CN-122024567-A - Real-time mapping and state prediction method for hydrogen fuel cell teaching device

CN122024567ACN 122024567 ACN122024567 ACN 122024567ACN-122024567-A

Abstract

The invention provides a real-time mapping and state predicting method of a hydrogen fuel cell teaching device, which comprises the steps of synchronously acquiring physical variable data such as voltage, current, temperature, humidity and the like by utilizing a plurality of sensors, denoising and standardizing the data to realize structural time sequence feature extraction, constructing a fault cause-sensing response associated knowledge graph based on expert experience, carrying out state prediction and key factor attribution on observed data by fusing a graph attention network, outputting understandable early warning information to a teaching terminal by combining natural language generation and a visualization module, and dynamically fine-adjusting knowledge graph edge weight after receiving user feedback to realize self-adaptive optimization of interpretation logic.

Inventors

DUAN CHUNYAN
LIU YANG
Lian Jiasheng
HU WENYONG
WU YUEFANG
CHEN XIAOYUE
ZHOU NING
LI PAN
TANG RONG
DENG KAILUN
LIANG SHUIYING

Assignees

佛山职业技术学院

Dates

Publication Date: 20260512
Application Date: 20260119

Claims (10)

1. The method for mapping and predicting the state of the hydrogen fuel cell teaching device in real time is characterized by comprising the following steps: S1, collecting multi-dimensional time sequence data generated in the operation process of a hydrogen fuel cell teaching device, and labeling a corresponding physical variable type and a spatial position label for each sensing channel to form an original sensing data sequence; S2, denoising and normalizing the original sensing data sequence to generate a normalized time sequence feature vector; S3, constructing a knowledge graph based on a typical fault mode and a physical cause mechanism defined in an expert experience library; S4, inputting the standardized time sequence feature vector into a graph attention network model, wherein the graph attention network model takes a node structure of the knowledge graph as a topological basis, maps each physical variable into a corresponding observation node, calculates an influence coefficient of each node on a fuel cell health state prediction result, and outputs a health state prediction value and a corresponding attention distribution vector; s5, carrying out matching analysis on the attention distribution vector and causal edge weights in the knowledge graph, identifying fault related node combinations with highest contribution to the current prediction result, and generating a key influence factor set; s6, driving an explanation generation engine by utilizing the key influence factor set, converting the high contribution degree nodes and the associated paths thereof into trend early warning descriptions of natural language descriptions, and generating a dynamic evolution path diagram by combining a knowledge graph visualization module to form understandable output information; And S7, pushing the understandability output information to a teaching terminal interface, and receiving feedback labels of the user on the interpretation rationality.
2. The method for real-time mapping and status prediction of hydrogen fuel cell teaching device according to claim 1, wherein the step S7 further comprises: And S8, performing online fine adjustment on the edge weight in the knowledge graph based on the feedback label.
3. The method for real-time mapping and status prediction of hydrogen fuel cell teaching device according to claim 1, wherein the step S1 specifically comprises: acquiring real-time output signals of a plurality of types of sensors deployed in the hydrogen fuel cell teaching device, performing synchronous sampling based on a physical interface protocol of each sensing unit, acquiring original analog/digital signals under continuous time sequences, and generating an original sensing signal stream; Performing channel separation and time alignment processing on the original sensing signal flow, aligning the mixed transmission data flow according to a sampling time stamp according to a preset sensor topology configuration table, and distributing the mixed transmission data flow to a corresponding data channel to generate a single-channel time sequence signal set classified according to physical variables; Injecting attribute tags for each single-channel time sequence signal based on a pre-established sensing channel metadata mapping table, packaging tag information by using an XML format and binding the tag information with corresponding time sequence data to generate an original sensing data sequence with semantic labels; Writing the original sensing data sequence with semantic annotation into an edge data buffer area, temporarily storing data fragments of the last N sampling periods by adopting an annular queue structure, triggering time marking service, and adding UTC standard time stamps and local system clock offset for each record to generate a labeled original sensing data sequence with space-time consistency; And executing data integrity verification and abnormal channel detection, calculating the loss rate and out-of-range ratio of the current batch data based on the expected sampling frequency and the numerical dynamic range of each channel, and marking a channel as a state to be diagnosed and generating an alarm log if a certain channel is continuously lost and exceeds a preset period or the out-of-range ratio reaches a preset threshold value.
4. The method for real-time mapping and status prediction of a hydrogen fuel cell teaching device according to claim 3, wherein the multiple types of sensors comprise voltage sensors, current sensors, thermocouple temperature sensors, capacitive humidity sensors, thermal mass gas flow meters and piezoresistive pressure transducers.
5. The method for real-time mapping and status prediction of hydrogen fuel cell teaching device according to claim 1, wherein the step S2 specifically comprises: based on the original sensing data sequence with the tag, time sequence data of voltage, current, temperature, humidity, gas flow and pressure parameters corresponding to each sensing channel are obtained, high-frequency random noise is suppressed, and a preliminary smoothing time sequence signal is generated; Respectively calculating statistical characteristic parameters of each physical variable type in the preliminary smoothing time sequence signal in a training period, and executing Z-score standardization transformation based on the statistical characteristic parameters to generate standardized signal components; carrying out space-time alignment recombination on the standardized signal components according to the spatial position labels, constructing a multivariable synchronous feature matrix based on the topological structure relation of the fuel cell stack body, and generating a joint feature representation; Detecting and correcting abnormal value segments in the joint characteristic representation, positioning data points deviating from a normal fluctuation range by adopting a method for identifying outliers based on a quartile range, and compensating and repairing the data points by utilizing linear interpolation and combining effective data before and after moments to generate a continuous characteristic sequence; And slicing and stacking the continuous feature sequences according to a fixed time window to form a high-dimensional tensor form, and outputting the continuous feature sequences to form a standardized time sequence feature vector.
6. The method for real-time mapping and status prediction of hydrogen fuel cell teaching device according to claim 1, wherein the step S3 specifically comprises: Acquiring a preset typical fault mode set in the hydrogen fuel cell teaching device, extracting core dependent variables of each type of faults based on physical mechanisms and key induction factors corresponding to the faults recorded in an expert experience library, and generating a structural fault cause list; Based on the structured fault cause list, combining the collected and marked physical variable types and the spatial position labels, identifying sensor observation variables associated with each fault cause, mapping each physical variable into an observation node in a knowledge graph, mapping each type of fault type into a state node, and constructing a heterogeneous graph structure frame comprising two types of nodes; aiming at the heterogeneous graph structure framework, establishing a directed edge connection relation from an observation node to a state node based on a physical causal relation path described in an expert experience library, and giving an initial causal strength weight to each edge by using a field prior to generate a causal edge set with weight; carrying out consistency verification and normalization processing on the causal edge set with the weight, verifying the rationality of each causal path and the physical interpretability of weight distribution by adopting a mode of combining an expert scoring method with historical fault case backtracking analysis, carrying out pruning optimization on conflict or redundant edges, calculating a global causal stability coefficient and outputting a standardized knowledge graph structure; And (3) performing persistent storage on the standardized knowledge graph structure in a graph database format, distributing a unique semantic identifier for each node, establishing a node metadata index table, and generating a final version of 'fault cause-sensing response' associated knowledge graph.
7. The method of claim 6, wherein the set of typical failure modes includes membrane dehydration, catalyst poisoning, cathode flooding, proton exchange membrane degradation, and insufficient gas supply.
8. The method for real-time mapping and status prediction of hydrogen fuel cell teaching device according to claim 1, wherein the step S4 specifically comprises: Based on the constructed 'fault cause-sensing response' associated knowledge graph, extracting the connection relation between the node set and the edges, mapping each node into an embeddable graph structure unit, initializing the corresponding node feature vector according to the node type, and forming graph structure input representation; Carrying out time window slicing processing on the standardized time sequence feature vector, extracting a multidimensional sensing data sequence in a sliding window at the current moment, mapping the multidimensional sensing data sequence to a hidden space consistent with the node dimension of the knowledge graph through a linear projection layer, and generating an observation feature embedding vector; based on the graph structure input representation and the observation feature embedded vector, performing multi-head graph attention mechanism calculation, carrying out weighted aggregation on feature similarity between adjacent nodes in each attention head, normalizing by a softmax function to obtain influence weights of all nodes in the local neighborhood, and generating a node update state containing context awareness capability; Splicing and nonlinear transformation processing are carried out on the multi-head attention output result to obtain a final node characterization matrix, output nodes which are strongly related to the health state are selected, a health state predicted value is generated through full-connection layer regression calculation, and meanwhile, the original weight distribution of the last layer of attention head is reserved to form an attention distribution vector; and carrying out normalization and sequencing on the attention distribution vectors, extracting physical variable nodes with contribution degrees of N before ranking and corresponding weight values thereof, and generating a structured key influence factor list.
9. The method for mapping and predicting states in real time of a hydrogen fuel cell teaching device according to claim 8, wherein the graph attention network model is characterized in that a health state predicted value and a node contribution degree attention distribution are output through node feature normalization, topology structure embedding, multi-head attention mechanism and nonlinear activation, nodes with contribution degree ranking of 5-10 are screened, and a high contribution degree key influence factor subset is formed through causal edge weight weighted fusion.
10. The method for real-time mapping and status prediction of hydrogen fuel cell teaching device according to claim 1, wherein the step S5 specifically comprises: Based on each dimension value in the attention distribution vector, calculating the mapping consistency between the attention distribution vector and a corresponding physical variable node in a knowledge graph, and generating a normalized attention weight set; based on the normalized attention weight set, performing a dual-channel weighted fusion operation of attention weights and structured 'fault cause-sensing response' associated knowledge graph edge weights, calculating a composite influence intensity value, and generating a joint contribution degree scoring vector; based on the joint contribution degree scoring vector, a dynamic threshold segmentation algorithm is adopted to determine a significant influence node set, nodes with scores higher than a threshold value and topology connection paths thereof in a knowledge graph are identified, corresponding fault type nodes and upstream sensing variable nodes are extracted, and a high contribution degree node combination set is generated; Performing path tracing analysis on the high-contribution degree node combination set, reversely tracking from a prediction target node to an original sensing input node based on a directed causal edge defined in the knowledge graph, reconstructing a multi-hop reasoning path from observation abnormality to potential fault, packaging each complete causal path into a structured causal path tuple, and generating a candidate interpretation path set; Based on the candidate interpretation path set, the sorting optimization is executed by adopting a path simplicity and semantic integrity dual evaluation criterion, redundant paths and isolated node branches are removed, the most representative causal paths which are not overlapped with each other are reserved, the related nodes and side relations are integrated, and a final key influence factor set is generated.

Description

Real-time mapping and state prediction method for hydrogen fuel cell teaching device Technical Field The invention relates to the technical field of fuel cell state monitoring and intelligent fault prediction, in particular to a method for mapping and predicting states of a hydrogen fuel cell teaching device in real time. Background The hydrogen fuel cell is used as a novel clean energy technology and is widely applied to the fields of transportation, distributed power generation, education and scientific research and the like. Aiming at the problems of the running state and the health trend prediction of the hydrogen fuel cell, the prior art mainly surrounds the state monitoring, the fault diagnosis, the abnormality early warning and the like. On one hand, the traditional diagnosis method based on a physical model or expert rules is adopted, common faults are detected through means of threshold judgment, interpolation comparison and the like, but the method is difficult to effectively model state evolution under complex working conditions and multi-source variable coupling relation, and has limited flexibility and generalization capability. On the other hand, with the development of artificial intelligence and big data technology in recent years, intelligent monitoring and prediction schemes based on deep learning, time sequence modeling and black box prediction models become research hotspots increasingly, and the methods can automatically extract multidimensional time sequence characteristics, realize the identification and early warning of key fault trends such as voltage attenuation, film degradation, flooding and the like, and have higher prediction precision and adaptability; However, although the existing fuel cell failure trend prediction methods based on the black box model can obtain a relatively ideal prediction accuracy in actual operation, the main decision process of these models is highly complex and difficult to directly interpret. The mainstream system generally only outputs health indexes or fault risk probabilities, and can hardly reveal the relation between each input variable and the predicted result, and the influence path of each parameter cannot be clearly marked according to a physical causal mechanism. Particularly, in a teaching experiment scene, students and teachers expect to demonstrate and understand the cause of faults, the action mechanism of variables and the prediction logic circulation through the device, but the internal calculation process of a black box model is opaque, so that a learner is difficult to understand why the model makes a certain judgment, and the effect and the knowledge transfer efficiency of experimental teaching are limited; The prior art attempts to introduce generic post-mortem interpretation tools such as SHAP, LIME, etc. to account for depth models. However, such methods often deviate from the specific physical structure and operation mechanism of the fuel cell, do only neighborhood disturbance analysis based on data distribution, and lack interpretation logic fused with discipline knowledge. In addition, advanced methods such as mixed white-black box, federal learning or transfer learning and the like also fail to optimize for teaching friendliness and knowledge attribution chains, so that model interpretation barriers in fuel cell teaching scenes are difficult to effectively solve. Disclosure of Invention The invention provides a method for mapping and predicting states of a hydrogen fuel cell teaching device in real time in order to solve the technical problems. The technical scheme of the invention is realized by a method for mapping and predicting the state of the hydrogen fuel cell teaching device in real time, which comprises the following steps: S1, collecting multi-dimensional time sequence data generated in the operation process of a hydrogen fuel cell teaching device, wherein the data comprise voltage, current, temperature, humidity, gas flow and pressure parameters, and labeling each sensing channel with a corresponding physical variable type and a spatial position label to form an original sensing data sequence with labels; S2, denoising and normalizing the original sensing data sequence with the tag, eliminating high-frequency interference and dimensional difference in the sampling process, and generating a standardized time sequence feature vector as input data of a subsequent modeling; S3, based on a typical fault mode and a physical cause mechanism thereof defined in an expert experience library, constructing a structured 'fault cause-sensing response' associated knowledge graph, wherein nodes represent physical variables or fault types, edges represent causal influence relationships, and initial edge weights are set according to field prior to form a traceable interpretation prior frame; s4, inputting the standardized time sequence feature vector into a lightweight graph attention network model, mapping each physical variable in