A large model-based green hydrogen production equipment fault diagnosis method and device

Abstract

This invention relates to the fields of industrial equipment fault diagnosis and artificial intelligence technology, and in particular to a method and apparatus for fault diagnosis of green hydrogen production equipment based on a large-scale model. The method includes responding to a query request generated by maintenance personnel based on alarm logs; retrieving external knowledge related to the query semantics from a fault knowledge base of all green hydrogen production equipment based on the query request; acquiring multimodal data within a preset time window prior to the triggering time of the query request; extracting features from the multimodal data to obtain data trend features and semantic features; constructing enhanced prompts using the external knowledge, data trend features, and semantic features; inputting the enhanced prompts into a trained target large-scale language model to obtain root cause analysis results. This invention can apply a domain-specific large-scale language model to the field of green hydrogen production, fully utilizing the contextual understanding capabilities of the domain-specific large-scale language model to make the reasoning results more accurate.

Description

Technical Field

[0001] This invention relates to the fields of industrial equipment fault diagnosis and artificial intelligence technology, and in particular to a fault diagnosis method and apparatus for green hydrogen production equipment based on a large model. Background Technology

[0002] With the acceleration of the global energy transition, green hydrogen production technology (such as using renewable energy sources like wind and solar power to electrolyze water to produce hydrogen) is increasingly becoming a focus of attention. Core equipment for green hydrogen production, such as alkaline electrolyzer systems, has a complex structure, operates under harsh conditions, and requires long-term continuous operation. A malfunction can lead not only to production stoppages and economic losses but also potential safety accidents. Therefore, efficient and accurate fault diagnosis of green hydrogen production equipment is crucial.

[0003] Current fault diagnosis methods mainly fall into three categories: one is based on physical models, which requires the establishment of accurate mathematical models of the system. For nonlinear and strongly coupled green hydrogen production systems, modeling is difficult and costly. Another category is based on expert systems, which relies on expert experience rules, resulting in significant knowledge acquisition bottlenecks, poor adaptability, and difficulty in handling unknown faults. The third category is based on traditional machine learning or deep learning, which can learn features from data, but usually requires a large amount of high-quality labeled data, has poor model interpretability, and is difficult to effectively integrate multimodal data for natural language interaction.

[0004] In summary, there are currently no cases of using large-scale domain models in the field of green hydrogen production, and the inference accuracy of existing fault diagnosis methods is not high.

[0005] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0006] In response to the above-mentioned deficiencies or improvement needs of existing technologies, this invention proposes a fault diagnosis method and device for green hydrogen production equipment based on a large model. This method can apply a domain-wide large language model to the field of green hydrogen production, making full use of the contextual understanding capabilities of the domain-wide large language model to make the reasoning results more accurate.

[0007] The embodiments of the present invention adopt the following technical solutions: In the first aspect, the present invention provides a fault diagnosis method for green hydrogen production equipment based on a large model, specifically: responding to a query request generated by maintenance personnel based on alarm logs; Based on the query request, external knowledge related to the query semantics is retrieved from the fault knowledge base of green hydrogen production equipment. Obtain multimodal data within a preset time window prior to the time the query request is triggered, and extract features from the multimodal data to obtain data trend features and semantic features; The external knowledge, the data trend features, and the semantic features are combined to construct enhanced prompt words. These enhanced prompt words are then input into a trained target large language model to obtain root cause analysis results.

[0008] Preferably, the method further includes: The query request is optimized before retrieval by comprehensively using query rewriting and query expansion techniques, and the optimized query is converted into a query vector through a vectorization model. Each document in the fault knowledge base is encoded using the vectorization model to generate a document vector; Calculate the semantic similarity between the query vector and all document vectors, and return the top K results with the highest similarity as candidate results; The candidate results are retrieved and optimized using a re-ranking technique to generate the external knowledge.

[0009] Preferably, the multimodal data includes operational data, image data, and historical maintenance work orders, and the method further includes: Real-time data collection of operational data from all green hydrogen production equipment using sensors; real-time data collection of image data from all green hydrogen production equipment using image acquisition devices. The operational data is captured according to a preset time window; the image data is captured according to a preset time window and then log data is generated; based on the green hydrogen production equipment involved in the query request, historical maintenance work orders are imported from the feedback database. Based on the characteristics of multimodal data, feature extraction is performed on the extracted log data, the log data itself, and the historical maintenance work orders to obtain data trend features and semantic features.

[0010] Preferably, the method for training a target large language model further includes: The pre-trained basic large language model is fine-tuned using text corpus from the fault knowledge base to obtain a domain-adapted model. The domain adaptation model was fine-tuned using historical operating data and fault knowledge base of green hydrogen production equipment to obtain the target large language model.

[0011] Preferably, the method further includes: The enhanced prompt words are segmented to obtain the original input sequence containing multiple tokens; Each token in the original input sequence is converted into a high-dimensional vector and its position is encoded. The context-aware token representation is then iteratively computed through a multi-layer self-attention mechanism. The token represents the probability distribution of the output covering the entire original input sequence, wherein each token in the original input sequence corresponds to a probability value; The token with the highest probability value is appended to the end of the original input sequence, and the multi-layer self-attention mechanism is used for iterative calculation and probability output until the end of the output is marked. All generated tokens are then concatenated to obtain the root cause analysis result.

[0012] Preferably, the method further includes: A structured diagnostic report is generated based on the root cause analysis results; The diagnostic report is then fed back to the operations and maintenance personnel to obtain their assessment of the accuracy of the diagnostic results. The accuracy is associated with the diagnostic process and stored in a feedback database, and the parameters of the target large language model are periodically optimized and adjusted using the contents of the feedback database.

[0013] Preferably, the fault knowledge base includes a vector database, and the method further includes: Build a simulation system for all types of green hydrogen production equipment; The nonlinear coupling between the power generation side and the hydrogen production side in the green hydrogen production process is simulated, and the fault simulation is vectorized and stored in a vector database.

[0014] Preferably, the method further includes: The green hydrogen production equipment includes wind power generation equipment, photovoltaic power generation equipment, energy storage equipment, hydrogen production equipment, and their respective subsystems.

[0015] In a second aspect, the present invention provides a fault diagnosis device for a green hydrogen production equipment based on a large model, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the processor to perform the fault diagnosis method for a green hydrogen production equipment based on a large model in the first aspect.

[0016] Thirdly, the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors to perform the fault diagnosis method for green hydrogen production equipment based on a large model as described in the first aspect.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: it covers real-time data of all green hydrogen production equipment, making fault diagnosis more comprehensive, and supports natural language question answering, which greatly improves the usability of green hydrogen production systems, lowers the technical threshold for operation and maintenance personnel, makes human-machine collaboration more efficient, integrates fault knowledge bases such as equipment manuals and maintenance records with real-time data, applies a domain-wide language model to the field of green hydrogen production, and makes full use of the contextual understanding ability of the domain-wide language model to make the reasoning results more accurate.

[0018] Furthermore, through a feedback learning loop, the green hydrogen production fault diagnosis system based on a large model can continuously learn from field experience, adapt to new green energy hydrogen production models, achieve self-optimization, and become increasingly intelligent with use. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This invention provides a fault diagnosis system for green hydrogen production equipment based on a large model. Figure 2 This is a flowchart illustrating a fault diagnosis method for a green hydrogen production equipment based on a large model, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the fault knowledge base provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the method for optimizing a target big learning model provided in an embodiment of the present invention. Figure 5 This is a flowchart illustrating the feature extraction method provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the data acquisition layer provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of feature extraction provided in an embodiment of the present invention; Figure 8 This is a flowchart illustrating the method for obtaining external knowledge provided in an embodiment of the present invention; Figure 9 This is a flowchart illustrating the method for training a large learning model for training objectives provided in an embodiment of the present invention; Figure 10 This is a flowchart illustrating the method for generating root cause analysis results provided in an embodiment of the present invention; Figure 11This is a flowchart of the method for generating root cause analysis results provided in an embodiment of the present invention; Figure 12 A schematic diagram of a fault diagnosis device for a green hydrogen production equipment based on a large model, provided for an embodiment of the present invention; The accompanying figure is labeled as follows: 21: Processor; 22: Memory. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0023] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0024] In the description of this invention, the expression “A and / or B” (where A and B are used to formally represent specific features) will be used. The corresponding expression includes the following three combinations: only A, only B, and a combination of A and B.

[0025] As used in this invention, “about,” “approximately,” or “approximately” includes the stated value and the average value within an acceptable range of deviation from a particular value, wherein the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the particular quantity (i.e., the limitations of the measurement system).

[0026] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0027] Example 1: A fault diagnosis method for green hydrogen production equipment based on a large model is applied to a fault diagnosis system for green hydrogen production equipment based on a large model. (See reference...) Figure 1 The diagnostic system comprises a data acquisition layer, a core processing layer, a human-computer interaction layer, and a feedback learning layer. The core processing layer includes a data preprocessing module, a multimodal data fusion and analysis module, a retrieval-augmented generation (RAG) module, a fault knowledge base module, a large language model core module, and a diagnostic reasoning and output module. The data acquisition layer is used for collecting and monitoring wind power generation data within the industrial park. The Acquisition layer (RAG) is used to process multimodal data from SCADA (Supervisory and Advanced Computer-Aided Systems) systems, including photovoltaic SCADA, energy storage SCADA, and hydrogen production SCADA systems. The core processing layer connects the data acquisition layer, human-computer interaction layer, and feedback learning layer. It receives multimodal data from the data acquisition layer and preprocesses it using a data preprocessing model. The preprocessed data is then processed by a multimodal data fusion analysis module to extract features. Additionally, the RAG queries received requests in a fault knowledge base. The extracted features and query results are input into the core module of the large language model for inference. The inference results are then fed back to maintenance personnel through a human-computer interaction interface. The human-computer interaction layer facilitates information exchange between maintenance personnel and the system. The feedback learning layer periodically adds diagnostic results from maintenance personnel to the training set to fine-tune the parameters of the large language model.

[0028] like Figure 2 As shown, Embodiment 1 of the present invention provides a fault diagnosis method for green hydrogen production equipment based on a large model. The method specifically includes the following steps: Step 101: Respond to the query request generated by the operation and maintenance personnel based on the alarm log.

[0029] In one embodiment, log data is generated from collected image data. The log data includes the fire location and fire level. By setting filtering rules containing specific alarm keywords (such as "fault," "overheating," "abnormal," or fire level three) and / or regular expressions, when the collected log line content successfully matches the above rules, it indicates that an alarm log has been collected (e.g., "Wind turbine #1 gearbox overheating" or fire level three). At this time, the alarm log is pushed to maintenance personnel through the human-machine interface layer to inform them that an anomaly has occurred in the green hydrogen production equipment. The green hydrogen production equipment includes wind power generation equipment, photovoltaic power generation equipment, energy storage equipment, and hydrogen production equipment and their respective subsystems. Simultaneously, the diagnostic system provides a graphical interface or mobile application. The interactive interface displays the real-time status dashboard, historical data curves, and alarm list of the equipment, and includes a built-in chatbot interaction window. Through the interaction window, maintenance personnel can directly describe the fault phenomenon in natural language (e.g., "Wind turbine #1 gearbox overheating, please analyze the cause") and perform queries. This design lowers the barrier to entry for the diagnostic system.

[0030] Step 102: Based on the query request, retrieve external knowledge related to the query semantics from the fault knowledge base of green hydrogen production equipment.

[0031] In one embodiment, the fault knowledge base for green hydrogen production equipment includes equipment technical manuals, historical fault case libraries, Failure Mode and Effects Analysis (FMEA) tables, expert experience rule bases, and simulation data, among other knowledge in the field of green hydrogen production. (See also...) Figure 3The equipment technical manual refers to the official user manuals and technical specifications for all equipment in the green hydrogen production field; the historical fault case library refers to the records of equipment fault cases accumulated during the historical operation of all equipment in the green hydrogen production field, including the time of fault occurrence, equipment model, fault phenomenon, diagnosis process, handling measures and repair results; the FMEA table refers to the FMEA table conducted for the entire green hydrogen production field, which clearly lists the potential fault modes, occurrence frequency, severity, detection difficulty and risk priority; the expert experience rule library refers to the fault diagnosis and operation and maintenance decision rule library summarized by experts in the field of green hydrogen production, which expresses expert experience in the form of "If-Then" or decision tree; the simulation data refers to the equipment simulation operation data generated based on digital twins or physical models, including the sensor time-series response and abnormal state evolution trajectory under different operating conditions (such as sudden changes in light intensity, power grid fluctuations, electrolyzer overload, etc.). When the diagnostic system receives a query request from maintenance personnel, it can quickly locate similar historical cases, potential fault modes, expert handling suggestions, and relevant simulation analysis results by intelligently searching and matching the current query semantics in the multi-dimensional fault knowledge base that integrates equipment technical manuals, historical fault cases, FMEA analysis tables, expert experience rule base, and simulation model data.

[0032] Step 103: Obtain multimodal data within a preset time window before the time the query request is triggered, and extract features from the multimodal data to obtain data trend features and semantic features.

[0033] In one embodiment, the data acquisition layer of the diagnostic system, in addition to collecting log information in real time, also collects operational data and image data of the entire green hydrogen production equipment. When needed, it extracts historical maintenance work orders related to the green hydrogen production equipment from the feedback database. By extracting features from the aforementioned multimodal data, it obtains data trend features and semantic features reflecting the true state of the entire green hydrogen production equipment. Since the data acquisition layer continuously collects multimodal data from the entire green hydrogen production equipment, and the diagnostic system only initiates diagnostic judgment upon receiving a query request, a large amount of multimodal data has low value for diagnostic analysis of the current query request. This can be addressed by setting a preset time window to perform subsequent feature extraction only on multimodal data within a certain period between the triggering of the query request. This significantly reduces the real-time computing load and improves the timeliness of the diagnosis. The preset time window can be set based on historical diagnostic data.

[0034] Step 104: Construct enhanced prompt words by combining the external knowledge, the data trend features, and the semantic features. Input the enhanced prompt words into the trained target large language model to obtain the root cause analysis results.

[0035] In one embodiment, the external knowledge in step 102 provides prior knowledge in the field of green hydrogen production for diagnosis; the data trend features and semantic features in step 103 reflect the real status of all green hydrogen production equipment. The enhanced prompt words obtained by combining the two integrate multimodal real-time evidence and structured prior knowledge, and are the core inputs that drive the large model or diagnostic algorithm to achieve accurate and reliable root cause analysis. The target large language model is obtained by training an open-source or commercial large language model (LLM).

[0036] In one embodiment, the enhanced prompt words can be understood by the target language big model, transforming numerical trend features and semantic features into natural language descriptions. The aforementioned external knowledge, numerical trend features, and semantic features are then fused in the prompt word template engine according to a predefined logical structure. This template supports conditional branching, priority sorting, and token length constraint optimization to ensure that high information density content is preferentially retained within the maximum context window of the target language model.

[0037] In this embodiment, real-time data from all green hydrogen production equipment is included, broadening the scope of fault diagnosis. It also supports natural language question answering, greatly improving the usability of the green hydrogen production system, lowering the technical threshold for maintenance personnel, and making human-machine collaboration more efficient. It integrates fault knowledge bases such as equipment manuals and maintenance records with real-time data, applying a domain-wide language model to the green hydrogen production field. By fully utilizing the contextual understanding capabilities of the domain-wide language model, the reasoning results are more accurate.

[0038] Furthermore, the system provides a feedback mechanism to optimize the parameters of the target large language model based on feedback from operations and maintenance personnel.

[0039] See Figure 4 This embodiment provides an optimization method, which specifically includes the following steps: Step 201: Generate a structured diagnostic report based on the root cause analysis results.

[0040] In one embodiment, the diagnostic reasoning and output module of the diagnostic system receives the reasoning results of the target large language model and transforms them into a clear, structured diagnostic report that is easy for maintenance personnel to understand. The diagnostic report includes diagnostic conclusions, root cause analysis, remedial recommendations, and predictive maintenance recommendations. The diagnostic conclusions, remedial recommendations, and predictive maintenance recommendations are all derived from the root cause analysis results of step 104.

[0041] Step 202: Feed back the diagnostic report to the operations and maintenance personnel to obtain their assessment of the accuracy of the diagnostic results.

[0042] In one embodiment, maintenance personnel obtain diagnostic reports returned by the diagnostic system through the human-machine interface layer, and perform processing and maintenance on the relevant green hydrogen production equipment according to the processing suggestions in the diagnostic reports. The accuracy of the maintenance results is then scored or marked (useful or useless). Feedback from the maintenance personnel is also fed back to the diagnostic system through the human-machine interface layer.

[0043] Step 203: The accuracy and diagnostic process are associated and stored in the feedback database, and the parameters of the target large language model are periodically optimized and adjusted using the contents of the feedback database.

[0044] In one embodiment, the diagnostic system automatically associates feedback data with corresponding diagnostic process data and stores it in a feedback database. It also periodically uses these new positive and negative samples to incrementally learn or fine-tune the target large language model, enabling the system to continuously evolve and iterate from practical applications.

[0045] like Figure 5 As shown, this embodiment provides a method for extracting features, which specifically includes the following steps: Step 301: Collect real-time operating data of the green hydrogen production equipment across the entire region using sensors; collect real-time image data of the green hydrogen production equipment across the entire region using image acquisition equipment.

[0046] In one embodiment, the data acquisition layer of the diagnostic system also deploys sensors and image acquisition devices (e.g., cameras) to continuously acquire operational data of the entire green hydrogen production system via sensors and to continuously acquire image data of the entire green hydrogen production system via cameras. See also Figure 6 The collected data includes sensor data, alarm and fault data, maintenance work order information, and video surveillance images for each wind turbine in the wind farm; sensor data, alarm and fault data, maintenance work order information, and video surveillance images for all photovoltaic power plants; sensor data, alarm and fault data, maintenance work order information, and video surveillance images for all energy storage plants; and sensor data, alarm and fault data, maintenance work order information, and video surveillance images for all hydrogen production plants.

[0047] Step 302: Extract the running data according to a preset time window; generate log data after extracting the image data according to a preset time window; import historical maintenance work orders from the feedback database based on the green hydrogen production equipment involved in the query request.

[0048] In one embodiment, with the time T that triggers the query request as the endpoint, operating parameters and image data within a time window of [T - Δt, T] are extracted from the historical database, where [T - Δt, T] is a preset time window. For the target equipment (e.g., a wind turbine) involved in the query request, historical maintenance work orders related to the wind turbine are retrieved.

[0049] In one embodiment, flame features (e.g., flame color is typically orange, red, or yellow; flame shape is irregular; edges dynamically change; and flames flicker at a certain frequency) and smoke features (e.g., smoke color is gray or white; smoke shape is diffused; smoke texture is typically semi-transparent and blurry; and smoke rises or spreads) are used. First, the model is trained using supervised methods such as color thresholding, motion detection, or texture analysis to classify fire severity levels. After acquiring image data, the trained model is used to determine whether candidate regions in the image data are flames or smoke. Temporal information (e.g., flickering flames and smoke diffusion in consecutive frames) can be combined to reduce false alarms. The identified fire locations and fire severity levels (pre-defined fire severity levels corresponding to flame and smoke features) are then compiled into log data.

[0050] Step 303: Based on the characteristics of multimodal data, extract features from the truncated log data, the log data, and the historical maintenance work orders to obtain data trend features and semantic features.

[0051] In one embodiment, see Figure 7 Multimodal data can be categorized into numerical data and textual data based on their characteristics. For numerical data (i.e., operational data), time-series features (such as mean, variance, trend, and waveform) are extracted to obtain data trend characteristics. For example, for active power data collected every 15 minutes over the past 24 hours from a wind farm, the mean can be calculated to reflect the overall power output level, and the variance can be used to measure the intensity of power fluctuations. Rising or falling trends can be extracted through linear fitting or moving averages, and then the Fast Fourier Transform (FFT) can be used to further analyze the data. Transform (FFT) or wavelet analysis can be used to identify periodic waveform features. Similarly, for the daytime irradiance and power generation sequence of a photovoltaic power station, the slope change rate from sunrise to sunset can be extracted as the light response trend feature. For the time series of the charge and discharge state of energy storage equipment, its daily maximum or minimum value, average charge and discharge rate, and number of cycles can be statistically analyzed. The hydrogen production rate time series data of hydrogen production equipment can be used to calculate the local mean and standard deviation through a sliding window, and combined with first-order difference to judge the dynamic behavior of start-up, shutdown, or load adjustment. These time series features together constitute a high-dimensional representation of the operating state of a multi-energy coupled system, providing an input basis for subsequent fault diagnosis.

[0052] For text data (i.e., historical maintenance work orders and alarm logs), key information is extracted and semantically encoded from multimodal data to obtain semantic features. For example, structured key information such as "fault type: pitch motor overheating", "occurrence time: 2025-12-15 14:23:07", and "equipment number: WTG-089" are extracted from the alarm logs of wind turbine generators. Similarly, fields such as "work content: inverter replacement", "executor: Zhang San", "associated equipment: INV-PV-204", and "completion status: closed loop" are extracted from photovoltaic power plant operation and maintenance work orders.

[0053] like Figure 8 As shown, this embodiment provides a method for obtaining external knowledge, which specifically includes the following steps: Step 401: Optimize the query request before retrieval by comprehensively using query rewriting and query expansion techniques, and convert the optimized query into a query vector through a vectorization model.

[0054] In one embodiment, query rewriting and query expansion are both techniques for optimizing user queries. First, query rewriting and expansion semantically enhance and standardize the user's input query request. Then, the enhanced content is mapped to a computable semantic space. Query rewriting includes spelling and grammar correction, entity normalization, and structural simplification (i.e., removing modal particles and redundant words). Query expansion includes synonym expansion and hypernym expansion. The text resulting from query rewriting and expansion is richer in information and more standardized in expression than the original user input.

[0055] In one embodiment, a query encoder equipped with a pre-trained dual-tower Transformer model (such as SBERT or SimCSE) transforms the optimized query into a query vector. These models map semantically similar sentences to positions close to each other in the vector space. After multi-layered self-attention mechanisms, the model aggregates the contextual information of the entire sequence, ultimately outputting a high-dimensional dense vector (e.g., a 768-dimensional floating-point vector). This vector is not a simple set of keywords, but a mathematical representation of the overall semantics of the query, containing comprehensive information about concepts, intent, and context. This vectorization process ensures that even if the query and the knowledge base document do not share the same keywords, as long as the semantics are similar, their vectors will be highly similar.

[0056] Step 402: Encode each document in the fault knowledge base using the vectorization model to generate a document vector.

[0057] The vectorization model is the trained vectorization model from step 401, at which point all document vectors and query vectors are in the same semantic vector space.

[0058] Step 403: Calculate the semantic similarity between the query vector and all document vectors, and return the top K results with the highest similarity as candidate results.

[0059] In one embodiment, in this semantic vector space, cosine similarity is used to calculate the semantic similarity between the query vector and all document vectors, and the top K results with the highest similarity are returned as candidate results.

[0060] Step 404: Use re-ranking technology to retrieve and optimize the candidate results to generate the external knowledge.

[0061] In one embodiment, the candidate results retrieved in step 402 are the most semantically relevant to the query. A re-ranking technique is used to select one or more results from the candidate results that are most accurate, useful, and relevant to the user's context. When employing the re-ranking technique, the ranking of candidate results can be increased or decreased based on factors such as exact matching, generality, and complexity. For example, if the document content corresponding to a candidate result precisely contains the error code or model number in the user's query, the ranking of the candidate result corresponding to that document should be significantly improved. If the user's query addresses a general problem, the ranking of root cause analysis and general solution processes covering most cases should be increased, while the ranking of candidate results targeting a highly specific edge case should be decreased. If multiple candidate results are relatively relevant to the user's context, the ranking of candidate results with simpler steps can be increased, while the ranking of candidate results with more complex steps can be decreased.

[0062] like Figure 9 As shown in the figure, this embodiment provides a method for training a large model of a target domain, which specifically includes the following steps: Step 501: Fine-tune the pre-trained basic large language model using the text corpus in the fault knowledge base to obtain the domain-adapted model.

[0063] In one embodiment, the basic large language model is trained with general data and lacks knowledge of the green hydrogen production domain. A secondary training of the basic large language model, such as LLaMA (Large Language Model Meta AI) or ChatGLM (Generative Language Model), using a fault knowledge base for all green hydrogen production equipment, enables the domain-adaptive model to understand professional terminology in the green hydrogen production domain (such as "proton exchange membrane electrolyzer," "alkaline electrolyzer," "rectifier," "deoxygenation unit," etc.), equipment names, and process flow diagram descriptions. This allows the domain-adaptive model to master the text structure and logic of fault descriptions, diagnostic reports, maintenance manuals, and safety procedures, establishing a preliminary domain knowledge framework for the model and making its output text more stylistically and content-wise similar to that of domain experts.

[0064] Step 502: Fine-tune the domain adaptation model using historical operating data of the green hydrogen production equipment to obtain the target large language model.

[0065] In one embodiment, a domain-adaptive model with existing domain knowledge is further trained to learn the actual, dynamic, and numerical operating and fault modes of the equipment. This fine-tuning allows for the association of textual information with actual data. For example, the target large language model can not only understand the term "gearbox overheating" but also associate it with specific data. This enables the target large language model to process and analyze trend data, identify abnormal patterns in the data, predict trends, and explain these data patterns using professional language (e.g., "The current decrease in hydrogen purity is strongly correlated with the increase in cooling water inlet temperature over time; it is recommended to check the heat exchanger efficiency"). This significantly improves the accuracy and practicality of the model in fault prediction, root cause analysis, and operational optimization suggestions.

[0066] When preparing the training set, for each specific historical case in the feedback database, historical operational data, numerical data, and textual data related to the case can be obtained, their features extracted, and prompt words containing external knowledge, trend features, and semantic features constructed. Based on the root cause analysis results already identified in the feedback database, the prompt words are combined with the corresponding root cause analyses to form high-quality enhanced prompt words and root cause training pairs. Using these training pairs to construct the training set, the basic large language model can be fine-tuned for domain adaptation, gradually optimizing its analysis and reasoning capabilities. Ultimately, a target large language model that can accurately understand business scenarios and intelligently diagnose the root causes of problems can be trained. like Figure 10 As shown in the figure, this embodiment provides a method for generating root cause analysis results from a target large language model. The method specifically includes the following steps: Step 601: Perform word segmentation on the enhanced prompt words to obtain the original input sequence containing multiple tokens.

[0067] In one embodiment, the enhanced prompt word is segmented. During the segmentation process, function words without clear semantic functions are removed, and all words with actual meaning are retained. Each generated token carries specific semantic information. For example, if the enhanced prompt word is "Hydrogen production has decreased by 2% in the last five days, and the No. 1 gearbox of the wind turbine is overheating, please help analyze the possible causes.", the original input sequence obtained after segmentation is "recently", "five days", "hydrogen production", "decline", "2%", "No. 1 wind turbine", "gearbox", "overheating", "help", "analyze", "possible", "cause". In this case, one word in the original input sequence is used as a token.

[0068] Step 602: Convert each token in the original input sequence into a high-dimensional vector and perform position encoding. Iteratively calculate the context-aware token representation through a multi-layer self-attention mechanism.

[0069] In one embodiment, each text token is mapped to a high-dimensional vector through the embedding layer of the target large language model. This vector represents its underlying semantic and syntactic information. Since the self-attention mechanism itself cannot perceive order, the model generates a specific positional encoding vector for each token's position. Finally, the semantic vector and positional encoding vector of each token are fused together to form the model's complete encoded representation of the input sequence. This vector sequence is fed into the main body of the model, which consists of hundreds of Transformer decoder layers. Each layer performs complex attention mechanisms and feedforward neural network calculations to ensure that the vector of each token can fully incorporate the information of the entire context. Finally, the model outputs the hidden state vector of the last position, which contains all the information needed to predict the next token.

[0070] Step 603: Based on the token representation, output the probability distribution covering the entire original input sequence, wherein each token in the original input sequence corresponds to a probability value.

[0071] In one embodiment, the hidden state vector from step 602 is passed through a linear layer called the "language model head" to output a probability distribution covering the entire vocabulary. Each token has a probability value representing its likelihood of becoming the next token. For example, "limited power operation": 0.70, "inverter failure": 0.20, "generator failure": 0.05; "low wind speed": 0.05.

[0072] Step 604: Take the token with the highest probability value as the end of the original input sequence, continue to perform multi-layer self-attention mechanism iterative calculation and probability output until the output end mark, concatenate all generated tokens to obtain the root cause analysis result.

[0073] In one embodiment, the token with the highest probability value in step 603, such as "limited power operation," is appended to the end of the original input sequence. This new sequence is then fed back into the model, and steps 602 and 603 are repeated to predict the next token after "limited power operation." This process continues until the model generates a special token indicating the end, or the preset maximum generation length is reached. All generated tokens are then concatenated and converted back into human-readable text (root cause analysis results). See also... Figure 11 The specific process of generating root cause analysis results using the target large language model is as follows: the enhanced prompt words are converted into tokens, each token is semantically understood and its positional order is injected, each token is iteratively considered (including considering context and comprehensive consideration), and the token with the highest probability value in each iteration is generated until a special token indicating the end is generated, or the preset maximum generation length is reached. All generated tokens are concatenated and output back to human-readable text. The target large language model can be an open-source LLaMA that has been fine-tuned and trained in the domain.

[0074] To clearly describe this solution, the following uses an alkaline water electrolysis hydrogen production system as an example to provide a detailed explanation of the workflow of this invention.

[0075] A hydrogen production station supporting a wind power, photovoltaic, and energy storage site has recently experienced a slow decline in hydrogen production in its alkaline electrolyzer system, but the emergency shutdown alarm has not been triggered. The data acquisition and preprocessing module continuously collects system data, including: power generation side data, voltages of each electrolysis cell, total current, hydrogen production flow rate, readings of hydrogen purity analyzers, inlet and outlet temperatures of cooling water, frequency of circulation pumps, etc. At the same time, an alarm log is collected: "Over-temperature of the gearbox of Fan 1#". The operation and maintenance personnel noticed the abnormality and entered a natural language query in the chat box of the human-machine interaction interface: "The hydrogen production has decreased by 2% in the last five days, and the gearbox of Fan 1# is over-temperature. Please help analyze the possible reasons." The large language model core module received the query and immediately initiated the RAG process. It first vectorized the query and retrieved 4 most relevant historical cases (such as "Over-temperature of the gearbox of wind power generation equipment leads to a decrease in hydrogen production", "Over-temperature of the inverter of photovoltaic power generation equipment leads to a decrease in hydrogen production", "Decrease in the cooling efficiency of hydrogen production equipment leads to a decrease in hydrogen production", "Membrane electrode pollution of hydrogen production equipment") and 3 expert rules from the fault knowledge base module. At the same time, the multi-modal data fusion analysis module generated trend charts for data such as hydrogen production flow rate, power generation of Fan 1#, and gearbox temperature in the last five days, and parsed the semantics of the alarm log. The retrieved knowledge, numerical data trend characteristics, and text semantic characteristics were jointly constructed into a rich prompt and input to the fine-tuned domain large language model. After comprehensively analyzing all the information, the domain LLM deduced that the core reason was "Over-temperature of the gearbox of wind power generation equipment, power reduction operation, causing a decrease in the input power of the electrolyzer, and ultimately leading to a decrease in electrolysis efficiency (hydrogen production)". The diagnosis reasoning and output module generated a detailed report based on this result, including the diagnosis conclusion: Over-temperature alarm of the gearbox of wind power generation equipment, power reduction operation (Confidence level: 80%). Root cause analysis: Over-temperature of the gearbox of wind power generation equipment, power reduction operation, input power of the electrolyzer, resulting in a decrease in hydrogen production. Disposal suggestions: ① Check the cooling water quality of the gearbox; ② Clean the heat exchanger; ③ Observe whether the power generation of the generator set returns to normal after cleaning; Predictive maintenance: It is recommended to perform a preventive cleaning on the cooling system of the gearbox of wind power generation equipment once a month. Feedback and learning: The operation and maintenance personnel checked and cleaned the heat exchanger according to the suggestions, and the hydrogen production subsequently returned to normal. He clicked the feedback button of "Diagnosis is accurate" on the interface. The feedback learning module stored this successful case together with all process data in the database for subsequent optimization and iteration of the model.

[0076] Embodiment 2: Based on the method for fault diagnosis of green hydrogen production equipment based on a large model provided in the foregoing embodiment, the present invention further provides a device for fault diagnosis of green hydrogen production equipment based on a large model that can be used to implement the above method, as Figure 12The diagram shown is a schematic representation of the device architecture according to an embodiment of the present invention. The device for fault diagnosis of a green hydrogen production equipment based on a large model in this embodiment includes one or more processors 21 and a memory 22. Wherein, Figure 12 Take a processor 21 as an example.

[0077] Processor 21 and memory 22 can be connected via a bus or other means. Figure 12 Taking the example of a connection between China and Israel via a bus.

[0078] The memory 22, as a non-volatile computer-readable storage medium for fault diagnosis of green hydrogen production equipment based on a large model, can be used to store non-volatile software programs and non-volatile computer-executable programs, such as the fault diagnosis method for green hydrogen production equipment based on a large model in the aforementioned embodiments. The processor 21 executes various functional applications and data processing of the fault diagnosis device for green hydrogen production equipment based on a large model by running the non-volatile software programs, instructions, and modules stored in the memory 22, thereby realizing the fault diagnosis method for green hydrogen production equipment based on a large model in the aforementioned embodiments.

[0079] Memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 22 may include memory remotely located relative to processor 21, which can be connected to processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0080] The program instructions / modules are stored in memory 22 and, when executed by one or more processors 21, execute the fault diagnosis method for green hydrogen production equipment based on a large model as described in the previous embodiments.

[0081] This invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors, for example... Figure 12 One of the processors 21 enables the one or more processors to execute the large-model-based green hydrogen production equipment fault diagnosis method in the foregoing embodiments.

[0082] It is worth noting that the information interaction and execution process between the modules and units in the above-mentioned device and system are based on the same concept as the processing method embodiment of the present invention. For details, please refer to the description in the method embodiment of the present invention, and will not be repeated here.

[0083] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for green hydrogen production equipment based on a large model, characterized in that, include: Responding to query requests generated by operations and maintenance personnel based on alarm logs; Based on the query request, external knowledge related to the query semantics is retrieved from the fault knowledge base of green hydrogen production equipment. Obtain multimodal data within a preset time window prior to the time the query request is triggered, and extract features from the multimodal data to obtain data trend features and semantic features; The external knowledge, the data trend features, and the semantic features are combined to construct enhanced prompt words. These enhanced prompt words are then input into a trained target large language model to obtain root cause analysis results.

2. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The method further includes: The query request is optimized before retrieval by comprehensively using query rewriting and query expansion techniques, and the optimized query is converted into a query vector through a vectorization model. Each document in the fault knowledge base is encoded using the vectorization model to generate a document vector; Calculate the semantic similarity between the query vector and all document vectors, and return the top K results with the highest similarity as candidate results; The candidate results are retrieved and optimized using a re-ranking technique to generate the external knowledge.

3. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The multimodal data includes operational data, log data, and historical maintenance work orders. The method further includes: Real-time data collection of operational data from all green hydrogen production equipment using sensors; real-time data collection of image data from all green hydrogen production equipment using image acquisition devices. The operational data is captured according to a preset time window; the image data is captured according to a preset time window and then log data is generated; based on the green hydrogen production equipment involved in the query request, historical maintenance work orders are imported from the feedback database. Based on the characteristics of multimodal data, feature extraction is performed on the extracted log data, the log data itself, and the historical maintenance work orders to obtain data trend features and semantic features.

4. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The method for training a large language model also includes: The pre-trained basic large language model is fine-tuned using text corpus from the fault knowledge base to obtain a domain-adapted model. The domain adaptation model was fine-tuned using historical operating data and fault knowledge base of green hydrogen production equipment to obtain the target large language model.

5. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The method further includes: The enhanced prompt words are segmented to obtain the original input sequence containing multiple tokens; Each token in the original input sequence is converted into a high-dimensional vector and its position is encoded. The context-aware token representation is then iteratively computed through a multi-layer self-attention mechanism. The token represents the probability distribution of the output covering the entire original input sequence, wherein each token in the original input sequence corresponds to a probability value; The token with the highest probability value is appended to the end of the original input sequence, and the multi-layer self-attention mechanism is used for iterative calculation and probability output until the end of the output is marked. All generated tokens are then concatenated to obtain the root cause analysis result.

6. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The method further includes: A structured diagnostic report is generated based on the root cause analysis results; The diagnostic report is then fed back to the operations and maintenance personnel to obtain their assessment of the accuracy of the diagnostic results. The accuracy is associated with the diagnostic process and stored in a feedback database, and the parameters of the target large language model are periodically optimized and adjusted using the contents of the feedback database.

7. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The fault knowledge base includes a vector database, and the method further includes: Build a simulation system for all types of green hydrogen production equipment; The nonlinear coupling between the power generation side and the hydrogen production side in the green hydrogen production process is simulated, and the fault simulation is vectorized and stored in a vector database.

8. The fault diagnosis method for green hydrogen production equipment based on a large model according to claim 1, characterized in that, The method further includes: The green hydrogen production equipment includes wind power generation equipment, photovoltaic power generation equipment, energy storage equipment, hydrogen production equipment, and their respective subsystems.

9. A fault diagnosis device for green hydrogen production equipment based on a large model, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the processor for performing the fault diagnosis method for a large-scale green hydrogen production equipment according to any one of claims 1-8.

10. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the fault diagnosis method for green hydrogen production equipment based on a large model as described in any one of claims 1-8.

Images