impedance change amount needs to be calculated. For the real part of the impedance change amount, the real part change rate is calculated, which is the difference between the real-time impedance real part and the reference impedance real part, and the ratio of the difference to the reference impedance real part, which can quantify the change amplitude of the resistance characteristics. Similarly, the imaginary part change rate is calculated to reflect the change of the reactance characteristics. At the same time, the impedance angle change value is obtained by the difference between the real-time impedance angle and the reference impedance angle, and the change of the impedance angle can be directly related to the charge distribution state of the electrode surface. These three types of derived parameters strengthen the feature expression of impedance change from different dimensions, and form a complement to the corrected harmonic feature subset.

[0088] Finally, the parameters are spliced in the preset order, which needs to be determined in combination with the priority of fault identification, and the harmonic feature parameters with higher sensitivity to faults are placed in the front sequence, and then the three types of impedance derived parameters are arranged in sequence, to ensure that the spliced real-time harmonic distortion feature vector is stable in structure and complete in information. After splicing, the integrity of the vector is checked by a data verification tool. If there is a missing or abnormal parameter, it needs to be reprocessed from the corresponding extraction link until the required real-time harmonic distortion feature vector is generated.

[0089] In some embodiments, in S3, the process of calculating the spectral deviation of the real-time harmonic distortion feature vector from the reference harmonic distortion feature vector under the corresponding working condition comprises:

[0090] Based on the current real-time working current density of the electrolytic cell, a corresponding reference harmonic distortion feature vector is matched from the pre-set multi-working-condition reference parameter library.

[0091] The real-time harmonic distortion feature vector and the reference harmonic distortion feature vector are aligned dimension by dimension, and the relative deviation values of each corresponding dimension parameter are calculated.

[0092] According to the contribution degree of each dimension parameter to fault recognition, a weight coefficient is pre-set, the relative deviation values of each dimension parameter are multiplied by the corresponding weight coefficient, and then summed to obtain a comprehensive spectral deviation; the Euclidean distance between the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector is calculated as an auxiliary deviation index, and the comprehensive spectral deviation and the auxiliary deviation index jointly constitute the spectral deviation.

[0093] The first step of implementation is to construct a multi-working-condition reference parameter library and complete real-time matching. The core of this step is to adapt the reference data to the current running condition of the electrolytic cell. When constructing the multi-working-condition reference parameter library, the real-time working current density of the electrolytic cell is taken as the core basis for division, and the rated current density range is evenly subdivided into multiple subintervals, each subinterval corresponding to a specific working condition. The interval division density meets the requirement that the difference between reference parameters under adjacent working conditions is less than the measurement error threshold, ensuring that there is no working condition blind area. For each working condition, the electrolytic cell needs to be operated stably at this current density for at least 30 minutes, during which the harmonic distortion feature vector is collected every 1 second. After 10 groups of complete data are continuously collected, the outliers are removed, and then the average value is taken according to the dimension to obtain the reference harmonic distortion feature vector under this working condition. After the reference vectors of all working conditions are collected, they are arranged in ascending order of current density and stored in the multi-working-condition reference parameter library. The upper and lower limits of the current density of each working condition and the collection time, environmental temperature, and other auxiliary information of the reference vector are recorded in the library, which facilitates subsequent tracing and calibration.

[0094] In real-time matching, the current real-time working current density of the electrolytic cell is first obtained through the current sensor, and is compared with the current density interval of each working condition in the multi-working condition reference parameter library. If the real-time current density falls within the interval of a working condition, the corresponding reference harmonic distortion characteristic vector of the working condition is directly called. If the real-time current density is at the boundary of two adjacent working conditions, the reference vector is calculated by using the linear interpolation method. Based on the reference vectors of the two adjacent working conditions, the distance between the real-time current density and the two interval boundaries is weighted, the closer the distance, the greater the weight, and the reference vector suitable for the current current is obtained by weighted calculation, avoiding the sudden change of the reference data under the boundary working condition. If the real-time current density exceeds the rated range, the preset extreme working condition reference vector in the multi-working condition reference parameter library is called. The vector is obtained in advance by simulating the extreme working condition test, ensuring that there is still a reference under special operating conditions.

[0095] After completing the reference matching, the real-time harmonic distortion characteristic vector and the reference harmonic distortion characteristic vector need to be aligned and the deviation is calculated. First, the dimension correspondence of the two vectors is determined. The dimension number of the real-time harmonic distortion characteristic vector and the reference harmonic distortion characteristic vector is consistent, and the parameter definition of each dimension corresponds, for example, the 1st dimension is the 3rd harmonic amplitude distortion rate of the high-frequency probe signal, the 2nd dimension is the 3rd harmonic phase shift value of the high-frequency probe signal, and so on. If the dimension number does not match, the calculation is immediately suspended, and the parameter splicing order is checked in the real-time vector generation link, or the integrity of the working condition vector in the multi-working condition reference parameter library is checked to exclude errors in the data transmission or storage process. After alignment, the relative deviation value is calculated dimension by dimension. The calculation logic is: (real-time dimension parameter value - reference dimension parameter value) ÷ reference dimension parameter value. This calculation method can convert the parameter deviation of different dimensions into a dimensionless relative value, which is convenient for subsequent cross-dimension comparison and weighting. For some reference parameters that may have zero values (such as harmonic amplitudes in the extremely low frequency band), a small replacement value needs to be set in advance to avoid the denominator being zero in the calculation. The size of the replacement value should be much smaller than the parameter value under normal working conditions, so as to ensure that the influence on the deviation result can be ignored.

[0096] The preset of the weight coefficient takes the contribution degree of each dimensional parameter to fault recognition as the core basis, and the contribution degree is determined through a large number of fault simulation tests. During the test, common faults of the electrolytic cell are artificially simulated, and the change amplitude of each dimensional parameter when each fault occurs is recorded. The greater the change amplitude, the higher the sensitivity of the dimension to the type of fault, and the greater the contribution degree. The sensitivity data of all fault types are statistically analyzed to calculate the average sensitivity coefficient of each dimensional parameter, and then standardized to a weight value between 0 and 1. The higher the sensitivity of the dimension, the greater the weight coefficient. For example, the 3rd harmonic amplitude distortion rate of the high-frequency probe signal has the largest change amplitude in the electrode scaling fault, and its weight coefficient is higher than that of other dimensions. The phase parameter of the low-frequency calibration signal after correction has relatively low sensitivity to faults, and the weight coefficient is relatively small. The weight distribution of each parameter under typical faults is shown in Table 1.

[0097] Table 1

[0098]

[0099] The calculation of the comprehensive spectrum offset is the cumulative result of the relative deviation value of each dimension multiplied by the corresponding weight coefficient. After the calculation is completed, it needs to be compared with the preset calculation threshold. If the result exceeds the threshold, it means that the difference between the real-time parameter and the reference has reached a level that needs to be concerned. If it is lower than the threshold, it needs to be checked whether there are multiple low-deviation dimensions superimposed to avoid ignoring the overall abnormality due to the small deviation of a single dimension. The auxiliary offset index is calculated by the Euclidean distance. This index can reflect the overall distance between two vectors in a multi-dimensional space, making up for the deficiency of the comprehensive spectrum offset which focuses on sensitive dimensions. When calculating, the two vectors are regarded as points in a multi-dimensional space. The straight-line distance between the two points is calculated by the formula. The greater the distance, the more significant the overall difference. The calculation result of the Euclidean distance does not need to be weighted. It is directly used as an auxiliary index together with the comprehensive spectrum offset to form the final spectrum offset. The presentation form of the two is the comprehensive spectrum offset, which is convenient for the subsequent fault judgment link to refer to both local sensitive differences and overall comprehensive differences.

[0100] A real-time data verification mechanism can be embedded in the entire calculation process. After each spectrum offset calculation is completed, the calculation results of this time and the previous time are compared. If the difference exceeds the allowed fluctuation range, the real-time parameter and the reference parameter are retrieved again for secondary calculation to exclude the abnormality caused by the instantaneous error of the sensor. This multi-step and strong verification calculation method can ensure that the spectrum offset accurately reflects the running deviation of the electrolytic cell and provides reliable quantitative basis for the subsequent fault tendency type judgment.

[0101] In some scenarios, the amplitude of the composite excitation signal may deviate due to slight fluctuations of the power module. Such deviation may indirectly cause the amplitude of the real-time harmonic distortion feature vector to drift, so additional amplitude fluctuation compensation can be considered to improve the accuracy of the spectrum offset.

[0102] Specifically, when constructing the multi-working condition reference parameter library, the corresponding composite excitation signal reference amplitude is associated with each reference harmonic distortion feature vector, and the amplitude is the actual signal amplitude when collecting the reference data. During real-time calculation, the real-time amplitude of the current composite excitation signal is recorded synchronously, which is compared with the reference amplitude of the corresponding working condition to obtain an amplitude correction factor, and the correction factor is the reference amplitude ÷ real-time amplitude.

[0103] Before calculating the relative deviation value dimension by dimension, each dimension parameter of the real-time harmonic distortion feature vector is multiplied by the correction factor to complete the amplitude drift correction, and then the relative deviation, the comprehensive frequency spectrum offset and the Euclidean distance are calculated according to the original logic. If the difference between the real-time amplitude and the reference amplitude exceeds ±3%, the excitation amplitude fluctuation needs to be marked in the calculation report, which is convenient for subsequent tracing of the deviation source. This compensation mechanism can eliminate the influence of signal source fluctuation, and make the frequency spectrum offset more truly reflect the difference in electrical characteristics of the electrolytic cell itself.

[0104] In some embodiments, in S3, the process of determining the fault tendency type and the fault area of the electrolytic cell in combination with the energy ratio of different frequency bands and the impedance change amount includes:

[0105] Based on the time-frequency spectrum, the energy values of each frequency band are extracted, the energy ratio of any two frequency bands is calculated, and at least two groups of different frequency band energy ratios are obtained;

[0106] The preset fault determination rule library is called, and the different fault tendency types and corresponding fault areas and the associated characteristics of the frequency spectrum offset-different frequency band energy ratio-impedance change amount are pre-stored in the fault determination rule library;

[0107] The different frequency band energy ratios, the frequency spectrum offset and the impedance change amount are substituted into the fault determination rule library, the real part change rate and / or the imaginary part change rate of the impedance change amount are used to preliminarily lock the fault category, the different frequency band energy ratios and the auxiliary offset indicators of the frequency spectrum offset are combined to match the specific fault tendency type, and finally the corresponding fault area is located according to the abnormal energy distribution position of each frequency band.

[0108] The first step of implementation is to extract and calculate the energy ratio of different frequency bands based on the time-frequency spectrogram. The energy distribution of different frequency intervals in the time-frequency spectrogram corresponds to the electrical signal response characteristics of different components of the electrolytic cell. The energy of the high frequency band mainly reflects the microstate of the electrode surface, the energy of the medium frequency band is related to the ion migration of the electrolyte, and the energy of the low frequency band is related to the overall structural stability of the electrolytic cell. First, the time-frequency spectrogram is divided into four independent intervals of high frequency band, medium frequency band, low frequency band and basic sweep frequency band according to the frequency range of the composite excitation signal. The division boundary needs to be consistent with the frequency threshold separated in the early filtering. Then, the total energy value of all frequency points in each frequency band is calculated by the energy integration algorithm. The calculation of energy value needs to exclude the noise interference area in the time-frequency spectrogram, and only the frequency points with signal amplitude higher than the noise threshold are integrated to ensure the authenticity of the energy value.

[0109] To fully reflect the abnormal energy distribution, at least two groups of different frequency band energy ratios need to be calculated, such as the ratio of high frequency band energy value to medium frequency band energy value and the ratio of low frequency band energy value to basic sweep frequency band energy value. The former can reflect the correlation of electrode microstate and electrolyte migration characteristics, and the latter can reflect the matching degree of overall structural stability and full-band impedance characteristics. During the calculation process, the real-time value of the energy ratio is recorded, and it is compared with the standard energy ratio under the baseline working condition. If the difference exceeds the energy ratio warning threshold, it is marked as an energy abnormality feature, which provides a key reference for subsequent judgment.

[0110] The fault judgment rule library is the core basis for realizing standardized judgment, and its construction needs to rely on a large number of fault simulation tests and actual operation data accumulation. In the test stage, typical faults of PEM electrolytic water hydrogen production system are simulated artificially, including resistance type faults (such as electrode scaling, poor contact), reactance type faults (such as proton exchange membrane damage, electrode polarization anomaly) and comprehensive type faults (such as multi-component coordinated failure). For each fault type, record the frequency spectrum shift range (including comprehensive shift and auxiliary shift indicators), real / imaginary part change rate proportion of impedance change, and abnormal characteristics of different frequency band energy ratios, and mark the specific area corresponding to the fault (such as anode area, cathode area, membrane component area). For example, when simulating an anode scaling fault, the real part change rate proportion of impedance will significantly increase, the energy ratio of high frequency band to medium frequency band will be more than 2 times the standard value, and the comprehensive frequency spectrum shift will exceed the threshold. These associated information will be stored as a judgment rule in the library. The rule library needs to be stored by fault category (resistance type, reactance type, comprehensive type). Each rule contains trigger conditions and judgment results. The trigger condition is the joint range of three parameters, and the judgment result is the explicit fault tendency type and fault area.

[0111] When entering the judgment link, the logic of gradually advancing by first locking the major category, then matching the type, and finally positioning the area needs to be followed. The first step is to lock the major category of fault by the real part change rate and / or the imaginary part change rate of the impedance change amount: calculate the ratio of the real part change rate to the imaginary part change rate in the impedance change amount, if the real part change rate accounts for more than 70%, it is preliminarily locked as a resistance type fault, the core feature of this type of fault is the abnormal resistance characteristic of the electrolytic cell; if the imaginary part change rate accounts for more than 70%, it is preliminarily locked as a reactance type fault, reflecting the problem of the reactance characteristic of the electrolytic cell; if both accounts for between 40% and 60%, it is determined as a comprehensive type fault, prompting multiple types of electrical characteristics abnormality. This step of judgment can quickly narrow down the fault range and avoid subsequent analysis falling into disorder.

[0112] The second step is to match the specific fault tendency type by combining the energy ratio of different frequency bands with the auxiliary offset amount index of the frequency spectrum offset amount. Retrieve all rules corresponding to the major category of fault in the rule library, and compare the real-time acquired energy abnormality features and the auxiliary offset amount index with the trigger conditions in the rules one by one. If it has been locked as a resistance type fault, and the real-time high frequency band and medium frequency band energy ratio is greater than the standard value, and the Euclidean distance of the auxiliary offset amount index is within the trigger range of the electrode scaling rule, it is matched as the fault tendency type of electrode scaling; if the energy ratio shows that the low frequency band and the basic sweep frequency band ratio is abnormal, and the auxiliary offset amount index meets the contact failure rule, it is matched as the corresponding type. If multiple rules partially match during the matching process, the similarity of the real-time parameters and the trigger conditions of each rule needs to be calculated, and the rule with the highest similarity is selected as the judgment result, and it is marked as suspected multiple fault superposition, prompting further verification in the future.

[0113] The third step is to locate the fault area according to the energy abnormality distribution position of each frequency band. There is a corresponding relationship between the time axis of the time-frequency spectrum and the space monitoring nodes of the electrolytic cell, through the signal acquisition sub-nodes arranged in the key areas of the anode, cathode, membrane assembly, etc. of the electrolytic cell, the signals of different time segments in the time-frequency spectrum can be associated with specific areas. If the signal segment of the high frequency band energy abnormality corresponds to the acquisition sub-node of the anode area, it is determined that the fault area is the anode; if the medium frequency band energy abnormality corresponds to the acquisition sub-node of the membrane assembly area, it is located as a membrane assembly fault. During the positioning process, cross verification needs to be carried out in combination with the common occurrence area of the fault type, such as electrode scaling fault occurring in the anode area, proton exchange membrane damage fault occurring in the membrane assembly area, through the double confirmation of the energy abnormality position and the associated area of the fault type, the accuracy of the fault area positioning is ensured.

[0114] After the determination is completed, a determination report containing the fault category, the fault tendency type, the fault area, and the core abnormal parameter is generated, wherein the core abnormal parameter marks the key indicators triggering the determination rule. At the same time, the determination result is compared with historical similar fault data. For example, if it is found that the matching degree of the determination result with the historical data is lower than 80%, a secondary determination process is started to re-call the parameters for analysis to avoid misjudgment caused by special working conditions interference. This multi-parameter coordinated and step-by-step determination method can fully exert the complementary advantages of each parameter, realize accurate identification of the fault tendency type and the area, and provide clear direction for subsequent early warning and processing.

[0115] In some embodiments, in S3, after the fault tendency type and the fault area of the electrolytic cell are determined, the steps of fault early warning verification and monitoring parameter adjustment are further included, specifically including:

[0116] Based on the determination result matching the preset verification incentive strategy, the parameters of the composite excitation signal are adjusted according to the fault tendency type. If it is a resistance type fault, the amplitude ratio of the low frequency calibration signal is increased. If it is a reactance type fault, the sweep frequency range of the basic sweep signal is expanded, and the adjustment amplitude and the frequency spectrum offset are positively correlated.

[0117] The adjusted composite excitation signal is re-applied to the electrolytic cell, and S2 to S3 are repeated to obtain a secondary determination result. The fault tendency type, the fault area, and the frequency spectrum offset of the two determinations are compared. If the consistency is higher than the consistency threshold, the fault early warning is triggered. If the consistency is lower than the consistency threshold, it is marked as a suspected interference.

[0118] When the fault tendency type and the fault area of the electrolytic cell are determined through the previous process, the preset verification excitation strategy can be matched based on the determination result first. The core logic of the verification excitation strategy is to strengthen the detection signal for the fault characteristics. Different fault types correspond to different electrical characteristic abnormalities of the electrolytic cell. The parameters of the composite excitation signal need to be adjusted to amplify the signal response corresponding to the abnormality, so that the secondary determination is more targeted. In implementation, the fault category locked in the previous stage is first determined: if it is determined to be a resistance type fault, the core feature of this type of fault is abnormal change of the equivalent resistance of the electrolytic cell, and the low-frequency calibration signal has the highest sensitivity to resistance change. Therefore, the amplitude ratio of the low-frequency calibration signal in the composite excitation signal needs to be increased. The specific operation is to reduce the amplitude ratio of the basic sweep signal, the high-frequency and medium-frequency probe signals under the premise that the total amplitude of the composite excitation signal does not exceed the safety threshold, and correspondingly increase the amplitude ratio of the low-frequency calibration signal, so as to ensure that the low-frequency signal can more clearly capture the resistance change characteristics; if it is determined to be a reactance type fault, the core is equivalent reactance abnormality. The wide frequency coverage feature of the basic sweep signal can comprehensively reflect the change law of the reactance with frequency, so the sweep frequency range of the basic sweep signal needs to be extended appropriately to the high and low frequency on the basis of the original range to cover the frequency interval where the reactance abnormality may occur. The adjustment amplitude needs to be linked with the frequency spectrum offset. The greater the frequency spectrum offset, the more significant the difference between the real-time state and the reference state, and the adjustment amplitude of the signal parameter also increases accordingly, so as to ensure that the optimized signal can effectively excite the fault characteristics, while avoiding excessive adjustment leading to operation fluctuation of the electrolytic cell.

[0119] After completing the parameter adjustment of the composite excitation signal, the adjusted signal needs to be re-applied to the electrolytic cell according to the same time sequence logic as the initial measurement. Before application, the output parameters of the signal generation module need to be checked to confirm that the sweep frequency range, amplitude ratio, frequency and other parameters are consistent with the adjustment requirements, and at the same time, it is ensured that the signal injection interface and timing are the same as the initial measurement, so as to avoid introducing new errors due to changes in application conditions. Then the whole process of signal acquisition and processing is repeated: multi-dimensional operating parameters are collected and each response component is separated, time-frequency domain joint processing is performed on the voltage response signal to generate a time-frequency spectrum, real-time harmonic distortion feature vectors and impedance change quantities are extracted based on the time-frequency spectrum, the frequency spectrum offset is calculated and combined with the frequency band energy ratio to determine the fault tendency type and area, and finally the secondary determination result is obtained. In the secondary determination process, all data acquisition sensors and data processing algorithm parameters are consistent with the initial ones, only the composite excitation signal parameters are different, so that the difference between the two determination results is only caused by the real state of the fault or interference factors, and the comparability is ensured.

[0120] After the secondary determination result is generated, a comparative analysis needs to be performed from three core dimensions: first, the consistency of the fault tendency type, checking the specific fault types of the two determinations, such as anode fouling and anode fouling being consistent, and anode fouling and contact failure being inconsistent; second, the consistency of the fault region, confirming whether the fault regions of the two determinations completely coincide or are in the same core region, such as both pointing to the anode region being consistent, and pointing to the anode and cathode respectively being inconsistent; third, the consistency of the spectral shift, calculating the difference of the comprehensive shift in the two spectral shifts, and if the difference is within the allowable fluctuation range, it is considered consistent. After the comparison is completed, a preset consistency threshold needs to be judged, which is determined through historical data statistics, and is usually set to at least two of the three dimensions being completely consistent, and the difference in the third dimension being within an acceptable range. If the consistency of the two determinations is higher than the threshold, it means that the fault features are stable and exist, and the possibility of transient interference is ruled out, at which time a fault warning is triggered, and the warning information needs to clearly include the fault tendency type, fault region and spectral shift data consistent in the two determinations; if the consistency is lower than the threshold, it means that the initial determination may be affected by transient interference, such as power grid voltage fluctuation, temporary false alarm of the sensor, and the determination needs to be marked as suspected interference, without triggering a warning, but the operating parameters at the time of interference, such as real-time working current density, temperature, and power grid voltage, need to be recorded to provide data for subsequent interference identification model optimization.

[0121] For the suspected interference situation, the monitoring parameters need to be further adjusted: appropriately shorten the parameter update period of the composite excitation signal, for example, from the regular 100 milliseconds to 50 milliseconds, to improve the response speed to signal changes; at the same time, increase the collection frequency of the high-frequency probe signal, and use its sensitivity to micro changes to capture possible interference signal characteristics. If the same type of determination result does not appear again within the next 10 minutes, the suspected interference mark is removed; if suspected interference occurs multiple times, a sensor self-checking process needs to be triggered to check whether there is a sensor fault.

[0122] In some embodiments, the time-frequency domain joint processing of the voltage response signal in S2 also includes local current density collection and signal verification processes, specifically including:

[0123] Collecting local current density data corresponding to the voltage response signal through multiple current collection nodes arranged on the anode and cathode surfaces of the electrolytic cell;

[0124] Calculating the mean and standard deviation of the current density of each node based on the collected local current density data to obtain the local current density distribution characteristics;

[0125] The local current density distribution characteristics are associated with the separated basic sweep signal response components, and if the local current density standard deviation exceeds the preset standard deviation range, the amplitude of the voltage response signal in the corresponding region is corrected, and if the local current density standard deviation is within the preset standard deviation range, the voltage response signal is confirmed to be effective.

[0126] In a specific application process, the preset standard deviation range is determined by a benchmark working condition test, and the specific method is as follows: after the electrolytic cell is delivered or overhauled, at least 24 hours of continuous full load and stable operation test is carried out. During this period, the data of all current collection nodes are continuously collected at high frequency. The standard deviation of the current density of each node during the whole test period is calculated, and the distribution of the standard deviation of all nodes is calculated. The preset standard deviation range is set as the maximum value of the standard deviation of all nodes, or the maximum value multiplied by a safety factor. This range represents the limit of the uniformity of the current distribution of the electrolytic cell in the best health state. For a standard 50kW PEM electrolytic cell, the local current density standard deviation of the anode surface is usually less than 5mA / cm 2 Therefore, the preset standard deviation range can be set as [0, 5.5]mA / cm 2 , wherein 5.5 is the upper limit value considering the measurement error.

[0127] Implementing this scheme first requires arranging current collection nodes on the anode and cathode surfaces of the electrolytic cell. As the core area of electrochemical reaction, the uniformity of current distribution is directly related to the reaction efficiency and electrode loss, so the nodes need to uniformly cover the effective working area of the electrode. The effective working area refers to the area of the electrode that participates in proton exchange and electron transfer, and the invalid area at the edge of the electrode is excluded. The arrangement interval is determined in combination with the electrode size, for example, a 5cm basic interval is used, and if there are structural protrusions or areas prone to fouling on the electrode surface, the interval needs to be reduced to 2cm to ensure more intensive current data collection in key areas.

[0128] The collection of local current density data is synchronized with the voltage response signal, because the corresponding relationship between the two in the time dimension is the basis for association verification. When the composite excitation signal is injected into the electrolytic cell, the data collection module generates a synchronous trigger signal, which simultaneously controls the voltage sensor and the current collection node to start collection. The collection frequency is set to match the frequency of the composite excitation signal, ensuring that each voltage response signal data point can correspond to a local current density data point, avoiding association deviation caused by collection timing misalignment. During the collection process, the working state of each current collection node needs to be monitored in real time, and if a node has signal interruption or data jump, it is immediately marked as an abnormal node, and a preset backup node data completion is enabled to ensure the continuity of the collected data.

[0129] When preprocessing the collected local current density data, the first task is to eliminate abnormal fluctuation data. Bubble adhesion, electrolyte flow impact and other situations that may occur during the operation of the electrolytic cell can cause the local current density to jump instantaneously. Such data is not a reflection of the true electrical characteristics of the electrode and needs to be processed through a smoothing filter algorithm. A 5-data-point moving average window is used to filter real-time data. The average value of the data in the window is calculated in chronological order, with recent data having slightly higher weight and distant data having slightly lower weight. This both preserves the dynamic change characteristics of the data and filters transient interference. After filtering, the 3σ criterion is used to identify outliers: the mean and standard deviation of the filtered data are calculated, and data that exceeds the mean ± 3 times the standard deviation range is determined to be abnormal fluctuation data and is eliminated. The positions of the eliminated data are completed by linear interpolation of the adjacent normal data to ensure the integrity of the data sequence.

[0130] After preprocessing, the mean and standard deviation of the current density of each node are calculated to obtain the local current density distribution characteristics. The mean reflects the overall current level of the electrode surface and matches the real-time working current density of the electrolytic cell, which can be used as a preliminary reference for data effectiveness. The standard deviation reflects the dispersion of the current density of each node. The smaller the dispersion, the more uniform the current distribution on the electrode surface, and vice versa, indicating abnormal local current concentration or sparsity. When calculating, the anode and cathode should be counted separately because the electrochemical reaction characteristics of the anode and cathode are different, and the normal range of current distribution differs. Separate statistics can improve the accuracy of subsequent verification.

[0131] The core logic of the correlation verification is based on the derived relationship of Ohm's law. The local current density of a certain area of the electrode should be positively correlated with the amplitude of the voltage response signal corresponding to that area. If the current distribution is uniform and the voltage response amplitude fluctuates greatly, it indicates that the voltage signal may have measurement errors or interference. In practice, the electrode surface is divided into multiple independent regions according to the distribution of the current collection nodes. Each region corresponds to a group of current collection nodes and the basic sweep signal response component of that region. The basic sweep signal response component reflects the equivalent impedance characteristics of the region and, together with the local current density, forms a correlation chain of current-impedance-voltage. Subsequently, the amplitude of the separated basic sweep signal response component of each region is compared with the local current density distribution characteristics of the corresponding region. If the standard deviation of the local current density of a certain region is within the preset range, it indicates that the current distribution in that region is uniform, and the amplitude fluctuation of the corresponding basic sweep signal response component should be within the normal range. In this case, the voltage response signal of the region is confirmed to be valid. If the standard deviation of the local current density exceeds the preset range, it indicates that there is an abnormal current distribution in the region, and the amplitude of the corresponding voltage response signal may be distorted due to current concentration or sparsity, which needs to be corrected.

[0132] The preset standard deviation range can be determined by a benchmark operation test: in the benchmark state of the normal operation of the electrolytic cell, the local current density data is continuously collected for 24 hours, the maximum and minimum values of the standard deviation are calculated for different regions of the anode and cathode respectively, and the interval is taken as the preset standard deviation range of the region. The ranges of different regions may be different due to the differences in electrode structure, for example, the preset range of the edge region of the cathode is slightly wider than that of the center region, because the edge is easily affected by the electrolyte flow and appears slight current fluctuation. When correcting, the correction amplitude is determined according to the difference between the local current density standard deviation and the preset range. The greater the difference, the more serious the abnormal current distribution, and the higher the distortion degree of the voltage response signal. The correction amplitude is also increased accordingly, and the voltage response signal amplitude of the region is adjusted through the preset proportional coefficient to ensure that the corrected voltage signal can truly reflect the matching relationship between the impedance characteristics and the current distribution of the region.

[0133] After verification and correction, the effective or corrected voltage response signals of each region are integrated to form a complete voltage response signal data set, and then the time-frequency domain transformation and feature extraction steps are continued. The local current density distribution characteristics, verification results and correction parameters of each region are recorded during the whole process. If the standard deviation of the current density of a certain region frequently exceeds the standard, mark this region as a key monitoring region and increase the analysis weight of this region in the subsequent fault judgment. This signal verification mechanism based on local current density can verify the authenticity of the voltage response signal from the perspective of electrode surface current distribution, effectively eliminate the distortion of the voltage signal caused by local current abnormality, provide a high-quality data basis for subsequent time-frequency domain analysis, and further improve the reliability of the whole monitoring system.

[0134] In addition, the relative humidity of the electrolytic cell operating environment can change the local current conduction characteristics by affecting the thickness of the adsorbed water film on the electrode surface, thereby interfering with the measurement accuracy of the local current density. Therefore, a humidity compensation mechanism can be considered to further improve the signal verification accuracy.

[0135] Specifically, a micro temperature and humidity sensor can be installed in the non-reaction area near the electrode, the sensor and the current collection node are kept 1 centimeter apart to avoid mutual interference, and the data collection needs to be synchronized with the local current density and the voltage response signal. A relative humidity-current density compensation coefficient correspondence table is established through preliminary tests: under different humidity conditions, the difference between the standard current density signal and the actual measured signal is collected, and the compensation coefficient of each humidity interval is calculated. When the humidity is higher than the benchmark humidity by 5%, the compensation coefficient is slightly greater than 1, which is used to correct the low measurement value caused by the thickening of the water film; when the humidity is lower than the benchmark humidity by 5%, the compensation coefficient is slightly less than 1, which is used to correct the high measurement value caused by the thinning of the water film.

[0136] In actual implementation, the temperature and humidity sensor data is called in real time, the corresponding compensation coefficients are matched to correct the local current density mean value and standard deviation, and then the correlation verification is carried out according to the original logic and the voltage response signal. If the humidity exceeds the reference humidity ± 10%, in addition to the conventional compensation, the humidity abnormality influence needs to be marked in the verification report, prompting the operation and maintenance personnel to pay attention to the state of the environmental control equipment. This way can eliminate the influence of humidity interference on the data, make the correlation verification of local current density and voltage signal more in line with the actual operating environment, and further improve the signal authenticity.

[0137] In some embodiments, the construction process of the harmonic feature subset also includes the extraction and integration steps of even harmonic parameters, specifically including:

[0138] For high-frequency probe signal response components, in addition to extracting 3rd and higher harmonic parameters, the amplitude distortion rate and phase shift value of 2nd and 4th even harmonics are additionally extracted;

[0139] For medium-frequency probe signal response components, in addition to extracting the amplitude ratio of fundamental wave and 2nd harmonic, the amplitude ratio of 4th even harmonic and fundamental wave and the frequency deviation of 4th even harmonic are additionally extracted;

[0140] The newly added amplitude distortion rate, phase shift value, amplitude ratio and frequency deviation of 2nd and 4th even harmonics are combined with each parameter of the original harmonic feature subset to form an extended harmonic feature subset containing even harmonic information;

[0141] The extended harmonic feature subset is corrected for temperature drift error, and then spliced with the three types of derived parameters of impedance change to form a real-time harmonic distortion feature vector containing even harmonic features.

[0142] In the harmonic feature analysis of PEM electrolytic water hydrogen production, the change of even harmonics, especially 2nd and 4th harmonics, is often directly related to hidden faults such as symmetry damage of proton exchange membrane and unevenness of electrode active area. This kind of feature is easy to be covered by high-order odd harmonics. Including it in the harmonic feature subset can significantly improve the comprehensiveness of fault identification, make up for the limitations of relying only on odd harmonic analysis, and make the real-time harmonic distortion feature vector more complete in reflecting the abnormality of electrolytic cell electrical characteristics.

[0143] In the implementation, the separated high-frequency probe signal response component and the medium-frequency probe signal response component are used as the core data source. The even harmonic parameters are extracted from the high-frequency probe signal response component. The triangular wave characteristics of the high-frequency probe signal contain rich harmonic components. The 3rd and higher harmonic parameters have been extracted, and this time, the 2nd and 4th even harmonics with relatively small amplitude are captured. These harmonics are more sensitive to the shedding of active substances on the electrode surface and the local thinning of the proton exchange membrane. In the operation, the frequency spectrum of the high-frequency response component is amplified by the frequency spectrum refinement algorithm. The amplitude of the even harmonic may be only one-tenth of the fundamental wave or even smaller, and it is easy to be covered by noise or high-order odd harmonics. The refined frequency spectrum can clearly show the peak position of the 2nd and 4th harmonics. After positioning the fundamental wave frequency, the frequency range of the 2nd and 4th even harmonics is locked according to the rule of fundamental wave frequency × 2 and fundamental wave frequency × 4. The amplitude distortion rate and the phase shift value of the two types of harmonics are calculated. The amplitude distortion rate is the ratio of the amplitude of the even harmonic to the amplitude of the fundamental wave, and the phase shift value is the difference between the phase of the even harmonic and the phase of the fundamental wave. In the calculation process, the harmonic amplitude needs to be noise filtered and corrected. The wavelet threshold denoising method is used to remove random interference in the frequency spectrum to ensure the accuracy of the parameters.

[0144] For the extraction of the even harmonics of the medium-frequency probe signal response component, the frequency spectrum characteristics of the square wave signal are combined. Although the amplitude of the even harmonic of the square wave is lower than that of the odd harmonic, it will fluctuate significantly when the electrolyte concentration gradient is abnormal. The amplitude ratio of the fundamental wave and the 2nd harmonic has been extracted, and this time, the related parameters of the 4th even harmonic are supplemented. The zero-crossing point detection of the time-domain waveform is used to determine the fundamental wave period of the medium-frequency signal, and the fundamental wave frequency is calculated. Then, the frequency of the 4th harmonic (fundamental wave frequency × 4) is located, and the amplitude and frequency values at this frequency point are extracted. The amplitude ratio of the 4th even harmonic and the fundamental wave is calculated. This ratio is in a stable interval when the electrolytic cell is running normally. If it suddenly increases or decreases, it usually indicates that the electrolyte is not evenly distributed on both sides of the membrane. At the same time, the frequency deviation of the 4th even harmonic from the theoretical frequency (fundamental wave frequency × 4) is calculated. The deviation value can reflect the local fluctuation of the electrode reaction rate. These two parameters are integrated with the existing fundamental wave-2nd harmonic amplitude ratio and 2nd harmonic frequency deviation to form the complete even harmonic characteristics of the medium-frequency signal.

[0145] After the extraction of even harmonic parameters of the two types of signals, the new parameters need to be merged with the original harmonic feature subset to construct an extended harmonic feature subset. Before merging, all parameters are dimensionally unified to ensure consistency in the calculation of amplitude parameters, uniformity in the calculation of phase parameters, and uniformity in the unit of frequency parameters. The merging order is divided into high-frequency signal features and medium-frequency signal signal features, and the internal high-frequency features are arranged in the order of 2nd harmonic, 4th harmonic, and 3rd and higher harmonics. The internal medium-frequency features are arranged in the order of 2nd harmonic and 4th harmonic. This order not only conforms to the logic of harmonic order, but also facilitates subsequent tracing of feature abnormalities by signal source, avoiding analysis errors caused by parameter confusion.

[0146] The extended harmonic feature subset still needs to be corrected for temperature drift error. The correction logic continues the method of taking the low-frequency calibration signal response component as the reference, but needs to adjust the compensation strategy for the temperature sensitivity of even harmonics. The coefficient of the phase shift of even harmonics affected by temperature is different from that of odd harmonics, and a temperature-even harmonic phase compensation coefficient correspondence table needs to be established through preliminary experiments. Standard even harmonic signals are collected at different temperatures, and the correlation between phase shift and temperature is recorded to obtain exclusive compensation coefficients for each temperature interval. When correcting, the real-time phase shift of the low-frequency calibration signal is first extracted, the temperature compensation reference value is calculated, and then the exclusive compensation coefficients of even harmonics are combined to accurately correct the phase shift values of 2nd and 4th harmonics in the extended subset; at the same time, according to the correlation between temperature and even harmonic amplitude, the amplitude distortion rate and amplitude ratio are adjusted synchronously to ensure that the corrected extended subset completely eliminates temperature interference.

[0147] Finally, the corrected extended harmonic feature subset is spliced with the three types of derived parameters of impedance changes to form a real-time harmonic distortion feature vector containing even harmonic features. The splicing order follows the principle of fault sensitivity priority, placing the high-frequency 2nd and 4th harmonic parameters with higher sensitivity to hidden faults in the front of the vector, followed by medium-frequency even harmonic parameters, high-order odd harmonic parameters, and finally impedance-derived parameters, ensuring that the artificial intelligence model can prioritize high-value features during reasoning. After splicing, data integrity verification is required to check for missing or abnormal values. If a certain even harmonic parameter exceeds the normal range, the parameter is marked and a secondary extraction process is started to avoid affecting the quality of the vector due to single extraction errors.

[0148] During the process, a dedicated quality evaluation mechanism for even harmonic parameters can be established to determine the reasonableness of the new parameters by comparing them with historical normal data. If the even harmonic amplitude distortion rate exceeds the normal range for three consecutive extractions, even if other parameters are normal, a preliminary warning is triggered to indicate the possible existence of hidden faults.

[0149] In some embodiments, S4 is the final execution link of the fault warning, and the core is to realize accurate decision-making through multi-parameter fusion and intelligent reasoning. The effective construction of the feature vector and the adaptive adjustment of the lightweight artificial intelligence model are the key to improving the response speed and accuracy of the warning, ensuring that the running characteristics of the electrolytic cell can be quickly matched, and avoiding the inference lag problem of general models.

[0150] When implementing, the selection and splicing logic of the characteristic parameters are first determined. The characteristic parameters with high correlation to faults need to be extracted from multi-dimensional running parameters, including the electrolyte temperature difference at the inlet and outlet of the electrolytic cell, the pressure difference on both sides of the proton exchange membrane, and the temperature distribution extreme value on the electrode surface. These parameters can directly reflect the thermal and mechanical state of the equipment, and are complementary to the frequency spectrum offset and impedance change. When splicing, the parameters are combined in the order of electrical characteristic parameters-thermal characteristic parameters-mechanical parameters, that is, the frequency spectrum offset and impedance change are arranged first, and then the temperature difference, pressure difference, and temperature distribution extreme value are connected. This ensures that similar parameters are concentrated in the vector, facilitating the model to quickly extract associated features.

[0151] The standardization processing of the feature vector uses the Z-score method, which converts all parameters to standardized data with a mean of 0 and a standard deviation of 1 by calculating the mean and standard deviation of each parameter. Before processing, abnormal values in the parameters need to be removed. If a parameter value exceeds the range of ±2 times the standard deviation of the mean, the mean of the parameter is replaced to avoid the influence of extreme data on the model inference.

[0152] The lightweight artificial intelligence model is based on an improved MobileNet architecture, with a focus on embedding a working condition adaptation attention module to achieve adaptive adjustment, addressing the poor adaptability of general models to different load conditions of the electrolytic cell. This module can dynamically adjust the weights of each feature channel of the model based on the real-time working current density of the electrolytic cell: when in low load conditions, it increases the channel weight of low-frequency related parameters; when in high load conditions, it strengthens the weight of high-frequency feature channels, allowing the model to focus on key fault signals under the current working condition. During model training, a mixed data set of normal working conditions and typical fault working conditions is used, covering feature vectors and corresponding fault labels under different loads and temperature conditions. Through transfer learning, the training period is shortened while preserving the lightweight characteristics of the model, ensuring that the response time during inference is controlled within 50 milliseconds, meeting real-time warning needs.

[0153] After inputting the standardized feature vector into the model, it is weighted by the working condition adaptation attention module, then passes through 3 layers of depth separable convolution to extract high-order features, and finally outputs the probability distribution of 6 typical faults such as electrode scaling, membrane damage, and poor contact through a fully connected layer. The probability distribution is presented in percentage form, and the sum of the probabilities of each type of fault is 100%, making it easy to visually determine the dominant fault risk.

[0154] The hierarchical early warning mechanism is executed according to the probability distribution result: if the probability of a certain type of fault exceeds 80%, it is determined as a first-level early warning, and the shutdown protection program is triggered immediately, and the fault position and type information are sent to the operation and maintenance terminal; the probability is between 50%-80% for a second-level early warning, the system automatically reduces the electrolytic cell working current density to the safe range, and starts the online monitoring frequency doubling mode; the probability is less than 50% for a third-level early warning, only the abnormal parameters are recorded, and no intervention is made to the operation, but the key attention items are marked in the operation log.

[0155] After the early warning is triggered, the system automatically associates the historical fault database, if the matching degree of the current fault type and the historical cases exceeds 90%, the historical processing scheme is pushed to the operation and maintenance personnel; if it is a new type of fault, the current feature vector and the operation data are saved, and samples are provided for subsequent optimization of the model. During the whole process, the model reasoning result and the early warning level are displayed in real time on the monitoring interface, and are stored in the local database, so that the fault information can be traced. Such a model optimized by adapting to the working condition can improve the reasoning accuracy while ensuring lightweight, and combined with the hierarchical early warning mechanism, the rapid response and scientific disposal of faults can be realized.

[0156] The above is only the preferred embodiment of the present application and the technical principle applied. The present application is not limited to the specific embodiments described herein, and various obvious changes, readjustments and substitutions made by those skilled in the art will not deviate from the protection scope of the present application. Therefore, although the present application has been described in more detail through the above embodiments, the present application is not limited to the above embodiments, and more other equivalent embodiments can be included without deviating from the concept of the present application.

Claims

1. A method for monitoring and early warning of key parameters in the PEM water electrolysis hydrogen production process, characterized in that, Includes the following steps: S1. Under the constant current steady-state operation condition of the electrolytic cell, a composite excitation signal is generated and injected into both ends of the electrolytic cell. The composite excitation signal includes a basic frequency sweep signal, a high-frequency probe signal, a medium-frequency probe signal and a low-frequency calibration signal. The amplitude of the composite excitation signal is related to the real-time operating current density and average temperature of the electrolytic cell. S2. Collect the multi-dimensional operating parameters of the electrolytic cell and perform time-frequency domain joint processing on the voltage response signal in the multi-dimensional operating parameters to obtain the time-frequency spectrum. S3. Based on the time spectrum diagram, extract the real-time harmonic distortion feature vector and impedance change of the electrolytic cell, calculate the spectral offset between the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector under the corresponding operating condition, and combine the energy ratio of different frequency bands and the impedance change to determine the fault tendency type and fault area of ​​the electrolytic cell. S4. The spectral offset, the impedance change, and the feature parameters in the multi-dimensional operating parameters are concatenated into a feature vector. After standardization, the vector is input into a lightweight artificial intelligence model for inference, and the probability distribution of various faults is output. Based on the probability distribution, graded early warning and corresponding handling measures are executed.

2. The method according to claim 1, characterized in that, In step S1, the basic frequency sweep signal is a sine wave. The frequency sweep range is determined by dividing the load interval according to the real-time operating current density of the electrolytic cell. When it is in the first load interval, the frequency sweep range is the first frequency sweep interval. When it is in the second load interval, the frequency sweep range is extended from the first frequency sweep interval to the second frequency sweep interval. When it is in the third load interval, the frequency sweep range changes to the third frequency sweep interval. The minimum frequency of the third frequency sweep interval is less than or equal to the maximum frequency of the first frequency sweep interval, and the maximum frequency of the third frequency sweep interval is greater than the maximum frequency of the second frequency sweep interval. The first load interval, the second load interval, and the third load interval are adjacent to each other and increase sequentially. The high-frequency probe signal is a triangular wave, the medium-frequency probe signal is a square wave, and the low-frequency calibration signal is a sine wave. The amplitude of the composite excitation signal is first set to a base value that increases with the increase of the real-time operating current density. Then, the base value of the amplitude is adjusted to compensate for the deviation of the average temperature from the reference temperature. The greater the temperature deviation, the greater the compensation.

3. The method according to claim 2, characterized in that, In step S1, the sweep rate of the basic sweep signal is a fixed value, the frequency of the high-frequency probe signal is higher than the maximum sweep frequency of the basic sweep signal, the frequency of the intermediate frequency probe signal is within the sweep range of the basic sweep signal, and the frequency of the low-frequency calibration signal is lower than the minimum sweep frequency of the basic sweep signal. The amplitude of the high-frequency probe signal is greater than the baseline amplitude value and has a fixed proportional relationship with the baseline amplitude value; the amplitude of the mid-frequency probe signal is greater than the baseline amplitude value and has another fixed proportional relationship with the baseline amplitude value; the amplitude of the low-frequency calibration signal is less than the baseline amplitude value and is a preset proportion of the baseline amplitude value. The duty cycles of the high-frequency probe signal and the intermediate-frequency probe signal are fixed values, and the duty cycle of the intermediate-frequency probe signal is less than that of the high-frequency probe signal.

4. The method according to claim 1, characterized in that, In step S2, the process of performing time-frequency domain joint processing on the voltage response signal includes: Based on the frequency range of each signal in the composite excitation signal, a targeted filtering algorithm is used to separate the response components that correspond one-to-one with the basic sweep frequency signal, the high-frequency probe signal, the intermediate frequency probe signal, and the low-frequency calibration signal. Perform time-frequency domain transformation on each response component to extract the time-frequency features of each response component, including instantaneous frequency, instantaneous amplitude and phase information; Based on the weight of each response component in fault monitoring, the time-frequency characteristics of each response component are weighted and fused to generate the time-frequency spectrum. The weight of each response component is related to the signal type corresponding to each response component.

5. The method according to claim 4, characterized in that, In step S3, the process of extracting the real-time harmonic distortion feature vector and calculating the impedance change includes: Based on the separated response components, the amplitude distortion rate and phase shift value of the third and higher harmonics are extracted for the high-frequency probe signal response components, and the amplitude ratio and frequency deviation of the fundamental and second harmonics are extracted for the mid-frequency probe signal response components. The amplitude distortion rate, the phase shift value, the amplitude ratio and the frequency deviation are combined into a harmonic feature subset. Using the phase reference of the low-frequency calibration signal response component as a reference, temperature drift error correction is performed on each parameter in the harmonic feature subset, and the equivalent impedance parameter of the electrolytic cell is calculated based on the full-band impedance spectrum data of the basic swept frequency signal response component, thereby obtaining the impedance change relative to the reference operating condition. The impedance change is calculated to obtain three types of derived parameters: the real part rate of change, the imaginary part rate of change, and the impedance angle change. The harmonic feature subset corrected for temperature drift error is then concatenated with the three types of derived parameters in a preset order to form the real-time harmonic distortion feature vector.

6. The method according to claim 5, characterized in that, In step S3, the process of calculating the spectral offset between the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector under the corresponding operating condition includes: Based on the current real-time operating current density of the electrolytic cell, the corresponding reference harmonic distortion feature vector is matched from the preset multi-condition reference parameter library; The real-time harmonic distortion feature vector is aligned dimension by dimension with the reference harmonic distortion feature vector, and the relative deviation values ​​of the corresponding dimension parameters are calculated. Based on the contribution of each dimension parameter to fault identification, a weighting coefficient is preset. The relative deviation value of each dimension parameter is multiplied by the corresponding weighting coefficient and then summed to obtain the comprehensive spectrum offset. The Euclidean distance between the real-time harmonic distortion feature vector and the reference harmonic distortion feature vector is calculated as an auxiliary offset index. The comprehensive spectrum offset and the auxiliary offset index together constitute the spectrum offset.

7. The method according to claim 6, characterized in that, In step S3, the process of determining the fault tendency type and fault area of ​​the electrolytic cell by combining the energy ratio of different frequency bands and the impedance change includes: Based on the time-spectrum diagram, extract the energy values ​​of each frequency band, calculate the energy ratio of any two frequency bands, and obtain at least two sets of energy ratios for different frequency bands; Retrieve a preset fault determination rule base, which contains pre-stored correlation features between different fault tendency types and corresponding fault regions and spectral offset, energy ratio of different frequency bands, and impedance change. Substitute the energy ratio of different frequency bands, the spectral offset, and the impedance change into the fault judgment rule base. First, the fault category is initially identified by the proportion of the real part change rate and / or the imaginary part change rate of the impedance change. Then, the specific fault tendency type is matched by the auxiliary offset index of the energy ratio of different frequency bands and the spectral offset. Finally, the corresponding fault area is located according to the abnormal energy distribution location of each frequency band.

8. The method according to claim 7, characterized in that, In step S3, after determining the fault tendency type and fault area of ​​the electrolytic cell, the steps of fault early warning verification and monitoring parameter adjustment are also included, specifically: Based on the judgment result, the preset verification excitation strategy is matched, and the parameters of the composite excitation signal are adjusted according to the fault tendency type. If it is a resistive fault, the amplitude ratio of the low frequency calibration signal is increased; if it is a reactive fault, the sweep frequency range of the basic sweep frequency signal is expanded. The adjustment amplitude is positively correlated with the spectral offset. The adjusted composite excitation signal is reapplied to the electrolytic cell, and S2 to S3 are repeated to obtain a second judgment result. The fault tendency type, fault area and spectral offset of the two judgments are compared. If the consistency is higher than the consistency threshold, a fault warning is triggered. If it is lower than the consistency threshold, it is marked as suspected interference.

9. The method according to claim 4, characterized in that, While performing time-frequency domain joint processing of the voltage response signal in S2, it also includes a local current density acquisition and signal verification process, specifically including: Local current density data corresponding to the voltage response signal are collected by setting multiple current acquisition nodes on the anode and cathode surfaces of the electrolytic cell. The mean and standard deviation of the current density at each node are calculated based on the collected local current density data to obtain the local current density distribution characteristics. The correlation between the local current density distribution characteristics and the separated basic sweep frequency signal response components is verified. If the local current density standard deviation exceeds the preset standard deviation range, the amplitude of the voltage response signal in that region is corrected accordingly. If the local current density standard deviation is within the preset standard deviation range, the voltage response signal is confirmed to be valid.

10. The method according to claim 5, characterized in that, The construction of the harmonic feature subset also includes the extraction and integration of even-order harmonic parameters, specifically including: For the response components of high-frequency probe signals, in addition to extracting the parameters of the third and higher harmonics, the amplitude distortion rate and phase shift value of the second and fourth even harmonics are also extracted. For the response components of the intermediate frequency probe signal, in addition to extracting the amplitude ratio of the fundamental wave to the second harmonic, the amplitude ratio of the fourth even harmonic to the fundamental wave and the frequency deviation of the fourth even harmonic are also extracted. The amplitude distortion rate, phase offset, amplitude ratio, and frequency deviation of the newly added 2nd and 4th even harmonics are combined with the parameters of the original harmonic feature subset to form an extended harmonic feature subset containing even harmonic information. Temperature drift error correction is applied to the extended harmonic feature subset, and then it is concatenated with three types of derived parameters of impedance change to form a real-time harmonic distortion feature vector containing even-order harmonic features.

Images