Hydrogen production system fault diagnosis method of KAN network based on Kolmogorov-Arnold time attention

Abstract

The invention discloses a hydrogen production system fault diagnosis method of a KAN network based on Kolmogorov-Arnold time attention, and relates to the technical field of fault diagnosis. The invention aims to solve the problems of low fault diagnosis accuracy and poor real-time performance of the existing fault detection method of the hydrogen production system. The method comprises the steps that the actual current, the actual voltage and the actual temperature of each working time point in the working process of a to-be-detected electrolytic cell are obtained and preprocessed; inputting the preprocessed actual current and actual temperature of the working time of the to-be-detected electrolytic cell into the hydrogen production system fault detection model to obtain a predicted voltage; obtaining the residual error of the actual working voltage of the to-be-measured electrolytic cell by using the predicted voltage and the preprocessed actual voltage; and judging whether the current hydrogen production system breaks down or not by using the residual error of the actual working voltage of the to-be-detected electrolytic cell. The method is used for fault detection of the hydrogen production system.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a hydrogen production system fault diagnosis method based on a Kolmogorov-Arnold temporal attention KAN network. Background Art

[0002] With the continuous growth of global energy demand and the increasing severity of environmental pollution, hydrogen energy, as a clean and sustainable energy source, has gradually become a key component of energy transformation. Hydrogen energy, with its high energy density, zero emissions, and broad application prospects, shows great potential, particularly in transportation, industry, and energy storage. The production, storage, and transportation of hydrogen are key links in hydrogen energy applications. Hydrogen production technology, as the core of the hydrogen energy industry chain, directly determines hydrogen supply capacity. Currently, there are several major hydrogen production technologies, including water electrolysis, natural gas reforming, and coal gasification. Water electrolysis, in particular, has garnered widespread attention in recent years because it uses electricity to decompose water into hydrogen and oxygen. With the rapid development of renewable energy sources (such as wind and solar energy), water electrolysis is considered a green, environmentally friendly, and efficient method for hydrogen production. However, with the large-scale application of hydrogen electrolysis equipment, equipment failures have become increasingly prominent, impacting the stability, reliability, and efficiency of hydrogen production. Therefore, improving the fault diagnosis capabilities of hydrogen production systems is crucial to ensure system safety, reduce maintenance costs, and improve production efficiency.

[0003] Fault diagnosis technology plays a crucial role in hydrogen electrolysis equipment. By monitoring the equipment's operating status in real time, it can promptly detect and identify system anomalies, preventing further failures and enabling prompt remedial action. This not only helps improve the operating efficiency and service life of hydrogen electrolysis equipment, but also ensures system safety, reduces downtime, and enhances production stability.

[0004] Currently, fault diagnosis methods based on physical models and signal processing are commonly used in hydrogen electrolysis devices. However, physical model-based fault diagnosis methods rely on precise mathematical models and system laws, which are computationally intensive and require a high level of prior knowledge of the system. Signal processing-based fault diagnosis methods address this issue, but they primarily infer fault conditions by analyzing signal changes in sensor data, often limiting their accuracy and real-time performance. Consequently, current fault diagnosis methods for hydrogen electrolysis devices still suffer from low accuracy and poor real-time performance. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of low fault diagnosis accuracy and poor real-time performance in the existing hydrogen production system fault detection methods, and propose a hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network.

[0006] The hydrogen production system fault diagnosis method based on Kolmogorov-Arnold temporal attention KAN network is as follows:

[0007] S1. Obtaining the actual current, actual voltage, and actual temperature at each working time point during the working process of the electrolytic cell to be tested, and preprocessing the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested to obtain the preprocessed actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested;

[0008] S2. Inputting the actual current and actual temperature of the electrolyzer to be tested during the pre-processed working time into the hydrogen production system fault detection model to obtain a predicted voltage;

[0009] S3, using the predicted voltage and the actual voltage after preprocessing to obtain the residual r' of the actual working voltage of the electrolytic cell to be tested i ;

[0010] S4, using the residual r' of the actual working voltage of the electrolytic cell to be tested i Determine whether the current hydrogen production system has a fault.

[0011] Furthermore, the step S1 of obtaining the actual current, actual voltage, and actual temperature at each working time point during the working process of the electrolytic cell to be tested, and preprocessing the actual current, actual voltage, and actual temperature during the working time of the electrolytic cell to be tested, is specifically as follows:

[0012] First, remove abnormal values in the actual current, actual voltage, and actual temperature of the electrolytic cell during operation;

[0013] Then, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after removing abnormal values are smoothed respectively;

[0014] Finally, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be measured after smoothing are normalized to obtain the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be measured after preprocessing.

[0015] Furthermore, the abnormal values in the actual current, actual voltage, and actual temperature of the electrolytic cell to be tested are removed respectively during the working time, specifically:

[0016] Draw the box plots corresponding to the actual working voltage, actual temperature and actual current data of the electrolytic cell to be tested respectively, and remove the data falling outside the upper and lower edges.

[0017] Furthermore, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after removing abnormal values are smoothed respectively, specifically as follows:

[0018]

[0019] in, is the smoothed value of the data on the electrolytic cell working time n, i is the working time index, γ is the smoothing parameter, is the original data value of the electrolytic cell at working time n;

[0020] The original data value is the actual current value, actual voltage value or actual temperature value of the electrolytic cell.

[0021] Furthermore, the residual r' of the actual working voltage of the electrolytic cell to be tested is obtained by using the predicted voltage and the actual voltage after pretreatment in S3. i , specifically:

[0022]

[0023] in, is the predicted voltage at operating time i, U' el,i is the actual voltage at operating time i after preconditioning.

[0024] Furthermore, the residual r' of the actual working voltage of the electrolytic cell to be tested is used in S4. i Determine whether the current hydrogen production system has a fault, specifically:

[0025] First, obtain the residual r' of the actual working voltage of the electrolyzer i The standard deviation σ' is:

[0026]

[0027] Then, determine whether the hydrogen production system has a fault based on σ':

[0028] If the residual r' is within the range of [-3σ',3σ'] i If the number is greater than the preset percentage, it means that the current hydrogen production system is normal; otherwise, it means that the current hydrogen production system is faulty.

[0029] Furthermore, the hydrogen production system fault detection model is obtained by the following method:

[0030] Step 1: Simulate and generate the operating voltage, temperature, and current of the electrolyzer in the hydrogen production system;

[0031] Step 2: The electrolytic cell operating voltage, electrolytic cell temperature, and electrolytic cell operating current generated by the simulation are combined into a data set, and the data set is divided into a training set and a test set. The training set is used to train the Kolmogorov-Arnold temporal attention based KAN network to obtain a trained Kolmogorov-Arnold temporal attention based KAN network;

[0032] Step 3: Use the test set to test the trained Kolmogorov-Arnold temporal attention-based KAN network to obtain the predicted operating voltage;

[0033] Step 4: Use the predicted working voltage obtained in step 3 and the working voltage generated by simulation to determine whether the trained KAN network based on Kolmogorov-Arnold time attention is saved. If saved, save the trained KAN network based on Kolmogorov-Arnold time attention as a hydrogen production system fault detection model; otherwise, return to step 1.

[0034] Furthermore, the KAN network based on Kolmogorov-Arnold temporal attention includes: a temporal attention module based on Kolmogorov-Arnold enhancement, a feature fusion module, and a KAN module;

[0035] The Kolmogorov-Arnold enhanced time attention module extracts timing features using the electrolytic cell operating temperature and operating current;

[0036] The feature fusion module is used to multiply the time series feature by the working temperature of the electrolytic cell to obtain the temperature feature, and multiply the time series feature by the working current of the electrolytic cell to obtain the current feature;

[0037] The KAN module uses current characteristics and temperature characteristics to obtain the electrolytic cell operating voltage.

[0038] Furthermore, the KAN module uses the current characteristics and temperature characteristics to obtain the electrolytic cell operating voltage, specifically:

[0039]

[0040] Among them, U el is the working voltage of the electrolytic cell, U r is the reversible voltage; I el is the current characteristic, r1 is the fixed ohmic resistance of the electrolytic cell, and r2 is the temperature-dependent ohmic resistance; T el is the temperature characteristic; k el is the activation overvoltage parameter, k T3 is the mass transfer overpressure parameter, A is the surface area of the electrolytic cell electrodes, kt1 It is a parameter that reflects the effect of temperature increase on reducing activation overvoltage, k t2 It is a parameter that reflects the effect of temperature changes on electrolyte conductivity and ion diffusion rate.

[0041] Furthermore, in step 4, the predicted operating voltage obtained in step 3 and the operating voltage generated by simulation are used to determine whether the trained KAN network based on Kolmogorov-Arnold time attention is saved. If it is saved, the trained KAN network based on Kolmogorov-Arnold time attention is saved as a hydrogen production system fault detection model; otherwise, return to step 1, specifically:

[0042] Step 4.1. Using the electrolytic cell working voltage U generated by simulation el,i and predicted operating voltage Get U el,i and The residuals are:

[0043]

[0044] in, is the predicted operating voltage at operating time i, U el,i is the operating voltage at operating time i generated by simulation;

[0045] Step 42: Determine whether the number of residual data in the range of [-3σ, 3σ] is greater than a preset percentage. If so, use the currently trained KAN model as a hydrogen production system fault detection model and save it; otherwise, return to step 1;

[0046] Where σ is the residual r i The standard deviation of .

[0047] The beneficial effects of the present invention are:

[0048] The present invention proposes a method for diagnosing hydrogen production system faults. The present invention first uses the Kolmogorov-Arnold temporal attention model to extract features of the electrolyzer's operating voltage and temperature, and then uses the KAN network model to obtain the electrolyzer's predicted voltage. The present invention establishes a hydrogen production system fault diagnosis model based on the relationship between the electrolyzer's current, temperature, and voltage. The hydrogen production system fault diagnosis model utilizes the characteristics that when the hydrogen production system operates normally, the relationship between the electrolyzer's current, temperature, and voltage is not affected by interference, while when the hydrogen production system fails, the relationship between the electrolyzer's current, temperature, and voltage will change, so that it can subsequently accurately determine whether the hydrogen production system has failed. The present invention determines whether the hydrogen production system has failed based on the residual between the electrolyzer's actual operating voltage and the predicted voltage, and can obtain accurate fault diagnosis results in real time, thereby improving the accuracy and real-time performance of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Flowchart of the present invention;

[0050] Figure 2 Schematic diagram of the Pauta criterion. DETAILED DESCRIPTION

[0051] Specific implementation method 1: Figure 1 As shown, the specific process of the hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network in this embodiment is as follows:

[0052] S1. Using a current sensor, a voltage sensor, and a temperature sensor to obtain the actual current, actual voltage, and actual temperature at each working time point during the working process of the electrolytic cell to be tested, and preprocessing the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested to obtain the preprocessed actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested;

[0053] The preprocessing is specifically as follows:

[0054] First, remove abnormal values in the actual current, actual voltage, and actual temperature of the electrolytic cell during operation:

[0055] Draw the box plots corresponding to the actual working voltage, actual temperature and actual current data of the electrolytic cell to be tested respectively. The data falling outside the upper and lower edges are abnormal values;

[0056] Then, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after removing the abnormal values are smoothed respectively, specifically as follows;

[0057]

[0058] in, is the smoothed value of the data on the electrolytic cell working time n, i is the working time index, γ is the smoothing parameter, is the original data value at the working time n of the solution slot;

[0059] The original data value is an actual current value, an actual voltage value or an actual temperature value.

[0060] Finally, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after smoothing are normalized to obtain the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after preprocessing;

[0061] In this step, outliers may be caused by factors such as sensor failure or external interference. Noisy data can obscure the device's true operating status, and duplicate data can increase the data processing burden. Data cleaning can improve data quality and provide more accurate data support for subsequent analysis. Furthermore, data normalization is required to convert data of different dimensions and ranges to the same scale for easier input and processing by the neural network.

[0062] In the fault diagnosis process of the electrolytic hydrogen production device, data extraction is required for both simulation data and actual data. By installing various sensors on the device, such as current sensors, voltage sensors, and temperature sensors, various parameter data during the operation of the device are collected in real time. These sensors can accurately capture the operating status information of the device under different working conditions, including the magnitude of the current, the stability of the voltage, and the change in temperature. At the same time, it is also necessary to record information such as the operating conditions and environmental conditions of the device, such as the load conditions of the device, the external temperature, etc. These data will serve as the basis for subsequent normal mode model establishment and fault diagnosis. In the data extraction process, in order to ensure the accuracy and reliability of the data, the collected data needs to be preprocessed. The normal mode model in the present invention refers to a KAN network based on Kolmogorov-Arnold temporal attention that learns the mapping relationship between the voltage, temperature, and current of the electrolyzer.

[0063] S2. Inputting the actual current and actual temperature of the electrolyzer under test during the current working time after pretreatment into the hydrogen production system fault detection model to obtain a predicted voltage;

[0064] S3, using the predicted voltage and the actual voltage after pretreatment to obtain the residual r' of the actual working voltage of the electrolytic cell to be tested i , specifically:

[0065]

[0066] in, is the predicted voltage at operating time i, U' el,i is the actual voltage at working time i after pretreatment;

[0067] S4, using the residual r' of the actual working voltage of the electrolytic cell to be tested i Determine whether the current hydrogen production system is faulty, specifically:

[0068] If the residual r' is within the range of [-3σ',3σ'] i If the number is greater than the preset percentage, it means that the current hydrogen production system is normal; otherwise, it means that the current hydrogen production system is faulty;

[0069]

[0070] Where σ' is the residual r' of the actual working voltage of the electrolyzer i The standard deviation of , i is the working time index, and n is the total number of working time points.

[0071] Specific embodiment 2: The hydrogen production system fault detection model is obtained by the following method:

[0072] Step 1: Use Simulink in Matlab to simulate and generate the working voltage, temperature and current of the electrolyzer in the hydrogen production system;

[0073] Step 2: The electrolytic cell operating voltage, electrolytic cell temperature, and electrolytic cell operating current generated by the simulation in step 1 are used to form a data set, and the data set is divided into a training set and a test set in a ratio of 2:1. The training set is used to train a Kolmogorov-Arnold temporal attention-based KAN network to obtain a trained Kolmogorov-Arnold temporal attention-based KAN network;

[0074] The KAN network based on Kolmogorov-Arnold temporal attention includes: a temporal attention module based on Kolmogorov-Arnold enhancement, a feature fusion module, and a KAN module;

[0075] The Kolmogorov-Arnold enhanced time attention module is used to extract timing features according to the working temperature and working current of the electrolyzer;

[0076] The feature fusion module is used to multiply the time series feature by the working temperature of the electrolytic cell to obtain the temperature feature, and multiply the time series feature by the working current of the electrolytic cell to obtain the current feature;

[0077] The KAN module uses current characteristics and temperature characteristics to obtain the electrolytic cell operating voltage, specifically:

[0078] U el =F(T el ,I el )

[0079]

[0080] Among them, U el is the working voltage of the electrolytic cell, U r is the reversible voltage of the electrolytic cell that changes with temperature; I el is the current characteristic; r1 is the fixed ohmic resistance of the electrolytic cell, reflecting the positive resistance of the electrolytic cell material itself; r2 is the temperature-dependent ohmic resistance, reflecting the change in electrolyte / membrane conductivity; T el is the temperature characteristic; k el is the activation overvoltage parameter, kT3 is the mass transfer overpressure parameter, A is the surface area of the electrolytic cell electrodes, k t1 It is a parameter that reflects the effect of temperature increase on reducing activation overvoltage, k t2 It is a parameter that reflects the effect of temperature changes on electrolyte conductivity and ion diffusion rate.

[0081] Step 3: Use the test set to test the trained KAN network based on Kolmogorov-Arnold temporal attention to obtain the predicted working voltage

[0082] Step 4: Use the predicted operating voltage obtained in step 3 And the simulated operating voltage U el Determine whether the trained Kolmogorov-Arnold time attention-based KAN network is saved. If so, save the trained Kolmogorov-Arnold time attention-based KAN network as a hydrogen production system fault detection model. Otherwise, return to step 1, specifically:

[0083] Step 4.1. Using the electrolytic cell working voltage U generated by simulation el and predicted operating voltage Get U el and The residuals are:

[0084]

[0085] Step 4.2: Determine r within the range of [-3σ, 3σ] i Whether the amount of data is greater than the preset percentage, if it is, the currently trained KAN model is used as the hydrogen production system fault detection model and saved; otherwise, return to step 1;

[0086] Where σ is the residual r i The standard deviation of .

[0087] The present invention adopts the core idea of establishing a normal mode, and extracts the "invariant" of the normal mode of the electrolytic hydrogen production device from various interference factors such as operating conditions and environmental conditions and changes in various measurable parameters. This "invariant" can be a certain parameter or a certain mapping relationship, etc. It needs to meet two conditions: when the electrolytic hydrogen production device is operating normally, although various interference factors such as operating conditions and environmental conditions have undergone major changes, the "invariant" remains basically unchanged or changes very little; and when the electrolytic hydrogen production device fails, the "invariant" will change significantly. In order to extract such an "invariant", it is first necessary to perform feature selection on the collected data. The electrolytic device can be regarded as a nonlinear load, and studying its UI characteristics can improve gas production efficiency. The voltage and current characteristics of the electrolytic cell are determined by the electrolytic cell temperature. Due to its high nonlinearity, general curve fitting cannot be used to represent it.

[0088] Since there is not much actual data on the electrolytic hydrogen production device, the trained normal mode model cannot be accurately expressed. The present invention obtains a large amount of simulation data through a simulation device, and the KAN network based on Kolmogorov-Arnold temporal attention extracts simulation data to construct a normal mode model domain, which is close to the correct normal mode model, and then corrects the domain through actual data to make its expression effect accurate. The KAN neural network is a neural network based on the Kolmogorov-Arnold theorem, which has a strong function approximation ability and can effectively handle complex nonlinear relationships. By training the KAN neural network, it learns the intrinsic mapping relationship between characteristic parameters, thereby extracting "invariants" that meet the conditions. During the training process, the error back propagation algorithm is used to optimize the network parameters so that the "invariants" output by the network remain stable during normal operation and change significantly when a fault occurs. The normal mode model established in this way can accurately describe the characteristics of the electrolytic hydrogen production device under normal operating conditions, providing a benchmark for subsequent fault diagnosis.

[0089] If the trained KAN network can well characterize the mapping relationship between input data and output data, the distribution of residuals approximately follows the normal distribution N(0,σ 2 )。 N(0,σ 2 )’s probability density function is as follows Figure 2 As shown in the figure, δ is the standard score, and φ(δ) is the probability density function. The probability that the residual is within [-3σ, 3σ] is 99.73%. Therefore, the Pauta criterion (i.e., the 3σ criterion) can be used to determine the threshold for fault detection. For the sample to be detected, the trained network first calculates the fitted voltage value. Then, the residual is calculated between the predicted value and the actual value. If the absolute value of the residual does not exceed 3σ, the sample is detected as normal; otherwise, it is detected as a faulty sample.

[0090] The present invention designs a KAN (Kolmogorov-Arnold Network) neural network based on Kolmogorov-Arnold temporal attention, which can extract the normal mode and fault mode of the hydrogen electrolysis device in a complex multidimensional data environment, thereby achieving accurate fault diagnosis. The core idea of the normal mode extraction method is to extract the "invariant" of the normal mode of the gas turbine from various interference factors such as operating conditions and environmental conditions and changes in various measurable parameters. This "invariant" can be, but is not limited to, a certain parameter, a certain mapping relationship, etc. This "invariant" needs to meet the following two conditions:

[0091] First, when the device is operating normally, despite significant changes in various interference factors such as operating conditions and environmental conditions, the "invariant" remains basically unchanged or changes very little.

[0092] Second, when a device fails, this “invariant” can change significantly.

[0093] Once an "invariant" that satisfies the above two conditions is extracted, sensitive and robust gas turbine fault detection can be achieved by detecting changes in the "invariant." The present invention searches for the "invariant" in the normal mode of the device.

[0094] In fault diagnosis of hydrogen electrolysis devices, the KAN network can learn the normal operating mode of the system from large amounts of sensor data and promptly detect anomalies when faults occur. A significant advantage of the KAN network is its ability to effectively cope with data scarcity, especially in industrial environments where sensor data may be missing or noisy. Through its adaptive modeling properties, the KAN network can better extract effective information from incomplete data. This makes the KAN network highly robust in fault diagnosis of hydrogen electrolysis devices. Furthermore, when dealing with highly nonlinear and time-varying fault modes, the KAN network can automatically adjust its internal structure and parameters, ensuring that the model maintains high diagnostic accuracy. This provides powerful technical support for real-time monitoring and fault warning of hydrogen electrolysis devices. Compared with traditional methods, the KAN network not only provides more accurate fault diagnosis results, but also offers greater real-time performance, adapting to rapidly changing industrial environments.

[0095] Example: In order to verify the beneficial effects of the present invention, the present invention conducted the following experiments:

[0096] The input and output of the hydrogen production system normal mode model (KAN network based on Kolmogorov-Arnold temporal attention) are obtained as shown in Table 1:

[0097] Table 1

[0098]

[0099] Since there was no fault data, the model's detection performance on normal data was tested. The final detection results are shown in Table 2. The proposed hydrogen production system fault diagnosis method can effectively detect the normal operation of the hydrogen production system. Compared with traditional methods, the introduced Kolmogorov-Arnold temporal attention KAN architecture performs well in fault diagnosis. This demonstrates that the proposed method has high accuracy and robustness in fault detection.

[0100] Table 2

[0101]

Claims

1. A hydrogen production system fault diagnosis method based on Kolmogorov-Arnold temporal attention KAN network is characterized by The specific process of the method is: S1. Obtaining the actual current, actual voltage, and actual temperature at each working time point during the working process of the electrolytic cell to be tested, and preprocessing the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested to obtain the preprocessed actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested; S2. Inputting the actual current and actual temperature of the electrolyzer to be tested during the pre-processed working time into the hydrogen production system fault detection model to obtain a predicted voltage; S3, using the predicted voltage and the actual voltage after preprocessing to obtain the residual r' of the actual working voltage of the electrolytic cell to be tested i ; S4, using the residual r' of the actual working voltage of the electrolytic cell to be tested i Determine whether the current hydrogen production system has a fault.

2. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 1 is characterized in that: The step S1 is to obtain the actual current, actual voltage, and actual temperature at each working time point during the working process of the electrolytic cell to be tested, and pre-process the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested, specifically as follows: First, remove abnormal values in the actual current, actual voltage, and actual temperature of the electrolytic cell during operation; Then, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after removing abnormal values are smoothed respectively; Finally, the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be measured after smoothing are normalized to obtain the actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be measured after preprocessing.

3. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 2 is characterized in that: The method of removing abnormal values in the actual current, actual voltage, and actual temperature of the electrolytic cell during operation is specifically as follows: Draw the box plots corresponding to the actual working voltage, actual temperature and actual current data of the electrolytic cell to be tested respectively, and remove the data falling outside the upper and lower edges.

4. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 3 is characterized in that: The actual current, actual voltage, and actual temperature of the working time of the electrolytic cell to be tested after removing abnormal values are smoothed respectively, specifically as follows: in, is the smoothed value of the data on the electrolytic cell working time n, i is the working time index, γ is the smoothing parameter, is the original data value of the electrolytic cell at working time n; The original data value is the actual current value, actual voltage value or actual temperature value of the electrolytic cell.

5. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 4 is characterized in that: In S3, the residual r' of the actual working voltage of the electrolytic cell to be tested is obtained by using the predicted voltage and the actual voltage after pretreatment. i , specifically: in, is the predicted voltage at operating time i, U' el,i is the actual voltage at operating time i after preconditioning.

6. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 5 is characterized in that: The residual r' of the actual working voltage of the electrolytic cell to be tested is used in S4 i Determine whether the current hydrogen production system has a fault, specifically: First, obtain the residual r' of the actual working voltage of the electrolyzer i The standard deviation σ' is: Then, determine whether the hydrogen production system has a fault based on σ': If the residual r' is within the range of [-3σ',3σ'] i If the number is greater than the preset percentage, it means that the current hydrogen production system is normal; otherwise, it means that the current hydrogen production system is faulty.

7. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 6 is characterized in that: The hydrogen production system fault detection model is obtained by the following method: Step 1: Simulate and generate the operating voltage, temperature, and current of the electrolyzer in the hydrogen production system; Step 2: The electrolytic cell operating voltage, electrolytic cell temperature, and electrolytic cell operating current generated by the simulation are combined into a data set, and the data set is divided into a training set and a test set. The training set is used to train the Kolmogorov-Arnold temporal attention based KAN network to obtain a trained Kolmogorov-Arnold temporal attention based KAN network; Step 3: Use the test set to test the trained Kolmogorov-Arnold temporal attention-based KAN network to obtain the predicted operating voltage; Step 4: Use the predicted working voltage obtained in step 3 and the working voltage generated by simulation to determine whether the trained KAN network based on Kolmogorov-Arnold time attention is saved. If saved, save the trained KAN network based on Kolmogorov-Arnold time attention as a hydrogen production system fault detection model; otherwise, return to step 1.

8. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 7 is characterized in that: The KAN network based on Kolmogorov-Arnold temporal attention includes: a temporal attention module based on Kolmogorov-Arnold enhancement, a feature fusion module, and a KAN module; The Kolmogorov-Arnold enhanced time attention module extracts timing features using the electrolytic cell operating temperature and operating current; The feature fusion module is used to multiply the time series feature by the working temperature of the electrolytic cell to obtain the temperature feature, and multiply the time series feature by the working current of the electrolytic cell to obtain the current feature; The KAN module uses current characteristics and temperature characteristics to obtain the electrolytic cell operating voltage.

9. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 8 is characterized in that: The KAN module uses current characteristics and temperature characteristics to obtain the electrolytic cell operating voltage, specifically: Among them, U el is the working voltage of the electrolytic cell, U r is the reversible voltage; I el is the current characteristic, r1 is the fixed ohmic resistance of the electrolytic cell, and r2 is the temperature-dependent ohmic resistance; T el is the temperature characteristic; k el is the activation overvoltage parameter, k T3 is the mass transfer overpressure parameter, A is the surface area of the electrolytic cell electrodes, k t1 It is a parameter that reflects the effect of temperature increase on reducing activation overvoltage, k t2 It is a parameter that reflects the effect of temperature changes on electrolyte conductivity and ion diffusion rate.

10. The hydrogen production system fault diagnosis method based on the Kolmogorov-Arnold temporal attention KAN network according to claim 9 is characterized in that: In the fourth step, the predicted working voltage obtained in the third step and the working voltage generated by simulation are used to determine whether the trained KAN network based on Kolmogorov-Arnold time attention is saved. If it is saved, the trained KAN network based on Kolmogorov-Arnold time attention is saved as a hydrogen production system fault detection model; otherwise, return to step one, specifically: Step 4.

1. Using the electrolytic cell working voltage U generated by simulation el,i and predicted operating voltage Get U el,i and The residuals are: in, is the predicted operating voltage at operating time i, U el,i is the operating voltage at operating time i generated by simulation; Step 42: Determine whether the number of residual data in the range of [-3σ, 3σ] is greater than a preset percentage. If so, use the currently trained KAN model as a hydrogen production system fault detection model and save it; otherwise, return to step 1; Where σ is the residual r i The standard deviation of .

Images