Anion exchange membrane electrolysis system early abnormality diagnosis method and system based on fault map learning

Abstract

The invention provides a fault map learning-based early abnormality diagnosis method and system for an anion exchange membrane electrolysis system, and relates to the technical field of fault diagnosis. The method comprises the following steps of: acquiring multi-working-condition operation measurement and constructing a time sequence characteristic fragment, establishing a fault map containing parts, working conditions and measurement nodes by combining a part relationship and working condition dependency, training and calibrating a map learning model on the fault map, and extracting a self-adaptive baseline; and projecting the feature fragments to a fault map, outputting multi-level risk indexes, generating a comprehensive risk score and an anomaly candidate set, and implementing multi-scale threshold judgment and stability test in combination with a baseline to form early warning and disposal suggestions. And performing further attribution analysis on the abnormal candidate set, identifying key components, and reinjecting maintenance feedback to update the atlas and the model to form a continuously optimized knowledge base. According to the method, causal correlation modeling and interpretable diagnosis are realized, and the accuracy and timeliness of early abnormality identification of the AEM electrolysis system are effectively improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent fault diagnosis, in particular to an early abnormality diagnosis method and system for an anion exchange membrane electrolysis system based on fault atlas learning. BACKGROUND

[0002] Water electrolysis hydrogen production equipment is rapidly applied in renewable energy consumption and green chemical raw material supply. Anion exchange membrane electrolysis systems are concerned due to the advantages of using alkaline electrolyte and low catalyst cost. Existing engineering practices generally configure stack voltage, current, temperature, pressure, flow and gas purity measurement, and realize alarm through experience threshold, statistical control chart or single variable rule. Some systems introduce equivalent impedance measurement, data-driven regression or classifier to evaluate the health state. For serious abnormalities that have occurred, such as pump failure, leakage and obvious polarization rise, the traditional method can give an alarm and guide shutdown inspection.

[0003] With the expansion of electrolysis scale and the grid connection of renewable power fluctuation, the AEM electrolysis system emphasizes the timely identification of early, weak signal and complex coupled abnormalities to reduce unplanned shutdown and irreversible damage to the stack. Multi-source sensor fusion, mechanism-data hybrid modeling with causal prior embedding, online self-calibration model and explainable traceability analysis have become the mainstream direction. At the same time, the general diagnosis expression and knowledge sedimentation mechanism across devices and scenarios are evolving from single-device rule sets to transferable and incremental knowledge graphs and graph computing paradigm.

[0004] The existing diagnosis methods with single variable threshold or black box model as the core are difficult to represent the physical coupling and control coupling between components at the same time, difficult to amplify early degradation signs in the weak signal stage stably, and lack the ability to continuously inject historical maintenance knowledge and online new evidence. Statistical models are prone to mismatch under working condition migration, environmental change and sensor drift; pure mechanism threshold is prone to false alarm under complex load and water quality disturbance. The lack of a unified carrier to organize the multi-layer relationship of "component-working condition-measurement-cause prior" makes it difficult to balance the diagnosis accuracy, generalization ability and explainability, resulting in late early abnormality identification, conservative maintenance strategy and high operating cost. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide an early abnormality diagnosis method and system for an anion exchange membrane electrolysis system based on fault atlas learning, which solves the problem that the existing AEM electrolysis system diagnosis method lacks a unified causal correlation modeling and explainable inference mechanism, resulting in the inability to accurately identify early abnormalities in a timely manner.

[0006] To achieve the above purpose, the present application provides the following solutions:

[0007] An early anomaly diagnosis method for anion exchange membrane electrolysis systems based on fault graph learning includes:

[0008] The AEM electrolysis system was collected under different operating conditions. The time reference was unified, noise was removed and missing data was filled in. The data were then aggregated according to a sliding time window to form feature segments for diagnosis, so as to obtain time-series feature segments.

[0009] Based on the temporal feature segments, the component relationships and operating condition dependencies are combined, and a graph structure is established with component nodes, operating condition nodes and measurement nodes. The edges and edge weights are set according to physical coupling and causal priors, and a fault map is obtained.

[0010] Define and initialize a graph learning model on the fault map, and perform supervised training on the initialized graph learning model using historical labeled event fragments to obtain a trained model;

[0011] Based on the training model and using the time-series feature segments for online small-step calibration and extraction of baseline parameters for different working conditions and scenarios, a trained and calibrated graph learning model and adaptive baseline were obtained.

[0012] The temporal feature segments are projected onto the fault map and the graph learning model outputs node-level, edge-level and subgraph-level risk indicators to obtain risk scores and anomaly candidate sets.

[0013] Based on the risk score and the adaptive baseline, multi-scale threshold determination and stability test are performed to generate early anomaly warnings and corresponding handling suggestions, and early warning results and diagnostic conclusions are obtained.

[0014] Based on the diagnostic conclusions, attribution analysis is performed on the abnormal candidate set, outputting key components, key measurements, and key paths. The maintenance feedback and attribution results are then back-injected into the fault map and the trained and calibrated graph learning model to update parameters and priors, resulting in an updated knowledge base and an optimized fault map and graph learning model.

[0015] Preferably, the acquisition of operational measurements of the AEM electrolysis system under different operating conditions involves unifying the time reference, removing noise, and completing missing data for the operational measurements. These measurements are then aggregated according to a sliding time window to form feature segments for diagnostic purposes, resulting in time-series feature segments, including:

[0016] Voltage, current, temperature, pressure and flow were collected from multiple sources under different operating conditions of the electrolysis system to obtain raw measurement data;

[0017] The original measurement data was subjected to time standardization, noise removal, and missing data completion to obtain the processed data;

[0018] Based on the processed data, aggregation and statistical feature extraction were performed according to a sliding time window to obtain time-series feature segments.

[0019] Preferably, the step of establishing a graph structure based on the temporal feature segments, component relationships, and operating condition dependencies, using component nodes, operating condition nodes, and measurement nodes, and setting edges and edge weights based on physical coupling and causal priors, yields a fault map, including:

[0020] Based on the aforementioned time-series feature segments, components of the fuel cell stack, circulation loop, pump, heat exchanger, and separator system were identified, resulting in a component set.

[0021] Threshold determination and interval division are performed on the current, power, start-stop signals and water quality parameters in the time-series feature segments, and they are labeled as operating stages, load intervals and water quality levels to obtain operating condition status labels.

[0022] Based on the operating condition status labels, different operating condition elements were abstracted to obtain the operating condition set;

[0023] By mapping the key measurements in the time-series feature segments to measurement nodes, a measurement set is obtained.

[0024] An initial graph structure is established using the aforementioned component set, operating condition set, and measurement set as node sets;

[0025] Edges are generated on the initial graph structure based on physical coupling relationships and causal priors, and directions and weights are assigned to them, thus obtaining a fault map.

[0026] Preferably, the step of generating edges on the initial graph structure based on physical coupling relationships and causal priors, and assigning directions and weights to obtain a fault map, includes:

[0027] Based on the node set of the initial graph structure, according to the process connectivity and energy transfer paths between the fuel cell stack, pump, heat exchanger, separator and pipeline, node pairs with coupling were screened to obtain a candidate edge set;

[0028] Based on the node pairs in the candidate edge set, and combined with the causal priors of material flow and heat transfer, the direction of each candidate edge is determined, thus obtaining the directed candidate edge set.

[0029] Under different operating conditions, the temporal feature segments are used to evaluate the strength and determine the lag of the directed candidate edge set, calculate the correlation between nodes and the order of responses, and obtain the edge weights and lag parameters.

[0030] The edge weights and hysteresis parameters were subjected to consistency checks and normalization, and edges with low confidence were removed, resulting in a sparse set of directed weighted edges.

[0031] The sparsed set of directed weighted edges is incorporated into the initial graph structure to obtain the fault graph.

[0032] The expressions for the edge weights and hysteresis parameters are as follows:

[0033]

[0034] Among them, w ij Let τ be the weight of the directed edge from node i to node j; ij s is the time lag parameter of the directed edge; ij The cross-condition consistency coefficient represents the proportion of the directed relationship that holds stably under different load and water quality scenarios and is normalized to between zero and one. The non-negative part of the partial correlation coefficient between node i and node j under zero hysteresis is normalized from zero to one and used to characterize the instantaneous coupling strength. The dimensionless value of the transmission bucket from node i to node j after being normalized from zero to one is used to characterize the intensity of the directed causal information flow. The non-negative measure of the directional cross-correlation of node i relative to node j under lag τ, normalized from zero to one. The dimensionless value of the conditional transfer entropy from node i to node j under lag τ, normalized from zero to one. τ represents the set of lag values ​​to be used in the calculation; i and j are the different node indices in the graph; τ is the candidate lag value; the superscript + indicates taking the non-negative part.

[0035] Preferably, the step of performing online small-step calibration based on the trained model and utilizing the temporal feature segments to extract baseline parameters for different operating scenarios, resulting in a trained and calibrated graph learning model and adaptive baseline, includes:

[0036] Based on the trained model and the time-series feature segments of continuous arrival, an online sliding window is constructed according to a preset duration and stride to filter out abnormal segments and form a sample set that can be used for calibration, thus obtaining an online data window;

[0037] The load range, start / stop status, and water quality parameters in the online data window are determined and grouped to form a scenario division for parameter updates and baseline statistics, resulting in a scenario set;

[0038] Without changing the overall model structure, a subset of trainable parameters related to fault map inference is selected, small step size constraints and parameter drift constraints are applied, and incremental optimization is performed using the online data window as input to obtain an intermediate model in calibration.

[0039] Based on the aforementioned scenario set, the intermediate model in the calibration process is subjected to stability and generalization checks. If it passes the checks, it is confirmed, and a verified calibration model is obtained.

[0040] Under the aforementioned scenario set, robust statistics and extreme value truncation are performed on key measurements and risk scores based on the online data window to generate thresholds, rate of change thresholds, and continuous duration thresholds for each scenario, thereby obtaining an adaptive baseline.

[0041] The verified calibration model and the adaptive baseline are jointly solidified and used as the current effective configuration to obtain a trained and calibrated graph learning model and an adaptive baseline.

[0042] The expression for the trained and calibrated graph learning model is:

[0043]

[0044] in, The risk score output by the trained and calibrated graph learning model; Π ★ The parameter set is the result of training and calibration; ε is the set of directed edges in the fault map; κ(Δt) ij ;τ ij ) is the observation time difference Δt ij and hysteresis parameter τ ij The input time decay kernel function is used to characterize the dynamic effects of the edge; Δt ij τ is the time difference between node i and node j under the current observation; ij The time lag parameter for the edge; For parameter set Π * The feature representation mapping of node i; x i and x j These are the input features of the temporal feature segments at nodes i and j, respectively; For parameter set Π ★ The discriminant function is used to generate a risk score by integrating neighbor information and target node features; B is an adaptive baseline obtained from the statistics of the operating scenario, used for the calibration and stabilization of the risk score.

[0045] Preferably, the step of projecting temporal feature segments onto the fault map and having the graph learning model output node-level, edge-level, and subgraph-level risk indicators to obtain risk scores and anomaly candidate sets includes:

[0046] The temporal feature segments were aligned by node timestamps and associated with operating condition status labels to achieve unit unification and range standardization, thus obtaining the node input set.

[0047] Based on the directed weighted edges in the fault graph and the time lag parameter corresponding to each edge, the node input set is rearranged by time lag and sliced ​​by window to construct the edge-level messages and propagation sequence from the source node to the target node, thus obtaining the edge-level message set.

[0048] Using the node input set and the edge message set as input, a trained and calibrated graph learning model is invoked for forward inference, outputting node-level risk, edge-level risk and subgraph-level risk respectively, thus obtaining multi-level risk indicators;

[0049] Based on the adaptive baseline, the multi-level risk indicators are subjected to contextualized calibration and multi-scale smoothing, and aggregated into a single score according to the set rules to obtain a comprehensive risk score.

[0050] The comprehensive risk score was subjected to continuity threshold judgment, spatial consistency constraint and cross-scale consistency test, and nodes, edges and subgraphs that met the abnormal conditions were screened to obtain an abnormal candidate set;

[0051] The comprehensive risk score is bound to and archived with the anomaly candidate set to obtain the risk score and the anomaly candidate set.

[0052] Preferably, multi-scale threshold determination and stability testing are performed based on the risk score and the adaptive baseline to generate early anomaly alerts and corresponding handling suggestions, resulting in early warning results and diagnostic conclusions, including:

[0053] By comparing the comprehensive risk score with the adaptive baseline under the corresponding scenario, the proportion of exceeding the threshold and the trend of change were calculated according to the short-term window, medium-term window and long-term window respectively, and the multi-scale judgment results were obtained.

[0054] The multi-scale determination results were subjected to continuity checks and fluctuation suppression, requiring them to continuously exceed the baseline and maintain a monotonically increasing trend within a set time period, thus obtaining stability verification results.

[0055] Based on the stability verification results, an early anomaly alarm is triggered, and the handling strategy library is automatically matched in combination with the component category where the abnormal node, edge and subgraph are located, generating at least one handling suggestion among load reduction, flushing, switching working mode or shutdown inspection, thus obtaining the early anomaly alarm and corresponding handling suggestion;

[0056] The early anomaly alerts and corresponding handling suggestions were archived and associated with risk scores and anomaly candidate sets to obtain early warning results and diagnostic conclusions.

[0057] Preferably, the step of performing attribution analysis on the candidate anomalies based on the diagnostic conclusion, outputting key components, key measurements, and key paths, and then backfeeding the maintenance feedback and attribution results into the fault map and the trained and calibrated graph learning model to update parameters and priors, results in an updated knowledge base and an optimized fault map and graph learning model, including:

[0058] Based on the diagnostic conclusions and the anomaly candidate set, the involved nodes, edges and subgraphs were narrowed down and associated with working conditions. The attribution objects were divided according to component category and working condition scenario, resulting in the attribution object set.

[0059] The node contribution, edge contribution, and critical path contribution are calculated for the set of attribution objects. Taking into account edge weights, time lag, risk score increment, and spatial consistency, an attribution list sorted by importance is generated, resulting in an attribution candidate list.

[0060] Consistency checks and causal constraint tests were performed on the attribution candidate list to eliminate spurious correlations and unstable terms, resulting in key components, key measurements, and key paths.

[0061] The key components, key measurements and key paths are used to generate handling verification tasks and trigger the maintenance process. Operational data before and after handling, alarm clearing status and actual fault confirmation results are collected to obtain maintenance feedback.

[0062] Based on the maintenance feedback, the fault map is updated by back-injection, and the trained and calibrated graph learning model is updated with small-step incremental updates and sample expansion, resulting in an updated knowledge base and a fault map and graph learning model optimized by back-injection.

[0063] Preferably, the step of back-injecting and updating the fault map based on the maintenance feedback, and simultaneously performing small-step incremental updates and sample expansion on the trained and calibrated graph learning model, yields an updated knowledge base and a back-injected optimized fault map and graph learning model, including:

[0064] By comparing the actual fault confirmation, measurement changes before and after handling, and anomaly disappearance in the maintenance feedback with the anomaly candidate set, the validity of the causal relationship was determined, and the back-injection basis dataset was obtained.

[0065] Based on the aforementioned back-injection dataset, the edge direction, edge weight, and time lag in the fault graph are corrected, edges that are no longer valid are downweighted or removed, and the weights of confirmed causal paths are enhanced, resulting in an updated fault graph.

[0066] Using the aforementioned back-injection dataset as new samples, a small-step incremental update is performed on the trained and calibrated graph learning model to constrain the magnitude of parameter changes and maintain stability with historical knowledge, thus obtaining an intermediate model in calibration.

[0067] Cross-scenario validation and robustness checks were performed on the intermediate model in the calibration. Updates that failed the checks were rolled back to the previous version, while updates that passed were solidified, resulting in a graph learning model optimized by back-injection.

[0068] The updated fault map and the back-injection optimized graph learning model are archived together with the maintenance feedback to form a knowledge base that can be continuously accessed and traced, thus obtaining the updated knowledge base and the back-injection optimized fault map and graph learning model.

[0069] An early anomaly diagnosis system for anion exchange membrane electrolysis systems based on fault graph learning includes:

[0070] The data processing module is used to collect the operating measurements of the anion exchange membrane electrolysis system under different operating conditions, unify the time reference, remove noise and fill in missing data for the operating measurements, and aggregate them according to the sliding time window to form feature segments for diagnosis, so as to obtain time-series feature segments.

[0071] The graph construction module is used to establish a graph structure with component nodes, operating condition nodes and measurement nodes based on the temporal feature fragments, combined with component relationships and operating condition dependencies, and to set edges and edge weights based on physical coupling and causal priors to obtain the fault graph.

[0072] The model training module is used to define and initialize a graph learning model on the fault map, and to perform supervised training on the initialized graph learning model using historical labeled event fragments to obtain a trained model.

[0073] The model calibration module is used to perform online small-step calibration based on the trained model and using the time-series feature segments, and to extract baseline parameters for different working conditions and scenarios, thereby obtaining a trained and calibrated graph learning model and an adaptive baseline.

[0074] The risk inference module is used to project the time-series feature segments onto the fault map, and the graph learning model outputs node-level, edge-level and subgraph-level risk indicators to obtain risk scores and anomaly candidate sets.

[0075] The early warning judgment module is used to perform multi-scale threshold judgment and stability test based on the risk score and the adaptive baseline, generate early anomaly alarms and corresponding handling suggestions, and obtain early warning results and diagnostic conclusions.

[0076] The attribution back-injection module is used to perform attribution analysis on the abnormal candidate set based on the diagnostic conclusion, output key components, key measurements and key paths, and back-inject the maintenance feedback and attribution results into the fault map and the trained and calibrated graph learning model to update parameters and priors, thereby obtaining an updated knowledge base and a fault map and graph learning model optimized by back-injection.

[0077] The present invention discloses the following technical effects:

[0078] This invention provides a method and system for early anomaly diagnosis of anion exchange membrane electrolysis systems based on fault graph learning. The method includes: collecting operational measurements of the AEM electrolysis system under different operating conditions; unifying the time reference, removing noise, and completing missing data for the operational measurements; aggregating the measurements according to a sliding time window to form feature segments for diagnosis, thereby obtaining time-series feature segments; establishing a graph structure based on the time-series feature segments, combined with component relationships and operating condition dependencies, using component nodes, operating condition nodes, and measurement nodes; setting edges and edge weights based on physical coupling and causal priors to obtain a fault graph; defining and initializing a graph learning model on the fault graph; performing supervised training on the initialized graph learning model using historical labeled event segments to obtain a training model; and performing online fault diagnosis based on the training model and the time-series feature segments. By performing small-step calibration and extracting baseline parameters for different operating scenarios, a trained and calibrated graph learning model and adaptive baseline were obtained. Temporal feature segments were projected onto the fault map, and the graph learning model output node-level, edge-level, and subgraph-level risk indicators, resulting in risk scores and anomaly candidate sets. Based on the risk scores and the adaptive baseline, multi-scale threshold determination and stability testing were performed to generate early anomaly warnings and corresponding handling suggestions, resulting in warning results and diagnostic conclusions. Based on the diagnostic conclusions, attribution analysis was performed on the anomaly candidate sets, outputting key components, key measurements, and key paths. Maintenance feedback and attribution results were then back-injected into the fault map and the trained and calibrated graph learning model to update parameters and priors, resulting in an updated knowledge base and an optimized fault map and graph learning model. This invention constructs a fault map integrating components, operating conditions, and measurements, and combines graph learning model training, online small-step calibration, and adaptive baseline determination to achieve dynamic inference of multi-level risk indicators and early anomaly identification. Compared with existing schemes that rely on single thresholds or black-box models, it can overcome the technical problems of weak signal difficulty in identification, high false alarm rate, and lack of interpretability. Thus, it can achieve accurate early warning, critical path tracing, and knowledge closed-loop updating of anion exchange membrane electrolysis systems under complex operating conditions, effectively improving the timeliness, stability, and scalability of diagnosis. Attached Figure Description

[0079] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] Figure 1 A flowchart of an early anomaly diagnosis method for anion exchange membrane electrolysis system based on fault map learning is provided in an embodiment of the present invention.

[0081] Figure 2 This is a schematic diagram of an early anomaly diagnosis system for anion exchange membrane electrolysis system based on fault map learning, provided as an embodiment of the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] like Figure 1 As shown, this invention provides an early anomaly diagnosis method for anion exchange membrane electrolysis systems based on fault map learning, comprising:

[0085] Step 100: Collect the operating measurements of the AEM electrolysis system under different operating conditions, unify the time base, remove noise and complete missing data for the operating measurements, and aggregate them according to the sliding time window to form feature segments for diagnosis, so as to obtain time-series feature segments.

[0086] Step 200: Based on the temporal feature fragments, the component relationships and operating condition dependencies are combined, and a graph structure is established with component nodes, operating condition nodes and measurement nodes. The edges and edge weights are set according to physical coupling and causal priors to obtain the fault map.

[0087] Step 300: Define and initialize a graph learning model on the fault map, and perform supervised training on the initialized graph learning model using historical labeled event fragments to obtain a trained model;

[0088] Step 400: Based on the trained model and using the temporal feature segments, online small-step calibration is performed and baseline parameters for different working conditions are extracted, resulting in a trained and calibrated graph learning model and adaptive baseline.

[0089] Step 500: Project the temporal feature fragments onto the fault map and output node-level, edge-level and subgraph-level risk indicators by the graph learning model to obtain risk scores and anomaly candidate sets;

[0090] Step 600: Perform multi-scale threshold determination and stability test based on the risk score and the adaptive baseline to generate early anomaly alarms and corresponding handling suggestions, and obtain early warning results and diagnostic conclusions;

[0091] Step 700: Perform attribution analysis on the candidate set of anomalies based on the diagnostic conclusion, output key components, key measurements and key paths, and backfeed the maintenance feedback and attribution results to the fault map and the trained and calibrated graph learning model to update parameters and priors, thereby obtaining the updated knowledge base and the backfeed optimized fault map and graph learning model.

[0092] Furthermore, the specific implementation process of step 100 is as follows:

[0093] This embodiment collects multi-source measurement data under different operating conditions of the electrolysis system. The measured objects include key operating parameters such as voltage, current, temperature, pressure, and flow rate. The measurements cover typical operating conditions such as startup, steady state, variable load, and flushing to ensure that the collected data can characterize the operating characteristics of the fuel cell stack, circulation loop, and control unit under different states.

[0094] This embodiment performs unified time base processing on the collected raw measurement data, aligning measurements from different sampling frequencies and timestamps to a unified time axis to ensure temporal consistency across sensors. Based on this, outliers and noise signals are removed, abnormal abrupt changes caused by sensor drift or transient interference are eliminated, and missing measurement points are appropriately supplemented to ensure the continuity and availability of the data sequence.

[0095] This embodiment further employs a sliding time window to construct features based on the processed data. Specifically, the data is aggregated within a preset time window, and statistical and rate-of-change features within the window are extracted to form a time-series feature segment for diagnosis. This feature segment not only preserves the temporal evolution of the operational measurements but also reflects the dynamic trends of key parameters at different window scales.

[0096] Furthermore, the specific implementation process of step 200 is as follows:

[0097] This embodiment first identifies the core components of the electrolysis system based on time-series feature segments, including the fuel cell stack, circulation loop, pump, heat exchanger, and separator. By analyzing the correspondence between measured signals and component operating states, the functional roles of different components under specific operating conditions are determined, and each component is abstracted as a component node in the graph, forming a set of components that characterizes the physical composition of the electrolysis process. This step ensures that the graph structure has the ability to physically map the actual device at the node level.

[0098] This embodiment further determines the operating conditions of key operating signals in the time-series feature segments. These signals include current, power, start / stop status, and water quality parameters. By setting reasonable interval thresholds, these signals are labeled as operating stages, load intervals, and water quality levels, and different operating condition elements are abstracted accordingly to form an operating condition set. In this way, the graph structure can not only reflect the physical relationships between components but also reflect the differences in operating environment and conditions, thereby realizing joint modeling of the component layer and the operating condition layer.

[0099] In this embodiment, the selected key measurements are used as measurement nodes, which, together with the component set and the operating condition set, constitute a complete node set, establishing an initial graph structure. Based on this, directed edges are set according to physical coupling relationships such as material transport, energy coupling, and control logic, and combined with causal priors, assigning directionality and strength weights to the connections between nodes. Through the above process, a fault map is obtained that simultaneously reflects component structure, operating condition status, and measurement characteristics, providing a unified modeling carrier for subsequent graph learning inference and anomaly diagnosis.

[0100] Specifically, after determining the nodes of the initial graph structure, this embodiment first screens node pairs with physical coupling potential based on the process connectivity between the fuel cell stack and pumps, heat exchangers, separators, and pipelines, the actual flow direction of the medium, the heat transfer path, and the control execution chain, forming a candidate edge set. Then, based on the unidirectional prior of material flow and heat transfer, the candidate edges are assigned directions, and the strength and time lag of each directed edge are estimated under different operating scenarios based on time sequence feature segments. Specifically, in each operating scenario, the degree of synchronous change between the source node and the target node is robustly evaluated to obtain an instantaneous coupling strength index.

[0101] The directional correlation between the source node and the target node within a set of alternative lag ranges is examined to identify the lead-lag relationship where the source precedes the target. At the same time, the information transmission intensity under the same lag is calculated to characterize the direction and magnitude of the causal information flow, and the above three quantities are normalized to zero to one in each scenario. In order to suppress occasional fluctuations in individual scenarios, a cross-scenario consistency coefficient is also calculated, which represents the proportion of the directional relationship that is stably established in multiple scenarios and is mapped to zero to one.

[0102] After obtaining the above metrics, the strength of the edge is determined by the cross-scenario consistency coefficient, the instantaneous coupling strength, and the causal information flow strength. Any negative contribution is treated as zero to avoid directional conflict. The time lag parameter is obtained by weighting the directional relevance and the conditional information transmission strength within the candidate lag range, so that the lag values ​​with high frequency and strong causality contribute more.

[0103] After the estimation is completed, the strength and hysteresis of all directed edges under different scenarios are tested for consistency and normalized, including cross-scenario repeatability test, confidence assessment and extreme value truncation. Edges with low confidence or unstable in key scenarios are removed according to the preset sparsification criteria. Finally, the sparsified set of directed weighted edges is incorporated into the initial graph structure to form a fault map that can simultaneously characterize the relationship between the component layer, the operating condition layer and the measurement layer.

[0104] Among them, the cross-scenario consistency coefficient is derived from the proportion of directed relationships that pass the test in multiple scenarios and mapped in combination with strength stability; the instantaneous coupling strength is derived from the robust correlation measure under zero lag conditions and mapped to zero to one, which is used to reflect the degree of synchronous coupling; the causal information flow strength is derived from the normalized result of the information transmission measure, which is used to reflect the strength of directed causal effects; the directional correlation is a non-negative directional correlation measure under different lag values ​​and normalized to zero to one, which is used to reflect the degree of matching between lead and lag; the conditional information transmission strength is an information transmission measure based on conditional constraints under different lag values ​​and normalized to zero to one, which is used to suppress spurious correlations caused by co-driving.

[0105] The candidate range for the hysteresis value is set according to the equipment control cycle, fluid inertia and thermal process characteristics, such as covering several sampling window lengths. Usually, the lower limit is selected to be no less than the minimum response time of the control loop, and the upper limit is no more than the main process time constant in the operating scenario.

[0106] Furthermore, the specific implementation process of steps 300-400 is as follows:

[0107] In this embodiment, after initializing the training model, to prevent rapid model drift during actual operation and to maintain effectiveness despite changes in operating conditions, an online sliding window is first constructed based on the training model and continuously arriving temporal feature segments. The sliding window has a preset length and sliding step. This embodiment performs anomaly filtering within the window, removing unreasonable or abrupt data segments and retaining only reliable data segments for subsequent calibration, thereby forming an online data window that provides a stable data source for incremental updates.

[0108] After obtaining the online data window, this embodiment determines and groups the load range, start / stop status, and water quality parameters contained within the window, establishing a scenario division corresponding to different operating conditions. Under this division, this embodiment selects a subset of parameters closely related to fault spectrum inference as trainable objects. Without changing the overall structure of the original model, small step size constraints and parameter drift constraints are applied to gradually fine-tune the model. This constraint mechanism can effectively limit the parameter update amplitude, avoid significant deviations in model performance due to short-term anomalies, and obtain an intermediate model under calibration during continuous optimization.

[0109] This embodiment further performs stability and generalization checks on the intermediate model during calibration based on the aforementioned operating scenario division. The model is confirmed when it exhibits consistent inference ability across all scenarios, thus obtaining a validated calibration model. Simultaneously, this embodiment uses an online data window as a statistical basis to perform robust statistics and extreme value truncation on key measurements and risk scores, extracting risk thresholds, rate of change thresholds, and continuous duration thresholds for each operating scenario to form an adaptive baseline. Finally, this embodiment jointly solidifies the validated calibration model and the adaptive baseline as the currently effective configuration, thereby obtaining a trained and calibrated graph learning model and adaptive baseline.

[0110] Specifically, in this embodiment, forward inference is performed once within each time window using the fault map as a carrier: First, the temporal feature fragments of the target node and its neighboring nodes are read. Neighboring messages are organized based on the edge direction, edge strength, and time lag recorded in the directed edge set, and a time decay coefficient is calculated for each edge. The time decay coefficient is a weight between zero and one obtained based on the time difference of the current observation and the lag difference of the edge. The closer the time difference is, the larger the lag coefficient is, and the larger the gap is, the smaller the coefficient is, thereby emphasizing the influence of alignment with the physical response time. Then, the input features of each neighboring node are sent to the node representation mapping determined by the trained and calibrated parameter set to obtain a fixed-dimensional node representation. Finally, the neighborhood representation of the target node is weighted and aggregated by edge strength and time decay coefficient.

[0111] Based on this, the discriminant function receives the neighborhood representation and the target node's own features, and combines them with an adaptive baseline to give a risk score. The adaptive baseline is derived from long-term statistics of different scenarios and is used to calibrate the location and scale of risk distribution under different load, start-up and shutdown and water quality conditions, so that the risk scores remain comparable when the operating conditions change. The risk score is the abnormality quantification result output by this embodiment, which is used for subsequent threshold determination and early warning. The trained and calibrated parameter set comes from offline supervised training and online small step calibration, and includes the node representation mapping and the weights and regularization coefficients of the discriminant function. Its update range is limited to avoid drift.

[0112] The set of directed edges originates from the graph construction steps, recording the direction, strength, and time lag of each edge; the time decay coefficient ranges from zero to one, derived from the time difference and lag gap mapping, and can adopt robust forms such as piecewise inverse or exponential types; the time difference originates from the timestamp alignment of time-series feature segments, usually in seconds; the time lag originates from the aforementioned lag estimation process, with values ​​falling between the control loop response time and the main process time constant; the node representation mapping is used to compress the original measurements and derived features into a compact representation that can participate in graph inference, which can be achieved through graph convolution or graph attention without limitation;

[0113] The node input features are derived from the data processing and feature construction steps, and typically include voltage, current, temperature, pressure, flow rate and its rate of change, volatility and consistency indicators. The discriminant function is used to integrate the neighborhood representation, target node features and adaptive baseline to output a single risk score. It is trained with labeled events in the offline stage and calibrated with small steps in the online stage. For ease of implementation, the adaptive baseline maintains position and scale parameters according to the scenario. For example, in the scenario of high load and water quality deviation, a higher baseline position and a wider scale are set to reduce false alarms.

[0114] Furthermore, the specific implementation process of step 500 is as follows:

[0115] This embodiment first aligns the time-series feature segments according to the node timestamps to eliminate errors caused by different sampling frequencies and clock offsets, and associates each record with an operating condition status label to achieve dimensional unification and range standardization. For short-term missing measurements, nearest-neighbor interpolation is used, and for extreme outliers, robust truncation is used to ensure that the inputs of each node are comparable on the same time axis. Then, based on the directed weighted edges in the fault map and the time lag parameters corresponding to each edge, the source node data is lag-aligned and rearranged relative to the target node, and divided into continuous segments according to a preset length and stride. The segments simultaneously store key descriptions such as numerical level and rate of change, thereby forming an edge-to-edge propagation sequence for message passing, resulting in the node input set and the edge-level message set.

[0116] In this embodiment, after obtaining the node input set and the edge-level message set, a trained and calibrated graph learning model is invoked to perform forward inference: hierarchical message aggregation is performed along directed weighted edges, and the contribution of an edge is determined by its direction, strength, and temporal proximity after lag alignment; the target node integrates its recent features while aggregating neighborhood messages, and outputs node-level risk; edge-level risk is comprehensively evaluated based on the consistency of features at both ends of the source and target nodes, edge strength, and the order of responses after lag alignment; subgraph-level risk is constrained by physical connectivity or functional dependence, and nodes and edges that are simultaneously high-risk and temporally coherent are merged into connected subgraphs and the overall risk is calculated, ultimately obtaining three types of multi-level risk indicators: node-level, edge-level, and subgraph-level.

[0117] This embodiment further performs scenario-based calibration of multi-level risk indicators based on adaptive baselines. After adjusting the position and scale for different load intervals, start-up and shutdown phases, and water quality levels, it performs smoothing and robust aggregation on short-term, medium-term, and long-term time scales, and obtains a single comprehensive risk score according to preset aggregation rules. Subsequently, a continuity threshold judgment is performed on the comprehensive risk score, requiring that it continuously exceeds the scenario threshold within the minimum confirmation time and has a monotonically increasing trend. At the same time, spatial consistency and cross-scale consistency checks are applied to eliminate occasional noise or local false alarms. Nodes, edges, and connected subgraphs that meet the above conditions are marked as anomaly candidates and archived together with the corresponding comprehensive risk score, time interval, scenario label, and contribution ranking for subsequent early warning output and source tracing analysis, thereby obtaining the risk score and anomaly candidate set.

[0118] Furthermore, the specific implementation process of step 600 is as follows:

[0119] After calculating the risk score and adaptive baseline, this embodiment further conducts multi-scale threshold determination and stability testing. First, the comprehensive risk score is compared with the adaptive baseline under the corresponding scenario. The proportion exceeding the baseline is counted at three time scales: short-term window, medium-term window, and long-term window. The trend of risk score change is calculated to obtain the significance of anomalies at different time levels, thereby obtaining the multi-scale determination result.

[0120] This embodiment performs a stability test based on the multi-scale judgment results. Through continuity checks and fluctuation suppression, it requires that the comprehensive risk score continuously exceeds the baseline threshold within a set confirmation period and maintains a monotonically increasing or trending increase, thereby avoiding false alarms caused by single fluctuations or occasional anomalies. After the above screening, the stability verification result is obtained, which is used to determine whether the risk score has sufficient continuity and reliability.

[0121] This embodiment triggers early anomaly alarms based on stability verification results. Combining the component categories corresponding to the anomaly nodes, edges, and subgraphs, it automatically matches appropriate countermeasures from a pre-established handling strategy library, generating handling suggestions including load reduction operation, flushing operation, switching operating modes, or shutdown inspection. Finally, this embodiment archives and associates the early anomaly alarms and corresponding handling suggestions with risk scores and anomaly candidate sets, forming early warning results and diagnostic conclusions that can be used for traceability and decision-making.

[0122] Furthermore, the specific implementation process of step 700 is as follows:

[0123] In this embodiment, after outputting the diagnostic conclusion, to trace the source of the anomaly, the diagnostic conclusion is first matched with the anomaly candidate set. The scope of the involved nodes, edges, and subgraphs is then narrowed. Furthermore, correlation analysis is performed based on runtime conditions to categorize different types of components and their corresponding environmental scenarios, thereby defining the attribution object set. This approach ensures that the attribution analysis is targeted and complete at both the structural and runtime environment levels.

[0124] This embodiment, based on the attribution object set, further calculates the contribution of each node, edge, and critical path composed of nodes and edges to the overall anomaly. Specifically, by combining edge weights, time lags, changes in risk scores, and spatial consistency relationships among different components, the strength of each object's influence on risk formation is assessed, and an attribution candidate list ranked by importance is generated. Subsequently, consistency checks and causal constraint tests are performed on the attribution candidate list. False correlations and unstable items are eliminated using cross-time window repeated verification and counterfactual comparisons, ultimately yielding reliable critical components, critical measurements, and critical paths.

[0125] This embodiment, after identifying key components, key measurements, and critical paths, triggers the maintenance process as a verification task. Operational data, alarm resolution status, and actual fault confirmation results are collected before and after maintenance to form maintenance feedback. Based on this feedback, the confirmed causal relationships are used as the back-injection dataset to correct and optimize the edge directions, edge weights, and time lags in the fault graph, resulting in an updated fault graph. Simultaneously, the back-injection dataset is used as new samples to perform small-step incremental updates on the graph learning model. Cross-scenario validation and robustness checks ensure the stability of the updates. The validated model is solidified as a back-injection-optimized graph learning model and archived along with the updated fault graph and maintenance feedback, ultimately forming a continuously accessible and traceable knowledge base.

[0126] like Figure 2 As shown, this embodiment also provides an early anomaly diagnosis system for anion exchange membrane electrolysis systems based on fault map learning, including:

[0127] The data processing module is used to collect the operating measurements of the anion exchange membrane electrolysis system under different operating conditions, unify the time reference, remove noise and fill in missing data for the operating measurements, and aggregate them according to the sliding time window to form feature segments for diagnosis, so as to obtain time-series feature segments.

[0128] The graph construction module is used to establish a graph structure with component nodes, operating condition nodes and measurement nodes based on the temporal feature fragments, combined with component relationships and operating condition dependencies, and to set edges and edge weights based on physical coupling and causal priors to obtain the fault graph.

[0129] The model training module is used to define and initialize a graph learning model on the fault map, and to perform supervised training on the initialized graph learning model using historical labeled event fragments to obtain a trained model.

[0130] The model calibration module is used to perform online small-step calibration based on the trained model and using the time-series feature segments, and to extract baseline parameters for different working conditions and scenarios, thereby obtaining a trained and calibrated graph learning model and an adaptive baseline.

[0131] The risk inference module is used to project the time-series feature segments onto the fault map, and the graph learning model outputs node-level, edge-level and subgraph-level risk indicators to obtain risk scores and anomaly candidate sets.

[0132] The early warning judgment module is used to perform multi-scale threshold judgment and stability test based on the risk score and the adaptive baseline, generate early anomaly alarms and corresponding handling suggestions, and obtain early warning results and diagnostic conclusions.

[0133] The attribution back-injection module is used to perform attribution analysis on the abnormal candidate set based on the diagnostic conclusion, output key components, key measurements and key paths, and back-inject the maintenance feedback and attribution results into the fault map and the trained and calibrated graph learning model to update parameters and priors, thereby obtaining an updated knowledge base and a fault map and graph learning model optimized by back-injection.

[0134] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0135] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for early anomaly diagnosis of an anion exchange membrane electrolysis system based on fault graph learning, characterized in that, include: The AEM electrolysis system was collected under different operating conditions. The time reference was unified, noise was removed and missing data was filled in. The data were then aggregated according to a sliding time window to form feature segments for diagnosis, so as to obtain time-series feature segments. Based on the temporal feature segments, the component relationships and operating condition dependencies are combined, and a graph structure is established with component nodes, operating condition nodes and measurement nodes. The edges and edge weights are set according to physical coupling and causal priors, and a fault map is obtained. Define and initialize a graph learning model on the fault map, and perform supervised training on the initialized graph learning model using historical labeled event fragments to obtain a trained model; Based on the training model and using the time-series feature segments for online small-step calibration and extraction of baseline parameters for different working conditions and scenarios, a trained and calibrated graph learning model and adaptive baseline were obtained. The temporal feature segments are projected onto the fault map and the graph learning model outputs node-level, edge-level and subgraph-level risk indicators to obtain risk scores and anomaly candidate sets. Based on the risk score and the adaptive baseline, multi-scale threshold determination and stability test are performed to generate early anomaly warnings and corresponding handling suggestions, and early warning results and diagnostic conclusions are obtained. Based on the diagnostic conclusions, attribution analysis is performed on the abnormal candidate set, outputting key components, key measurements, and key paths. The maintenance feedback and attribution results are then back-injected into the fault map and the trained and calibrated graph learning model to update parameters and priors, resulting in an updated knowledge base and an optimized fault map and graph learning model.

2. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 1, characterized in that, The AEM electrolysis system's operational measurements under different operating conditions are collected. These measurements undergo time-base unification, noise removal, and missing data completion. The measurements are then aggregated using a sliding time window to form diagnostic feature segments, resulting in time-series feature segments, including: Voltage, current, temperature, pressure and flow were collected from multiple sources under different operating conditions of the electrolysis system to obtain raw measurement data; The original measurement data was subjected to time standardization, noise removal, and missing data completion to obtain the processed data; Based on the processed data, aggregation and statistical feature extraction were performed according to a sliding time window to obtain time-series feature segments.

3. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 1, characterized in that, The fault map is obtained by establishing a graph structure with component nodes, operating condition nodes, and measurement nodes based on the temporal feature segments, component relationships, and operating condition dependencies, and setting edges and edge weights according to physical coupling and causal priors. This graph includes: Based on the aforementioned time-series feature segments, components of the fuel cell stack, circulation loop, pump, heat exchanger, and separator system were identified, resulting in a component set. Threshold determination and interval division are performed on the current, power, start-stop signals and water quality parameters in the time-series feature segments, and they are labeled as operating stages, load intervals and water quality levels to obtain operating condition status labels. Based on the operating condition status labels, different operating condition elements were abstracted to obtain the operating condition set; By mapping the key measurements in the time-series feature segments to measurement nodes, a measurement set is obtained. An initial graph structure is established using the aforementioned component set, operating condition set, and measurement set as node sets; Edges are generated on the initial graph structure based on physical coupling relationships and causal priors, and directions and weights are assigned to them, thus obtaining a fault map.

4. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 3, characterized in that, The process of generating edges on the initial graph structure based on physical coupling relationships and causal priors, and assigning directions and weights, yields a fault map, including: Based on the node set of the initial graph structure, according to the process connectivity and energy transfer paths between the fuel cell stack, pump, heat exchanger, separator and pipeline, node pairs with coupling were screened to obtain a candidate edge set; Based on the node pairs in the candidate edge set, and combined with the causal priors of material flow and heat transfer, the direction of each candidate edge is determined, thus obtaining the directed candidate edge set. Under different operating conditions, the temporal feature segments are used to evaluate the strength and determine the lag of the directed candidate edge set, calculate the correlation between nodes and the order of responses, and obtain the edge weights and lag parameters. The edge weights and hysteresis parameters were subjected to consistency checks and normalization, and edges with low confidence were removed, resulting in a sparse set of directed weighted edges. The sparsed set of directed weighted edges is incorporated into the initial graph structure to obtain the fault graph. The expressions for the edge weights and hysteresis parameters are as follows: Among them, w ij Let τ be the weight of the directed edge from node i to node j; ij s is the time lag parameter of the directed edge; ij The cross-condition consistency coefficient represents the proportion of the directed relationship that holds stably under different load and water quality scenarios and is normalized to between zero and one. The non-negative part of the partial correlation coefficient between node i and node j under zero hysteresis is normalized from zero to one and used to characterize the instantaneous coupling strength. The dimensionless value of the transmission bucket from node i to node j after being normalized from zero to one is used to characterize the intensity of the directed causal information flow. The non-negative measure of the directional cross-correlation of node i relative to node j under lag τ, normalized from zero to one. The dimensionless value of the conditional transfer entropy from node i to node j under lag τ, normalized from zero to one. τ represents the set of lag values ​​to be used in the calculation; i and j are the different node indices in the graph; τ is the candidate lag value; the superscript + indicates taking the non-negative part.

5. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 1, characterized in that, The process of performing online small-step calibration based on the trained model and utilizing the temporal feature segments to extract baseline parameters for different operating scenarios yields a trained and calibrated graph learning model and an adaptive baseline, including: Based on the trained model and the time-series feature segments of continuous arrival, an online sliding window is constructed according to a preset duration and stride to filter out abnormal segments and form a sample set that can be used for calibration, thus obtaining an online data window; The load range, start / stop status, and water quality parameters in the online data window are determined and grouped to form a scenario division for parameter updates and baseline statistics, resulting in a scenario set; Without changing the overall model structure, a subset of trainable parameters related to fault map inference is selected, small step size constraints and parameter drift constraints are applied, and incremental optimization is performed using the online data window as input to obtain an intermediate model in calibration. Based on the aforementioned scenario set, the intermediate model in the calibration process is subjected to stability and generalization checks. If it passes the checks, it is confirmed, and a verified calibration model is obtained. Under the aforementioned scenario set, robust statistics and extreme value truncation are performed on key measurements and risk scores based on the online data window to generate thresholds, rate of change thresholds, and continuous duration thresholds for each scenario, thereby obtaining an adaptive baseline. The verified calibration model and the adaptive baseline are jointly solidified and used as the current effective configuration to obtain a trained and calibrated graph learning model and an adaptive baseline. The expression for the trained and calibrated graph learning model is: in, The risk score output by the trained and calibrated graph learning model; Π ★ The parameter set is the result of training and calibration; ε is the set of directed edges in the fault map; κ(Δt) ij ;τ ij ) is the observation time difference Δt ij and hysteresis parameter τ ij The input time decay kernel function is used to characterize the dynamic effects of the edge; Δt ij τ is the time difference between node i and node j under the current observation; ij Φ is the time lag parameter for the edge. Π *(x i ) is based on parameter set Π * The feature representation mapping of node i; x i and x j These are the input features of the temporal feature segments at nodes i and j, respectively; Ψ Π *(·) represents parameters based on the Π set. ★ The discriminant function is used to generate a risk score by integrating neighbor information and target node features; B is an adaptive baseline obtained from the statistics of the operating scenario, used for the calibration and stabilization of the risk score.

6. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 1, characterized in that, The process of projecting temporal feature fragments onto the fault map and having the graph learning model output node-level, edge-level, and subgraph-level risk indicators yields risk scores and anomaly candidate sets, including: The temporal feature segments were aligned by node timestamps and associated with operating condition status labels to achieve unit unification and range standardization, thus obtaining the node input set. Based on the directed weighted edges in the fault graph and the time lag parameter corresponding to each edge, the node input set is rearranged by time lag and sliced ​​by window to construct the edge-level messages and propagation sequence from the source node to the target node, thus obtaining the edge-level message set. Using the node input set and the edge message set as input, a trained and calibrated graph learning model is invoked for forward inference, outputting node-level risk, edge-level risk and subgraph-level risk respectively, thus obtaining multi-level risk indicators; Based on the adaptive baseline, the multi-level risk indicators are subjected to contextual calibration and multi-scale smoothing, and aggregated into a single score according to the set rules to obtain a comprehensive risk score. The comprehensive risk score was subjected to continuity threshold judgment, spatial consistency constraint and cross-scale consistency test, and nodes, edges and subgraphs that met the abnormal conditions were screened to obtain an abnormal candidate set; The comprehensive risk score is bound to and archived with the anomaly candidate set to obtain the risk score and the anomaly candidate set.

7. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 1, characterized in that, Based on the risk score and the adaptive baseline, multi-scale threshold determination and stability testing are performed to generate early anomaly alerts and corresponding handling suggestions, resulting in early warning results and diagnostic conclusions, including: By comparing the comprehensive risk score with the adaptive baseline under the corresponding scenario, the proportion of exceeding the threshold and the trend of change were calculated according to the short-term window, medium-term window and long-term window respectively, and the multi-scale judgment results were obtained. The multi-scale determination results were subjected to continuity checks and fluctuation suppression, requiring them to continuously exceed the baseline and maintain a monotonically increasing trend within a set time period, thus obtaining stability verification results. Based on the stability verification results, an early anomaly alarm is triggered, and the handling strategy library is automatically matched in combination with the component category where the abnormal node, edge and subgraph are located, generating at least one handling suggestion among load reduction, flushing, switching working mode or shutdown inspection, thus obtaining the early anomaly alarm and corresponding handling suggestion; The early anomaly alerts and corresponding handling suggestions were archived and associated with risk scores and anomaly candidate sets to obtain early warning results and diagnostic conclusions.

8. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 1, characterized in that, The process involves performing attribution analysis on the candidate anomalies based on the diagnostic conclusions, outputting key components, key measurements, and key paths, and then backfeeding the maintenance feedback and attribution results into the fault map and the trained and calibrated graph learning model to update parameters and priors. This results in an updated knowledge base and an optimized fault map and graph learning model, including: Based on the diagnostic conclusions and the anomaly candidate set, the involved nodes, edges and subgraphs were narrowed down and associated with working conditions. The attribution objects were divided according to component category and working condition scenario, resulting in the attribution object set. The node contribution, edge contribution, and critical path contribution are calculated for the set of attribution objects. Taking into account edge weights, time lag, risk score increment, and spatial consistency, an attribution list sorted by importance is generated, resulting in an attribution candidate list. Consistency checks and causal constraint tests were performed on the attribution candidate list to eliminate spurious correlations and unstable terms, resulting in key components, key measurements, and key paths. The key components, key measurements and key paths are used to generate handling verification tasks and trigger the maintenance process. Operational data before and after handling, alarm clearing status and actual fault confirmation results are collected to obtain maintenance feedback. Based on the maintenance feedback, the fault map is updated by back-injection, and the trained and calibrated graph learning model is updated with small-step incremental updates and sample expansion, resulting in an updated knowledge base and a fault map and graph learning model optimized by back-injection.

9. The method for early anomaly diagnosis of anion exchange membrane electrolysis system based on fault map learning according to claim 8, characterized in that, The fault map is updated by back-injection based on the maintenance feedback. Simultaneously, the trained and calibrated graph learning model is updated with small-step incremental updates and sample expansion, resulting in an updated knowledge base and a fault map and graph learning model optimized by back-injection, including: By comparing the actual fault confirmation, measurement changes before and after handling, and anomaly disappearance in the maintenance feedback with the anomaly candidate set, the validity of the causal relationship was determined, and the back-injection basis dataset was obtained. Based on the aforementioned back-injection dataset, the edge direction, edge weight, and time lag in the fault graph are corrected, edges that are no longer valid are downweighted or removed, and the weights of confirmed causal paths are enhanced, resulting in an updated fault graph. Using the aforementioned back-injection dataset as new samples, a small-step incremental update is performed on the trained and calibrated graph learning model to constrain the magnitude of parameter changes and maintain stability with historical knowledge, thus obtaining an intermediate model in calibration. Cross-scenario validation and robustness checks were performed on the intermediate model in the calibration. Updates that failed the checks were rolled back to the previous version, while updates that passed were solidified, resulting in a graph learning model optimized by back-injection. The updated fault map and the back-injection optimized graph learning model are archived together with the maintenance feedback to form a knowledge base that can be continuously accessed and traced, thus obtaining the updated knowledge base and the back-injection optimized fault map and graph learning model.

10. An early anomaly diagnosis system for anion exchange membrane electrolysis system based on fault graph learning, characterized in that, include: The data processing module is used to collect the operating measurements of the anion exchange membrane electrolysis system under different operating conditions, unify the time reference, remove noise and fill in missing data for the operating measurements, and aggregate them according to the sliding time window to form feature segments for diagnosis, so as to obtain time-series feature segments. The graph construction module is used to establish a graph structure with component nodes, operating condition nodes and measurement nodes based on the temporal feature fragments, combined with component relationships and operating condition dependencies, and to set edges and edge weights based on physical coupling and causal priors to obtain the fault graph. The model training module is used to define and initialize a graph learning model on the fault map, and to perform supervised training on the initialized graph learning model using historical labeled event fragments to obtain a trained model. The model calibration module is used to perform online small-step calibration based on the trained model and using the time-series feature segments, and to extract baseline parameters for different working conditions and scenarios, thereby obtaining a trained and calibrated graph learning model and an adaptive baseline. The risk inference module is used to project the time-series feature segments onto the fault map, and the graph learning model outputs node-level, edge-level and subgraph-level risk indicators to obtain risk scores and anomaly candidate sets. The early warning judgment module is used to perform multi-scale threshold judgment and stability test based on the risk score and the adaptive baseline, generate early anomaly alarms and corresponding handling suggestions, and obtain early warning results and diagnostic conclusions. The attribution back-injection module is used to perform attribution analysis on the abnormal candidate set based on the diagnostic conclusion, output key components, key measurements and key paths, and back-inject the maintenance feedback and attribution results into the fault map and the trained and calibrated graph learning model to update parameters and priors, thereby obtaining an updated knowledge base and a fault map and graph learning model optimized by back-injection.

Images