A Real-Time Fault Diagnosis Method for Proton Exchange Membrane Fuel Cells Across Loads Based on Electrochemical-Magnetic Cross-Integration

Abstract

This invention provides a real-time fault diagnosis method for proton exchange membrane fuel cells across loads based on electrochemical-magnetic cross-integration. It acquires broadband excitation electrochemical response signals and non-invasive magnetic field distribution signals after environmental interference removal; it performs frequency domain separation of the electrochemical signals to extract electrochemical mechanism features of proton transport, electrochemical activation, and mass transport processes; it constructs a magnetic sensor node graph structure for the magnetic field signals to extract load-invariant spatial features; and it performs bidirectional complementary fusion and adaptive weighting through a gated cross-attention mechanism. The fused features are then processed by a classifier to output fault diagnosis results. Based on the diagnosis results, health status assessment and preventative maintenance decisions are implemented, constructing a health management system to support reliable operation of the fuel cell throughout its entire lifecycle. This invention integrates electrochemical mechanism characterization with the load-invariant characteristics of the magnetic field, improving the real-time performance and accuracy of cross-load diagnosis and meeting the needs of online diagnosis under dynamic operating conditions.

Description

Technical Field

[0001] This invention relates to the field of fuel cell fault diagnosis technology, specifically to a real-time fault diagnosis method for proton exchange membrane fuel cells across loads based on electrochemical-magnetic cross-integration. Background Technology

[0002] Proton exchange membrane fuel cells (PEMFCs) have become a key power technology in the fields of new energy vehicles and distributed power generation due to their high energy conversion efficiency and zero emissions. However, under actual dynamic load conditions, the battery is prone to sudden failures such as flooding, membrane dehydration, and air starvation, leading to a sharp decline in performance and even a shortened lifespan. Therefore, developing fault diagnosis methods with real-time response capabilities and adaptability to load changes is of great significance for promoting the commercial application of PEMFCs.

[0003] Currently, PEMFC fault diagnosis is mainly divided into model-based methods and data-driven methods. Model-based diagnosis relies on accurate mechanistic modeling, but the strong coupling characteristics of multiple physics fields (electrochemical, thermodynamic, and fluid dynamics) within PEMFCs make model construction extremely difficult, and diagnostic stability is hard to guarantee in engineering applications. Data-driven methods, on the other hand, collect external lumped signals such as voltage, current, and temperature, and use deep learning networks to extract fault features, and have become the mainstream technical approach. However, external macroscopic signals are difficult to characterize the microscopic electrochemical processes inside the battery, such as proton transport, electrochemical activation, and mass transport, and the feature distribution shifts drastically with load changes, causing the model's diagnostic performance to drop sharply under cross-load conditions.

[0004] Patent CN121659036A, entitled "A Method, Device, Equipment, and Medium for Fault Diagnosis of Fuel Cells," proposes a fault diagnosis method that integrates internal partitioning features and external lumped features. It utilizes CNN to extract spatial local features, SEAM modules to enhance sensitive features, and LSTM to capture temporal dependencies. While this method improves diagnostic accuracy to some extent, it belongs to a purely data-driven paradigm. The extracted features lack physical constraints on the internal electrochemical mechanisms of the fuel cell, and are prone to shift under dynamic load changes.

[0005] Patent CN121862793A, "A Fault Diagnosis Method and Related Equipment for Proton Exchange Membrane Fuel Cells," discloses a physical information-guided transfer learning method. This method involves pre-training the source domain and then fine-tuning the parameters of the target domain, introducing the Nernst equation to calculate the deviation between the theoretical and measured voltages as a physical consistency loss. While this method has achieved good results across different devices or operating conditions, its input is still limited to external lumped signals such as voltage, current, temperature, and humidity. It cannot directly perceive the dynamic evolution of different electrochemical processes within the battery, and its adaptability to the real-time diagnostic needs of the same device under rapid dynamic load changes is insufficient.

[0006] In summary, current technologies have not yet formed a PEMFC diagnostic scheme that can simultaneously achieve real-time response, cross-load stability, and fault mechanism differentiation. While some methods can acquire internal electrochemical information, they are time-consuming and highly load-sensitive; others possess real-time performance and load invariance, but struggle to distinguish between faults with similar spatial effects, such as flooding and air starvation. Therefore, there is an urgent need to develop a technology that integrates multimodal physical signals, possesses inherent load invariance characteristics, and can accurately diagnose the electrochemical root causes of faults, to meet the practical needs of reliable, real-time, and accurate PEMFC diagnosis under dynamic load scenarios. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a real-time fault diagnosis method for proton exchange membrane fuel cells across loads based on electrochemical-magnetic cross-fusion, solving the technical problems of "insufficient real-time fault diagnosis of existing proton exchange membrane fuel cells under dynamic load conditions and low accuracy of cross-load scenario diagnosis".

[0008] To achieve the above objectives, the present invention is implemented using the following technical solution: This invention provides a real-time fault diagnosis method for proton exchange membrane fuel cells across loads based on electrochemical-magnetic cross-integration, comprising the following steps: Acquire broadband excitation electrochemical response signal and non-invasive magnetic field distribution signal of fuel cell; The broadband excitation electrochemical response signal is frequency domain separated to extract the electrochemical mechanism features corresponding to proton transport, electrochemical activation and mass transport processes, respectively. A graph structure with magnetic sensors as nodes is constructed for the non-invasive magnetic field distribution signal to extract load-invariant spatial features; The electrochemical mechanism features and the load-invariant spatial features are bidirectionally complementary and adaptively weighted by a gated cross-attention mechanism to obtain fused features; The fused features are input into the classifier, which outputs the fault diagnosis results. Based on the output fault diagnosis results, health management is carried out, a health status quantitative assessment and level classification strategy is constructed, maintenance decisions are generated, and health regulation is carried out in real time.

[0009] Furthermore, the acquisition of the broadband excitation electrochemical response signal includes: based on the characteristic frequencies of three types of electrochemical processes—proton transport, electrochemical activation, and mass transport—iteratively supplementing frequency points using an adaptive frequency selection method guided by the Fisher information matrix to obtain an optimal discrete frequency set; synthesizing the sinusoidal signals of the optimal discrete frequency set into a single broadband composite excitation signal, applying it to the fuel cell once, and simultaneously acquiring the AC response voltage time-domain signal.

[0010] Furthermore, the acquisition of non-invasive magnetic field distribution signals includes: attaching a magnetic sensor array to the outer surface of the cathode side of the fuel cell, covering the active reaction area of ​​the battery; acquiring the ambient magnetic field under zero load as a background reference, acquiring the original magnetic field data under steady-state operation of the battery, eliminating ambient magnetic field interference through background subtraction, and finally obtaining effective magnetic field distribution data generated only by the internal current response of the battery.

[0011] Furthermore, the excitation current amplitude at each frequency point in the excitation signal is distributed using an inverse Gaussian distribution, and the excitation current does not exceed 5% of the load current; wherein, the amplitude of the low-frequency band is increased, the amplitude of the mid-frequency band is decreased, and the amplitude of the high-frequency band is increased in the corresponding mass transfer process.

[0012] Furthermore, the extraction of load-invariant spatial features uses the measured values ​​of each magnetic sensor, the local magnetic gradient magnitude, and the normalized spatial coordinates as the initial features of the nodes. An adjacency matrix is ​​constructed based on the k-nearest neighbor strategy. A multi-layer graph convolutional network is used to perform deep aggregation of spatial features on the graph structure data. After global average pooling, the magnetic feature vector is obtained.

[0013] Furthermore, the extraction of electrochemical mechanism features includes: designing a frequency domain bandpass filter based on the characteristic frequency bands of the three types of electrochemical processes, and converting it into a time domain convolution kernel through inverse Fourier transform; convolving the broadband excitation electrochemical response signal with the convolution kernel corresponding to each process to separate the frequency response components of each electrochemical process; configuring a sub-convolutional network for each electrochemical process, extracting process-level features bound to the parameters of the equivalent circuit model, and concatenating them to obtain an electrochemical feature vector.

[0014] Furthermore, the sub-convolutional network includes three sub-convolutional networks corresponding to the proton transport process, the electrochemical activation process, and the mass transport process, respectively; wherein, the features extracted by the proton transport sub-network are related to the series resistance in the equivalent circuit model, the features extracted by the electrochemical activation sub-network are related to the charge transfer resistance and the parameters of the constant phase element, and the features extracted by the mass transport sub-network are related to the gas diffusion resistance and the parameters of the RLC parallel circuit.

[0015] Furthermore, the gated cross-attention mechanism includes: mapping electrochemical mechanism features and load-invariant spatial features to a shared semantic space using a trainable projection matrix; calculating a first cross-attention using electrochemical features as queries and magnetic features as bonds and values, and calculating a second cross-attention using magnetic features as queries and electrochemical features as bonds and values; after performing residual connection and layer normalization on the cross-attention output and the original features, generating a gated vector through the sigmoid function, and adaptively weighting and fusing the dual-modal features.

[0016] Furthermore, the fault diagnosis results include probability values ​​for four health states: normal, flooded, dehydrated, and starved. The classification confidence threshold is set to 0.85, and a secondary diagnosis is triggered when the confidence level is below the threshold.

[0017] Furthermore, the health management involves: constructing a quantitative assessment strategy for fuel cell health status based on the output probability values ​​of four health states (normal, flooded, dehydrated, and starved) and continuous time series historical data, integrating fault type, frequency of occurrence, and severity of single faults, and completing multi-level classification of health status; generating differentiated preventive maintenance decisions based on health status results and fault root cause characteristics, clarifying maintenance content, execution cycle, and resource priority; and simultaneously optimizing the gas metering ratio, humidification temperature, and operating temperature parameters of the battery in real time according to fault type and health status, achieving self-healing of minor faults and health status regulation.

[0018] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention shortens the traditional electrochemical impedance spectroscopy frequency-by-frequency scanning measurement from minutes to seconds by using Fisher information matrix-guided adaptive frequency selection and broadband excitation signal synthesis. Combined with the rapid detection of non-invasive magnetic field array, the end-to-end full-process diagnosis takes only 7.7 seconds, which significantly improves the real-time performance of fault diagnosis in dynamic load scenarios.

[0019] (2) This invention integrates the microscopic mechanism characterization capability of electrochemical impedance spectroscopy with the load-invariant characteristics of magnetic field mapping. It utilizes a gated cross-attention mechanism to perform bidirectional complementary fusion and adaptive weighting of electrochemical mechanism features and load-invariant spatial features. The diagnostic accuracy under cross-load conditions can reach up to 97%, effectively overcoming the shortcomings of traditional methods that experience a sharp decline in performance when the load changes.

[0020] (3) This invention extracts the electrochemical characteristics of three internal processes—proton transport, electrochemical activation, and mass transport—through frequency domain separation, and captures the spatial gradient information of the magnetic field using graph convolutional networks. This effectively distinguishes faults with similar spatial effects but different mechanisms, such as flooding and air starvation. Ablation experiments show that the misclassification rate is reduced to below 7%, significantly improving the accuracy of fault classification. This invention takes health status assessment as its core and preventive maintenance as its support, transforming traditional passive repair into proactive prevention, achieving refined health status characterization and differentiated maintenance, and reducing the risk of sudden downtime. Experiments show that this scheme can reduce unplanned downtime by more than 30%, reduce operation and maintenance costs by more than 22%, and improve the operational reliability and commercial value of fuel cell systems. Attached Figure Description

[0021] Figure 1 This invention provides a hardware platform for testing proton exchange membrane fuel cells.

[0022] Figure 2 This diagram shows the wideband current excitation signal of the battery under different load conditions.

[0023] Figure 3 This is a diagram showing the distribution of the battery's magnetic field under different load conditions.

[0024] Figure 4 The diagnostic accuracy of the method in this application is shown under different training-test data configurations.

[0025] Figure 5 This is a comparison chart of the diagnostic accuracy of different control models under different configurations in the ablation experiment.

[0026] Figure 6 Confusion matrix diagram of diagnostic results of each ablation model under different training-test data configurations.

[0027] Figure 7 This is a comparison chart of real-time signal acquisition and feature fidelity analysis.

[0028] Figure 8 This is a comparison chart of the model's fast convergence performance and cross-load diagnostic performance.

[0029] Figure 9 A performance comparison chart of different battery fault diagnosis methods. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1; refer to Figure 1 This embodiment constructs a proton exchange membrane fuel cell (PEMFC) testing hardware platform, which includes a PEMFC (proton exchange membrane fuel cell) operation testing unit, an electrochemical signal excitation and acquisition unit, a non-invasive magnetic field distribution detection unit, and a host computer data processing and diagnostic unit. Through the platform, the battery operating current density, gas metering ratio, humidification temperature, and operating temperature can be precisely adjusted to reproduce various operating states such as normal, flooded, dehydrated, and air-starved, as well as cross-load variable operating conditions.

[0032] The PEMFC operation test unit adopts a fuel cell test system, which integrates a gas supply subsystem, a humidification subsystem, a temperature control subsystem, and a single cell module. It can simulate the normal operation and various fault states of PEMFC under variable load conditions.

[0033] The electrochemical signal excitation and acquisition unit is configured as a chemical workstation, used to generate multi-sinusoidal current excitation signals with custom frequency components, and to complete EIS measurement and AC response voltage acquisition.

[0034] The non-invasive magnetic field distribution detection unit uses a Mag612 magnetic sensor matrix, which integrates 16 independent triaxial magnetic sensors arranged in a 4×4 array and closely attached to the outer surface of the PEMFC cathode side for real-time monitoring of the external magnetic field distribution of the battery.

[0035] The host computer data processing and diagnosis unit uses an industrial control computer equipped with a high-performance processor and a graphics processor to complete excitation signal synthesis, data preprocessing, network model training, and end-to-end fault diagnosis calculations.

[0036] Example 2; This embodiment provides a real-time fault diagnosis method for proton exchange membrane fuel cells across loads based on electrochemical-magnetic cross-integration. This embodiment uses the proton exchange membrane fuel cell test hardware platform of Embodiment 1.

[0037] This embodiment employs a time-division multiplexing method to acquire electrochemical response signals and magnetic field distribution signals, avoiding electromagnetic interference between the excitation signal and magnetic measurements. The entire implementation process is divided into three parts: signal acquisition, data preprocessing and sample set construction, and diagnostic network construction and training. The specific flow is as follows: S1. Acquire broadband excitation electrochemical signals using an adaptive frequency selection method guided by the Fisher Information Matrix (FIM). The steps are as follows: (1) Screening of key frequencies: Based on the characteristic frequencies of the three core electrochemical processes of proton transport, electrochemical activation and mass transport in PEMFC, 2kHz (proton transport), 100Hz (electrochemical activation) and 1Hz (mass transport) were used as the initial frequency set. The error propagation framework constructed by FIM was used to iteratively supplement the frequency points where the impedance prediction error of the whole band exceeded the threshold of 5%. Finally, 10 optimal discrete frequency points were obtained, namely 0.5Hz, 1Hz, 5Hz, 10Hz, 20Hz, 50Hz, 100Hz, 500Hz, 1000Hz and 2000Hz.

[0038] (2) Excitation signal amplitude design: The excitation current amplitude is allocated to each frequency point using an inverse Gaussian (Wald) distribution, strictly adhering to the safety constraint that the excitation current does not exceed 5% of the load current. For the typical operating condition of 0.6A / cm² current density and 15A load current, the excitation current amplitude at each frequency point is set to 100mA, 100mA, 100mA, 95mA, 45mA, 50mA, 85mA, 95mA, 100mA, and 100mA respectively. The amplitude is increased at the frequency points corresponding to the low-frequency quality transmission process to ensure the signal-to-noise ratio, the amplitude is appropriately reduced in the mid-frequency band to reduce the signal peak value, and the amplitude is increased in the high-frequency band to cope with the characteristics of battery amplitude response attenuation.

[0039] (3) The sinusoidal signals at 10 frequency points are combined into a single broadband composite excitation signal, which is applied to the PEMFC once through the electrochemical workstation, and the AC response voltage time domain signal is collected synchronously to complete the electrochemical data acquisition.

[0040] Figure 2 This diagram shows the wideband current excitation signal of the battery under different load conditions. Figure 2 (a) is 0.3 A / cm 2 Under load conditions, Figure 2 (b) is 0.6 A / cm 2 Under load conditions, Figure 2 (c) is 1.2 A / cm 2 Load conditions, such as Figure 2 As shown, the broadband excitation signal synthesized by this method has good load adaptability and can maintain stable excitation characteristics under different operating conditions.

[0041] S2. Acquiring non-invasive magnetic field distribution signals The non-invasive magnetic field distribution signal includes: magnetic field measurement and data matching.

[0042] The magnetic field measurement involves attaching a 4×4 magnetic sensor array to the outer surface of the cathode side of the proton exchange membrane fuel cell, covering the entire active reaction area of ​​the cell, and configuring a fixed sampling frequency to collect magnetic field signals. First, the ambient magnetic field signal under zero load is collected as a background reference, and then the original magnetic field data under the steady-state operation of the cell is collected. The ambient magnetic field interference is eliminated by background subtraction, and finally the effective magnetic field distribution data generated only by the internal current response of the cell is obtained.

[0043] The data matching involves matching each set of electrochemical response signals with the spatial distribution data of the magnetic field under the same operating conditions to form a one-to-one dual-modal diagnostic sample.

[0044] Figure 3 This is a graph showing the distribution of the battery's magnetic field under different load conditions. Figure 3(a)-(d) is 0.3A / cm 2 Under load conditions, Figure 3 (e)-(h) is 0.6 A / cm 2 Under load conditions, Figure 3 (i)-(l) is 1.2A / cm 2 Load conditions, such as Figure 3 As shown, under various load conditions, the magnetic field distribution is most uniform under normal conditions. The water-flooded, dehydrated, and air-starved states all exhibit different spatial distribution patterns, providing a basis for fault differentiation.

[0045] S3. Data preprocessing and sample set construction, including the following steps: (1) Preprocessing the collected dual-mode data: For the electrochemical signal, the frequency response components corresponding to the three processes of proton transport, electrochemical activation and mass transport are separated by bandpass filtering; for the magnetic field signal, the local magnetic gradient amplitude of each sensor node is calculated and the graph structure data is constructed by combining the spatial coordinates.

[0046] (2) Sample set construction: Sufficient independent dual-mode samples are constructed for different load conditions. Electrochemical samples are selected from non-overlapping AC response voltage time series data of fixed duration, and magnetic samples are matched with magnetic field spatial distribution data under steady state of the corresponding operating conditions. All samples are divided into training set and test set according to independent experimental batches according to a fixed ratio. The training set and test set samples are guaranteed to be non-overlapping throughout the process, effectively avoiding data leakage problems.

[0047] S4. Constructing and training an electrochemical-magnetic cross-integrated diagnostic network The electrochemical-magnetic cross-linked diagnostic network (EM-CrossNet) includes: an electrochemical feature extraction module, a GCN magnetic feature extraction module, a gated cross-attention fusion module, and a fault classification output module.

[0048] The ECM-guided electrochemical feature extraction module constructs sub-convolutional networks corresponding to proton transport, electrochemical activation, and mass transport, respectively. The output features of each sub-convolutional network are correlated with the parameters of the equivalent circuit model. The output results of each sub-network are fused to obtain an electrochemical feature vector that can characterize the internal reaction state of the battery.

[0049] The GCN magnetic feature extraction module constructs an undirected graph structure with each magnetic sensor as a node, using sensor measurements, local magnetic gradient magnitudes, and normalized spatial coordinates as initial node features, and builds an adjacency matrix based on the k-nearest neighbor strategy. A three-layer graph convolutional network is used to perform deep aggregation of spatial features, and after global average pooling, a magnetic feature vector f encoding the spatial distribution information of the current is obtained. m This feature has load invariance.

[0050] The gated cross-attention fusion module maps two types of features to a unified semantic space using a trainable projection matrix, generating corresponding feature tensors; it uses a bidirectional cross-attention mechanism to complete information interaction and complementarity, correct feature bias, and distinguish fault types; combined with residual connection and normalization processing, it obtains weight coefficients through an activation function, and outputs comprehensive diagnostic features after adaptive weighted fusion.

[0051] The fault classification output module sequentially inputs the fused features into the fully connected layer and the classifier, and outputs the probability values ​​of four health states of PEMFC: normal, flooded, dehydrated, and starved. The classification confidence threshold is set to 0.85. If the confidence level is lower than the threshold, a secondary diagnosis is triggered to ensure the reliability of the diagnosis results.

[0052] The electrochemical-magnetic cross-fusion diagnostic network is trained using a loss function that incorporates physical constraints from an embedded equivalent circuit model. This loss function is a weighted average of the supervised classification loss and the physical constraint loss. The ECM-guided physical constraint loss is based on ECM response voltage modeling: under broadband excitation, the total response voltage of the ECM model is the sum of the voltage contributions from three electrochemical process modules: proton transport, electrochemical activation, and mass transport. , ,in, V is the total response voltage, p is the electrochemical process, and V is the total response voltage. p For the response voltage, I is the broadband excitation signal, and Z is the response voltage. p Here is the analytical expression for the ECM impedance corresponding to the electrochemical process p. This is the set of fitting parameters for the ECM submodule corresponding to process p.

[0053] The physical constraint loss is constructed for each electrochemical process by building a loss term based on the correlation between the measured signal and the ECM predicted signal. , ,in, Let y be the physical constraint loss component of the p-th electrochemical process, where p is the quantization process. p Vp is the measured response voltage, and Vp is the simulated response voltage of the ECM submodule. For process p, the set of fitting parameters for the corresponding ECM submodule, Loss phys This represents the total loss due to physical constraints.

[0054] S5. Health Management and Maintenance (1) Quantitative grading of health status: The weighted average health score is calculated based on three consecutive diagnostic results. The time weights are set to 0.2, 0.3, and 0.5 in the order of diagnosis. Different weights are assigned to different fault types: 0 for normal status, 0.6 for dehydration fault, 0.8 for flooding fault, and 1.0 for air starvation fault. The comprehensive health score is calculated by combining the fault occurrence frequency correction coefficient. A health score ≥90 is set as excellent, 70≤health score<90 as good, 50≤health score<70 as warning level, and health score<50 as fault level.

[0055] (2) Generation of graded preventive maintenance decisions: For the excellent health status, routine inspections are carried out, including gas pipeline air tightness testing, sensor calibration and fuel cell appearance inspection; for the good health status, parameter calibration is carried out, including humidification system accuracy calibration, gas flow sensor calibration and load response characteristic testing; for the warning health status, partial maintenance is carried out, and corresponding components are checked according to the fault type: for dehydration faults, the focus is on checking the humidification system water circuit and atomizing nozzles; for water flooding faults, the focus is on checking the drain valve and gas diffusion layer permeability; for air starvation faults, the focus is on checking the air compressor and intake filter; for the fault health status, a shutdown overhaul is immediately carried out, the fuel cell is disassembled to check the membrane electrode, bipolar plate and seal condition and replace damaged parts.

[0056] (3) Optimization of operating parameters: For mild faults at the warning level, the operating parameters are adjusted in real time to achieve fault self-healing: for mild flooding faults, the cathode humidification temperature is lowered by 2-5℃, and the anode drainage cycle is shortened to 1 / 2 of the original cycle; for mild dehydration faults, the anode and cathode humidification temperatures are raised by 3-6℃, and the gas metering ratio is increased by 0.2-0.4; for mild air starvation faults, the air metering ratio is increased to 1.8-2.5 times, and the maximum load power is limited to 80% of the rated power. After the health status is restored to above 70, the rated operating parameters are gradually restored.

[0057] Example 3; Performance Verification and Comparison Experiment This embodiment uses the hardware platform described in Embodiment 1 and the diagnostic method described in Embodiment 2 to conduct experimental verification under three load conditions (0.3A / cm², 0.6A / cm², 1.2A / cm²) and four healthy states: normal, flooded, dehydrated, and starved.

[0058] 3.1 Performance Verification of Cross-Load Generalization To verify the cross-load generalization capability of this method, this invention designed six training-test configurations, including single-load training with cross-load testing and dual-load training with single-load testing.

[0059] Configuration 1: Training with a load of 0.3 A / cm², and testing with loads of 0.6 A / cm² and 1.2 A / cm²; Configuration 2: Training with a load of 0.6 A / cm², and testing with loads of 0.3 A / cm² and 1.2 A / cm²; Configuration 3: Training with a load of 1.2 A / cm², and testing with loads of 0.3 A / cm² and 0.6 A / cm²; Configuration 4: Training with loads of 0.6 A / cm² and 1.2 A / cm², and testing with loads of 0.3 A / cm²; Configuration 5: Training with loads of 0.3 A / cm² and 1.2 A / cm², and testing with loads of 0.6 A / cm²; Configuration 6: Training with loads of 0.3 A / cm² and 0.6 A / cm², and testing with loads of 1.2 A / cm².

[0060] Depend on Figure 4 It can be seen that the average diagnostic accuracy of configurations 1-3 can reach over 90%; the average diagnostic accuracy of configurations 4-6 can reach over 94%; and the diagnostic accuracy of configuration 5 (trained with high and low load data and tested with intermediate load data) is the highest, reaching 97%. This shows that even when trained with only a single load data, this method can still maintain an average diagnostic accuracy of about 90%, which is mainly due to the complementary advantages of dual-modal features: the electrochemical features extracted by the ECM-guided network can anchor the mechanism of the fault, and the magnetic spatial features extracted by the GCN have natural load invariance. The combination of the two can weaken the interference of load fluctuations on fault features, thereby alleviating the performance degradation problem of traditional electrochemical methods under cross-load scenarios.

[0061] 3.2 Ablation Experiment Verification Four sets of control models were set up: Model-1 (electrochemical features only), Model-2 (magnetic features only), Model-3 (using dual-modal feature extraction completely consistent with the present method, achieving feature fusion through simple vector concatenation, without cross attention and adaptive gating mechanism), and Model-4 (the complete network of the present invention). Group 1 uses the same load conditions for training and testing, Group 2 is a cross-load scenario of single-load training and dual-load testing, and Group 3 is a cross-load scenario of dual-load training and single-load testing.

[0062] Depend on Figure 5 and Figure 6It can be seen that Model-4 achieved the best performance in all three scenarios: same load, single-load training with cross-load testing, and dual-load training with single-load testing. The accuracy was 98.7% in the same load scenario and 92.4% and 97.8% respectively in the cross-load scenario, reducing the misclassification rate of flooding and air starvation to below 7%. In contrast, Model-1's cross-load performance was severely degraded, making it difficult to adapt to feature distribution shifts caused by load fluctuations; Model-2 maintained stable diagnostic accuracy in cross-load scenarios, but consistently suffered from high misclassification of flooding and air starvation faults, making it difficult to distinguish between these two types of faults with different mechanisms but similar spatial effects; Model-3's performance was between the two but still had significant errors. The ablation results confirm the necessity of dual-modal gating cross-attention fusion.

[0063] 3.3 Model Convergence and Real-time Diagnostic Performance The average total diagnostic time for the end-to-end process of this invention is 7.7s, including 0.52s for magnetic field acquisition (including 0.5s for settling time), 6s for broadband excitation and voltage acquisition (3 times the lowest frequency cycle), and 1.2s for data preprocessing and model inference.

[0064] Figure 7 This is a comparison chart of real-time signal acquisition and feature fidelity analysis. Among them, Figure 7 (a) is a comparison chart of the acquisition time and number of frequency points between traditional electrochemical impedance spectroscopy and the AC voltage sampling method of this invention. Figure 7 (a) It can be seen that compared with the traditional EIS's 33 frequency points and 247s full spectrum scan time, the signal acquisition time of this method is shortened by 97.6% without losing the fidelity of fault characteristics; Figure 7 (b) is a two-dimensional projection visualization of electrochemical features extracted based on electrochemical impedance spectroscopy. Figure 7 (c) is a two-dimensional projection visualization of electrochemical characteristics extracted based on AC voltage according to the present invention, by Figure 7 (b) and Figure 7 (c) It can be seen that this application not only shortens the signal acquisition time, but also the extracted features can achieve the same excellent clustering effect as traditional electrochemical impedance. The feature clusters of the four health states are completely separated without overlap or confusion, which verifies that this application can significantly shorten the acquisition time while completely preserving the fault-related information of the internal electrochemical process of PEMFC, meeting the real-time online diagnosis requirements under dynamic load scenarios.

[0065] Figure 8 This is a comparison chart of the model's fast convergence performance and cross-load diagnostic performance. Among them, Figure 8 (a) is a comparison of the number of training rounds and convergence time between the traditional CNN network and the ECM guided network of this invention. The ECM guided network of this method only needs 17 training rounds to converge, while the traditional CNN network needs 150 rounds, achieving a convergence speed improvement of nearly 10 times. Figure 8 (b) is a schematic diagram of the PEMFC fault diagnosis accuracy of the two diagnostic methods under different load scenarios (Group 1 is training and testing under the same load conditions; Group 2: cross-load scenarios, single load training and dual load testing; Group 3: cross-load scenarios, dual load training and single load testing). The average cross-load accuracy of the model under the physical constraints of this application is 8 to 18 percentage points higher than that of the pure CNN baseline, which not only speeds up the training efficiency but also improves the generalization ability of the model.

[0066] 3.4 Verification of the effectiveness of health management and maintenance This invention further constructs a health management system based on fault diagnosis results. To verify its actual effect, 30 groups of PEMFC single batteries with similar operating conditions were selected and randomly divided into a control group and an experimental group.

[0067] The control group adopted the traditional passive fault repair strategy (i.e., shutdown and repair after a fault occurs), while the experimental group adopted the proactive health management scheme proposed in this invention (implementing differentiated preventive maintenance and optimization of operating parameters based on diagnostic results).

[0068] After 2000 hours of continuous operation, the experimental results showed that the unplanned downtime of the experimental group was reduced by 32.6% compared with the control group, and the system operation and maintenance cost was reduced by 23.1%. At the same time, the performance degradation rate caused by the accumulation of faults in the experimental group was 41.3% lower than that in the control group. The above results indicate that the health management scheme constructed in this invention can effectively avoid sudden downtime accidents, reduce the total life cycle operation and maintenance cost, and significantly improve the operational reliability of the fuel cell power system.

[0069] 3.5 Comparison and Validation with Existing Methods This invention selects four mainstream methods for comparison under the same conditions: transfer learning method based on DC voltage (TL-DCV), transfer learning method based on AC voltage (TL-ACV), transfer learning method based on electrochemical impedance spectroscopy (TL-EIS), and convolutional neural network based on magnetic field distribution (CNN-Magnetic).

[0070] Figure 9 This is a performance comparison chart of different battery fault diagnosis methods. Among them, Figure 9 (a) Diagnostic accuracy under different load scenarios (Group 1: Same load scenario, training and testing under the same load conditions; Group 2: Cross-load scenario, single load training and dual load testing; Group 3: Cross-load scenario, dual load training and single load testing). Figure 9 (b) represents the data acquisition time for different methods, combined with... Figure 9 (a) and Figure 9(b) It can be seen that TL-DCV has a diagnostic accuracy of about 96% under the same load scenario and maintains an accuracy of about 90% under cross-load scenarios, but the acquisition time is as long as 600s; TL-ACV has good accuracy under the same load scenario and the acquisition time is only 6s, but the features it extracts are naturally coupled with the combined effects of load fluctuation and fault state, making it difficult to adapt to the working conditions of load changes under cross-load scenarios, and the accuracy drops significantly; the TL-EIS method has a diagnostic accuracy close to 100% under the same load scenario, but its impedance characteristics are highly sensitive to load fluctuations, and the accuracy drops below 60% when the operating conditions deviate from the training load range, while the traditional EIS full-spectrum scan takes a long time. The 247s time is insufficient for real-time online diagnosis; CNN-Magnetic's data acquisition time is only 0.1s, but its accuracy is below 80% under different load scenarios. The diagnostic accuracy under cross-load scenarios is not significantly lower than that under the same load scenario, but both water flooding and air starvation faults will cause internal local current decay, resulting in highly similar magnetic field spatial distribution patterns, which have obvious misclassification problems. The EM-CrossNet method in this application has the best overall performance. Through the cross-fusion of electrochemical and magnetic dual-mode gating, it not only achieves end-to-end diagnosis in the second level of 7.7s, but also maintains a high accuracy of over 96% under the same load and cross-load scenarios.

[0071] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No markings in the claims should be construed as limiting the scope of the claims.

Claims

1. A real-time fault diagnosis method for proton exchange membrane fuel cells across loads based on electrochemical-magnetic cross-integration, characterized in that, Includes the following steps: Acquire broadband excitation electrochemical response signal and non-invasive magnetic field distribution signal of fuel cell; The broadband excitation electrochemical response signal is frequency domain separated to extract the electrochemical mechanism features corresponding to proton transport, electrochemical activation and mass transport processes, respectively. A graph structure with magnetic sensors as nodes is constructed for the non-invasive magnetic field distribution signal to extract load-invariant spatial features; The electrochemical mechanism features and the load-invariant spatial features are bidirectionally complementary and adaptively weighted by a gated cross-attention mechanism to obtain fused features; The fused features are input into the classifier, which outputs the fault diagnosis results. Based on the output fault diagnosis results, health management is carried out, a health status quantitative assessment and level classification strategy is constructed, maintenance decisions are generated, and health regulation is achieved.

2. The method according to claim 1, characterized in that, The acquisition of the broadband excitation electrochemical response signal includes: based on the characteristic frequencies of three types of electrochemical processes—proton transport, electrochemical activation, and mass transport—iteratively supplementing frequency points using an adaptive frequency selection method guided by the Fisher information matrix to obtain an optimal discrete frequency set; synthesizing the sinusoidal signals of the optimal discrete frequency set into a single broadband composite excitation signal, applying it to the fuel cell once, and simultaneously acquiring the AC response voltage time-domain signal.

3. The method according to claim 1, characterized in that, The acquisition of non-invasive magnetic field distribution signals includes: attaching a magnetic sensor array to the outer surface of the cathode side of the fuel cell, covering the active reaction area of ​​the battery; acquiring the ambient magnetic field under zero load as a background reference, acquiring the original magnetic field data under steady-state operation of the battery, eliminating ambient magnetic field interference through background subtraction, and finally obtaining effective magnetic field distribution data generated only by the internal current response of the battery.

4. The method according to claim 2, characterized in that, The excitation current amplitude at each frequency point in the excitation signal is distributed using an inverse Gaussian distribution, and the excitation current does not exceed 5% of the load current; among which, the amplitude of the low-frequency band is increased, the amplitude of the mid-frequency band is decreased, and the amplitude of the high-frequency band is increased in the corresponding quality transmission process.

5. The method according to claim 1, characterized in that, The extraction of load-invariant spatial features uses the measured values ​​of each magnetic sensor, the local magnetic gradient magnitude, and the normalized spatial coordinates as the initial features of the nodes. An adjacency matrix is ​​constructed based on the k-nearest neighbor strategy. A multi-layer graph convolutional network is used to perform deep aggregation of spatial features on the graph structure data. After global average pooling, the magnetic feature vector is obtained.

6. The method according to claim 1, characterized in that, The extraction of electrochemical mechanism features includes: designing a frequency domain bandpass filter based on the characteristic frequency bands of three types of electrochemical processes, and converting it into a time domain convolution kernel through inverse Fourier transform; convolving the broadband excitation electrochemical response signal with the convolution kernel corresponding to each process to separate the frequency response components of each electrochemical process; configuring a sub-convolutional network for each electrochemical process, extracting process-level features bound to the parameters of the equivalent circuit model, and concatenating them to obtain an electrochemical feature vector.

7. The method according to claim 6, characterized in that, The subconvolutional network includes three subconvolutional networks corresponding to the proton transport process, the electrochemical activation process, and the mass transport process, respectively. Among them, the features extracted by the proton transport sub-network are related to the series resistance in the equivalent circuit model, the features extracted by the electrochemical activation sub-network are related to the charge transfer resistance and the parameters of the constant phase element, and the features extracted by the mass transport sub-network are related to the gas diffusion resistance and the parameters of the RLC parallel circuit.

8. The method according to claim 1, characterized in that, The gated cross-attention mechanism includes: mapping electrochemical mechanism features and load-invariant spatial features to a shared semantic space using a trainable projection matrix; calculating a first cross-attention using electrochemical features as queries and magnetic features as bonds and values, and calculating a second cross-attention using magnetic features as queries and electrochemical features as bonds and values; performing residual connection and layer normalization on the cross-attention output and the original features, and generating a gated vector using the sigmoid function to adaptively weight and fuse the dual-modal features.

9. The method according to claim 1, characterized in that, The fault diagnosis results include probability values ​​for four health states: normal, flooded, dehydrated, and starved. The classification confidence threshold is set to 0.85, and secondary diagnosis is triggered when the score is below the threshold.

10. The method according to claim 1, characterized in that, The health management is as follows: based on the output probability values ​​of four health states (normal, flooded, dehydrated, and starved) and historical data in continuous time series, a quantitative assessment strategy for the health status of fuel cells is constructed, which integrates fault type, frequency of occurrence, and severity of single fault, and a multi-level classification of health status is completed. Based on health status results and fault root cause characteristics, differentiated preventive maintenance decisions are generated, specifying maintenance content, execution cycle, and resource priority; Simultaneously, based on the fault type and health status, the gas metering ratio, humidification temperature, and operating temperature parameters of the battery are optimized in real time to achieve self-healing of minor faults and regulation of health status.

Images