Intelligent fault diagnosis system and method for high-pressure syringe

Abstract

The application provides a high-pressure injector fault intelligent diagnosis system and method, comprising a fault case database, a vector database, a large language model reasoning module and a man-machine interface; the fault case database is formed on the basis of a plurality of real maintenance cases, maintenance logs, maintenance codes and expert maintenance knowledge of multiple brands of high-pressure injectors, and is preprocessed through text cleaning, classification and structure standardization; the vector database converts the natural language content of the fault case database into a programmable key-value pair data structure in JSON format, and generates a vector index through FAISS construction; the large language model reasoning module adopts the technical architecture of retrieval enhancement to generate RAG and a large language model LLM, and is based on the Llama3-8B local offline lightweight small model as the man-machine interface. The application assists maintenance engineers in diagnosing the possible causes, exclusion solutions and matters needing attention of high-pressure injector faults.

Description

Technical Field

[0001] This invention relates to the fields of medical equipment fault diagnosis and artificial intelligence technology, specifically to an intelligent fault diagnosis system and method for high-pressure injectors. Background Technology

[0002] Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) have become important tools for clinical diagnosis. High-pressure injectors, as key auxiliary devices used in these large imaging systems, precisely control the flow rate, volume, and timing of contrast agents, significantly improving image quality and lesion detection rates. However, due to prolonged exposure to strong magnetic fields, X-ray radiation, and contrast agents, high-pressure injectors are prone to failure in their mechanical structures, hydraulic systems, drive circuits, control software, piping systems, and communication modules, requiring rapid diagnosis and troubleshooting. Currently, fault diagnosis of high-pressure injectors relies primarily on engineers' personal experience, equipment error code lookups, and manufacturer technical support, exhibiting significant problems such as strong reliance on experience, low identification rate of latent faults, and long repair response times.

[0003] Therefore, medical equipment repair units urgently need an intelligent diagnostic system and method for high-pressure injector malfunctions to solve the problem of rapid diagnosis and troubleshooting of high-pressure injector malfunctions. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnostic system and method for high-pressure injector faults, so as to realize intelligent diagnosis of high-pressure injector faults and provide diagnostic results.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a high-pressure injector fault intelligent diagnosis system, which includes a fault case database, a vector database, a large language model reasoning module, and a human-computer interaction interface.

[0007] The fault case database is based on several real repair cases, repair logs, repair codes and expert repair knowledge of multi-brand high-pressure injectors, and was formed through text cleaning, classification and structural standardization preprocessing.

[0008] The vector database uses JSON (JavaScript Object Notation) format to convert the natural language content of the fault case database into a programmable key-value pair data structure, and generates a vector index through FAISS (Facebook AISimilarity Search).

[0009] The large language model inference module adopts a technical architecture of Retrieval-Augmented Generation (RAG) and Large Language Model (LLM), and both it and the human-computer interaction interface are based on the local offline lightweight small model of Llama3-8B.

[0010] The diagnostic process of the system is as follows: After the user inputs a fault description through the human-computer interaction interface, the system preprocesses the input text, converts the fault phenomenon language into data vectors through Embedding, performs vector similarity retrieval in the vector database, recalls the top K similar cases, constructs prompt words containing system instructions, retrieval cases and user questions, and drives the large language model to generate diagnostic results containing fault cause analysis, troubleshooting steps suggestions and reference case links.

[0011] This invention establishes a complete technical framework for intelligent fault diagnosis of high-pressure injectors, realizing an intelligent and standardized fault diagnosis process. Through multi-module collaboration, it solves the problems of traditional high-pressure injector maintenance relying on manual experience, resulting in low diagnostic efficiency and poor accuracy. Relying on a local offline lightweight model, it balances diagnostic response speed with medical data security, avoiding the risk of data leakage in cloud deployments. By combining vector retrieval with large language model reasoning, it ensures the accuracy and practicality of diagnostic results.

[0012] Optionally, the fault case database is divided into fault modules and fault types. The fault modules include contrast agent injection module, pressure control module, pipeline system module, motor drive module, mainboard control module, and component matching module.

[0013] The fault types of the contrast agent injection module include unstable injection flow rate / fluctuation between fast and slow, deviation in injection dosage accuracy, and injection interruption / inability to start injection.

[0014] The fault types of the pressure control module include excessive injection pressure triggering an alarm, insufficient pressure / failure to build up pressure, and abnormal pressure deviation in the pressure chamber.

[0015] The fault types of the pipeline system module include pipeline leakage / contrast agent leakage, pipeline blockage / tangling, and air detection error / air intake in the pipeline;

[0016] The fault types of the motor drive module include motor not turning / power output failure, potentiometer error / abnormal motor speed, and motor overheating / protection mechanism triggering.

[0017] The fault types of the motherboard control module include power-on failure / system failure to start, irregular fault code errors, and communication interruption with imaging equipment.

[0018] The fault types of the component supporting modules include syringe installation detection failure and sensor failure / data acquisition abnormality.

[0019] This invention makes the fault case database clearer and more comprehensive, accurately corresponding to common clinical fault scenarios of high-pressure injectors; it facilitates the system to quickly locate the module and specific type of fault, improves the targeting of fault retrieval and diagnosis, and reduces invalid retrieval steps; at the same time, it provides more accurate classification basis for subsequent vector retrieval and model reasoning, further improving diagnostic accuracy.

[0020] Optionally, the preprocessing of the fault case database includes data cleaning, data labeling, and structuring.

[0021] The data cleaning process is used to remove missing key information, incorrectly labeled data, and duplicate or redundant data; correct inconsistent data formats; desensitize sensitive information; and retain core fields such as fault symptoms, fault causes, repair steps, and precautions.

[0022] The data annotation and structuring process are used to annotate historical fault data with fault type, fault phenomenon, fault root cause, maintenance steps and precautions, ensuring that each piece of data is complete and unambiguous in classification.

[0023] This invention cleanses data to remove invalid and redundant data, desensitizes sensitive information, and ensures data quality and medical data compliance. Through data annotation and structuring, it makes fault data standardized and unambiguous, facilitating subsequent vectorization processing and model learning, avoiding retrieval bias and inference errors caused by data disorder, and providing high-quality data support for the accurate operation of the entire system.

[0024] Optionally, the process of building the vector database includes: using a BERT pre-trained language model to vectorize the JSON-formatted structured knowledge base for the high-pressure injector fault domain, specifically through the following steps:

[0025] The input fault symptoms, fault causes, repair steps and precautions are segmented into words, generating a token sequence, and category and separator tags are added;

[0026] The token sequence is input into the BERT encoder, which learns contextual information through a multi-head self-attention mechanism and outputs the hidden layer features at each position.

[0027] The output features corresponding to the classification labels are taken as the semantic sentence vector of the entire text.

[0028] L2 normalization is performed on the semantic sentence vector.

[0029] This invention utilizes the context-aware encoding capability of the BERT model to transform unstructured text into low-dimensional vectors with semantic features, ensuring that semantically similar faulty texts are close in distance in the vector space, thus improving the accuracy of vector retrieval. Through a standardized vectorization process, the uniformity and standardization of vector data are guaranteed, providing support for efficient approximate nearest neighbor search, shortening retrieval time, and improving system response speed.

[0030] Optionally, the construction of the vector database also includes the following process:

[0031] Combine the fault symptoms, fault causes, repair steps, and precautions into a standard input text;

[0032] The BERT pre-trained language model is used to perform batch inference on all maintenance cases to generate a fault vector set in a unified format.

[0033] The fault vector and its corresponding metadata are stored in the FAISS vector database, and a cosine similarity index is established to support efficient approximate nearest neighbor search.

[0034] When a user inputs new fault information, the system vectorizes the query text and performs a similarity match with vectors in the database, returning the most relevant historical cases.

[0035] This invention integrates core fault information through text concatenation, enhancing the semantic integrity of vectors; batch vectorization improves data processing efficiency, adapting to large-scale applications with multiple cases; and the establishment of a cosine similarity index enables rapid semantic matching of fault texts, ensuring that the most relevant historical cases can be quickly recalled after a user inputs a fault, providing accurate reference for model reasoning and further improving diagnostic efficiency and accuracy.

[0036] Optionally, the large language model inference module is based on a small model built on the local offline lightweight model of Llama3-8B and pre-trained in the field of medical equipment. The module freezes the main backbone network parameters of the base through transfer learning technology, with a freezing ratio of not less than 95%, and only opens the feature extraction layer and attention layer. It transfers common knowledge in the field of general medical equipment maintenance to complete the basic cognitive adaptation of medical equipment maintenance scenarios.

[0037] A LoRA matrix with a rank of 8-64 is connected to the base attention layer. Only the weight parameters of this matrix are trained. Based on the fault case database, a prompting learning paradigm is used to complete the directional adaptation of the high-pressure injector fault diagnosis scenario.

[0038] A highly regular instruction fine-tuning dataset based on preprocessed data from a fault case database is constructed. Each instruction adopts a unified structured format. The input end is fixed with fault phenomena, equipment models and fault codes, and the output end is fixed with fault root cause analysis, step-by-step maintenance steps, required tools, precautions and risk warnings. This is used to constrain the model output paradigm and complete the training.

[0039] This application employs a lightweight fine-tuning strategy to achieve accurate model adaptation to knowledge in the high-pressure injector repair domain without consuming a large amount of computing power, thus solving the problems of general models lacking domain knowledge and having low inference accuracy. By combining transfer learning and few-sample fine-tuning, it maximizes the use of limited high-pressure injector-specific fault data, avoiding the problems of model overfitting and poor generalization ability caused by few-sample training. The standardized instruction format improves the model's efficiency in understanding input information and the standardization of output results.

[0040] Optionally, the large language model inference module achieves model illusion suppression through the synergistic effect of cue word engineering optimization and retrieval enhancement generation RAG technology, as follows:

[0041] Design a dedicated prompt word template for the field of high-pressure injector fault diagnosis, including system instructions, small sample constraint logic, input format specifications, and output format specifications. The small sample constraint logic embeds typical fault prompt examples. The input format supports three methods: text description, fault code input, and module selection. The output format is consistent with the output format of the instruction fine-tuning dataset, constraining the output logic and specifications of the large language model inference module.

[0042] The system receives fault information input by the user, and after preprocessing and vectorization, it initiates a cosine similarity retrieval to the vector database and recalls the top 3-10 similar cases. The similar cases, system instructions and user questions are concatenated into prompt words and input into the large language model reasoning module. The large language model reasoning module combines the retrieved cases with its own acquired knowledge to complete the reasoning and generate a diagnostic result containing the root cause of the fault and maintenance steps.

[0043] The root causes of faults and repair steps generated by the large language model inference module are verified. If logical contradictions exist, supplementary cases are retrieved a second time and diagnostic results are regenerated. The output of fault types and repair solutions that do not appear in the training samples and vector database by the large language model inference module is restricted to reduce the probability of hallucination.

[0044] This invention optimizes prompt word engineering to constrain the model's reasoning range, prevent excessive model divergence, and improve the relevance of reasoning. It also deeply binds model reasoning with a vector knowledge base using RAG technology, effectively suppressing model illusions and preventing the model from fabricating maintenance plans and fault causes, thus ensuring the safety and accuracy of maintenance guidance. Furthermore, the support for multiple input methods lowers the barrier to entry for engineers and enhances the system's operability.

[0045] Optionally, the system adopts a hybrid local deployment environment of a central processing unit and a lightweight graphics processor, and uses the Ollama and TextGenerationWebUI lightweight deployment framework to complete the deployment; the local offline lightweight small model of Llama3-8B is subjected to INT4 or INT8 quantization compression processing.

[0046] This invention adopts a local hybrid deployment environment adapted to the configuration of ordinary maintenance workstations, without the need for high-end computing power equipment, thus reducing system deployment costs and hardware thresholds; the lightweight deployment framework simplifies the deployment process, enabling one-click start-up and offline operation, and is adapted to the usage scenarios of clinical maintenance sites; model quantization compression reduces model video memory and memory usage without significantly reducing inference accuracy, improves local running speed, and ensures that response latency meets the real-time diagnostic needs of the site.

[0047] Optionally, the system divides the preprocessed sample data into a training set, a validation set, and a test set in a 6:2:2 ratio, and uses leave-one-out cross-validation and k-fold cross-validation to verify the model performance.

[0048] This invention achieves a scientific allocation of training, validation, and test sets through a reasonable sample partitioning ratio, ensuring the sufficiency of model training and the effectiveness of validation; the multiple cross-validation scheme maximizes the use of limited sample data, avoids sample waste and model overfitting, improves the model's generalization ability and diagnostic stability, and ensures that the model can achieve accurate diagnosis under different fault scenarios.

[0049] Secondly, the intelligent diagnosis method for high-pressure injector faults according to the present invention, applied to the intelligent diagnosis system for high-pressure injector faults described in this application, includes the following steps:

[0050] Step 1: Collect real-world repair cases, repair logs, fault codes, and expert repair knowledge from different brands of high-pressure injectors to obtain raw fault data;

[0051] Step 2: Perform text cleaning, information anonymization, and structural standardization preprocessing on the original fault data, divide the equipment functional modules and corresponding fault types, and construct the fault case database;

[0052] Step 3: Convert the natural language content in the fault case database into a JSON key-value pair data structure, use the BERT pre-trained language model to complete the text vectorization process, build a vector index based on FAISS, and construct the vector database.

[0053] Step 4: Select the Llama3-8B local offline lightweight small model as the base model, and use a combination of few-shot learning, LoRA lightweight fine-tuning and transfer learning to complete the model training and optimization, and combine prompt word engineering to complete the instruction adaptation.

[0054] Step 5: Obtain the fault description information input by the user through the human-computer interaction interface, and perform preprocessing operations on the input text;

[0055] Step 6: Convert the fault description text into a data vector using Embedding, perform vector similarity retrieval in the vector database, and recall a preset number of similar historical fault cases;

[0056] Step 7: Combine system commands, retrieved similar cases, and user troubleshooting issues to construct complete prompt keywords;

[0057] Step 8: Input the prompt words into the large language model inference module for inference calculation, and output the fault diagnosis results, which include fault cause analysis, troubleshooting steps suggestions, and reference case information.

[0058] This invention can assist maintenance engineers in quickly and accurately diagnosing the possible causes, troubleshooting solutions, and precautions for high-pressure injector malfunctions based on faulty equipment and its symptoms. It can effectively identify scenarios with the same malfunction symptoms but different causes and provide differentiated troubleshooting paths. Attached Figure Description

[0059] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings and tables, wherein:

[0060] Figure 1 This is a block diagram of the intelligent diagnostic system for high-pressure injector faults in this application;

[0061] Figure 2 This is an overall framework diagram of the intelligent diagnostic system for high-pressure injector faults in this application;

[0062] Figure 3 This is a diagram of the human-computer interaction interface of the intelligent diagnostic system for high-pressure injector faults in this application;

[0063] Figure 4 This is a bar chart comparing traditional repair and system-assisted repair in this application;

[0064] Figure 5 This is a graph of the model training loss function in this application;

[0065] Figure 6 This is a flowchart of the intelligent diagnostic method for high-pressure injector faults in this application. Detailed Implementation

[0066] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0067] like Figure 1 As shown, an intelligent fault diagnosis system includes a fault case database, a vector database, a large language model inference module, and a human-computer interaction interface. The fault case database is based on over 343 real-world maintenance cases, maintenance logs, maintenance codes, and expert knowledge covering leading high-pressure injectors from Bayer, Stellant, Ulrich, and Euligion, formed through text cleaning, classification, and structural standardization preprocessing. The vector database uses JSON (JavaScript Object Notation) format to convert the natural language content of the fault case database into a programmable key-value pair data structure, and generates vector indexes using FAISS (Facebook AI Similarity Search). The large language model inference module employs a Retrieval-Augmented Generation (RAG) and Large Language Model (LLM) technical architecture, and both it and the human-computer interaction interface are based on a local offline lightweight small model (Llama3-8B). The system's diagnostic process is as follows: After the user inputs a fault description through the human-computer interaction interface, the system preprocesses the input text, converts the fault phenomenon language into data vectors through Embedding, performs vector similarity retrieval in the vector database, recalls the top K similar cases, constructs prompt words containing system instructions, retrieval cases and user questions, and drives the large language model to generate diagnostic results containing fault cause analysis, troubleshooting steps suggestions and reference case links.

[0068] This application establishes a complete technical framework for intelligent fault diagnosis of high-pressure injectors, realizing an intelligent and standardized fault diagnosis process. Through multi-module collaboration, it solves the problems of traditional high-pressure injector maintenance relying on manual experience, resulting in low diagnostic efficiency and poor accuracy. Relying on a local offline lightweight model, it balances diagnostic response speed with medical data security, avoiding the risk of data leakage in cloud deployments. By combining vector retrieval with large language model reasoning, it ensures the accuracy and practicality of diagnostic results.

[0069] The following is a detailed description of this application:

[0070] I. Construction of the Fault Case Database

[0071] The fault case database is divided into fault modules and fault types. The fault modules include contrast agent injection module, pressure control module, pipeline system module, motor drive module, mainboard control module, and component matching module.

[0072] For example, 343 real repair cases of high-pressure injectors from leading tertiary hospitals and several partner hospitals, accumulated from 2019 to 2025, were collected. These cases covered first-line high-pressure injectors from companies such as Bayer Stellant, Euligia, Wandong Medical, United Imaging Healthcare, Mindray Medical, Shenzhen Anke, Shandong Weigao, and Nanjing Jusha. The cases encompassed six core modules: contrast agent injection, saline flushing, pressure control, tubing system, motor drive, and mainboard control. The focus was on common clinical fault types, including abnormal pressure, unstable injection flow rate, tubing leakage, fault code errors, startup failures, and communication anomalies. The distribution of fault samples is shown in Table 1. Based on this, all cases underwent structural standardization preprocessing to form a fault case database, as shown in Table 1. In Table 1, " / " represents "or".

[0073] Table 1 shows the distribution of fault samples:

[0074]

[0075] After obtaining the above historical fault data, the historical fault data is preprocessed, including data cleaning, data labeling, and structuring.

[0076] Data cleaning includes removing missing key information, incorrectly labeled data, and duplicate or redundant data; correcting inconsistent data formats; desensitizing sensitive information (hospital name, patient information, engineer personal information); and retaining core fields such as fault symptoms, fault causes, repair steps, and precautions.

[0077] Table 2 shows an example of data cleaning results:

[0078]

[0079] Data labeling and structuring: Historical fault data is labeled with dimensions such as fault type, fault phenomenon, fault root cause, maintenance steps and precautions, so that each piece of data information is complete and the classification is unambiguous.

[0080] Table 3 shows the data annotation dimensions and explanations.

[0081]

[0082] The processed sample data is divided into training, validation, and test sets in a 6:2:2 ratio. Leave-one-out cross-validation and k-fold cross-validation (k=5) are used to maximize the use of the limited sample data and avoid sample waste and overfitting.

[0083] II. Construction of a Knowledge Base for High-Pressure Injector Repair

[0084] Knowledge graph construction: The equipment structure, fault correlation and maintenance logic of high-pressure injectors are analyzed to construct a knowledge graph containing equipment modules, fault phenomena, fault root causes, maintenance solutions and auxiliary information, and to clarify the correlation between various faults and fault reasoning logic.

[0085] Vector Knowledge Base Construction: A lightweight vector database (such as Chroma or FAISS) was selected, and a BERT pre-trained language model was used to vectorize the structured knowledge base (JSON) for high-pressure injector faults, constructing a vector knowledge base for maintenance scenarios. This vector knowledge base enables semantic retrieval and rapid matching of fault texts, supports real-time retrieval of relevant maintenance knowledge by the model, improves inference accuracy and response speed, and adapts to local offline retrieval needs.

[0086] 1. Building a JSON-based structured database

[0087] A structured database for high-pressure injector fault repair is constructed using JSON format to achieve standardized, lightweight, and efficient storage and management of multi-source repair data. JSON, as a lightweight data exchange format, boasts advantages such as concise syntax, fast parsing, and easy integration with vector libraries and large models, making it suitable for local offline deployment and small-sample applications. This application standardizes and encapsulates 343 repair cases in a unified JSON format. Each record contains seven core fields: fault type, fault phenomenon, fault code, root cause, tools / parts, precautions, and equipment model, forming a hierarchical and complete structured data entry. All data, after anonymization, is stored in independent JSON files and collection-style JSON arrays, supporting fast local reading, retrieval, and parsing. Simultaneously, field constraint rules are established to ensure data format uniformity, completeness, and unambiguity. The JSON-based structured database can seamlessly integrate with vector databases and RAG modules, providing standardized input and output for the model, significantly improving knowledge retrieval efficiency and inference accuracy, and providing stable, secure, and efficient underlying data support for intelligent diagnostic systems.

[0088] 2. Principles and steps of text vectorization based on the BERT model

[0089] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer bidirectional encoder. Its vectorization principle involves using a multi-layer self-attention mechanism to perform context-aware encoding of text, mapping unstructured natural language text into low-dimensional, dense real-valued vectors, i.e., sentence vectors. These sentence vectors preserve the semantic information of the text, ensuring that semantically similar texts are close in distance within the vector space. In this application, the BERT model is used to vectorize texts such as descriptions of high-pressure injector malfunctions and repair cases, constructing a vector knowledge base to achieve semantic retrieval and similarity matching of fault texts, providing accurate domain knowledge support for subsequent local small-model inference.

[0090] The BERT pre-trained language model performs text vectorization, and the specific process is as follows:

[0091] (1) Text segmentation and sequence construction: The input text of fault phenomena, fault causes (including fault codes and root causes), maintenance steps and precautions are segmented to generate a token sequence acceptable to the model, and classification and separation tags are added.

[0092] (2) Contextual semantic encoding: Input the token sequence into the BERT encoder, learn the contextual association information through the multi-head self-attention mechanism, and output the hidden layer features at each position.

[0093] (3) Sentence vector generation: Take the output features corresponding to the classification label as the semantic sentence vector of the whole text, with a dimension of 768.

[0094] (4) Vector normalization: Perform L2 normalization on the semantic sentence vector to improve the stability of cosine similarity calculation.

[0095] 3. Vector Knowledge Base Construction Process

[0096] (1) Text structure concatenation: Concatenate fields such as fault phenomenon, fault cause, maintenance steps and precautions into standard input text.

[0097] (2) Batch vectorization: Use the BERT pre-trained language model to perform batch inference on all maintenance cases to generate a fault vector set in a unified format.

[0098] (3) Vector storage and index establishment: Store the fault vector and corresponding metadata into the FAISS vector database, establish a cosine similarity index, and support efficient approximate nearest neighbor search.

[0099] (4) Semantic retrieval: When a user inputs new fault information, the system vectorizes the query text and performs similarity matching with vectors in the database, returning the most relevant historical cases to provide retrieval support for the large language model. The intelligent operation and maintenance knowledge base based on deep learning has been applied in industrial scenarios, verifying the potential of semantic retrieval in improving fault handling efficiency.

[0100] III. Local Small Model Selection, Deployment, and Quantization Optimization

[0101] The core of this application is a local, lightweight, small model, rejecting cloud deployment. It prioritizes models with moderate parameter size, low hardware requirements, and open-source commercial viability, balancing inference performance and offline deployment feasibility.

[0102] 1. Model selection criteria: The parameter size should be controlled between 3B and 7B (small model range, suitable for local operation on ordinary desktops / laptops), support Chinese command fine-tuning, support quantization compression, and be open source with no commercial copyright risks. The alternative models include Llama3-8B (lightweight after quantization), Qwen2-7B, Phi-34K-7B, Mistral-7Bv0.3, etc. The optimal model was determined after local hardware adaptation testing. The specific model evaluation dimensions are shown in Table 4.

[0103] Table 4 shows the model evaluation dimensions.

[0104]

[0105] 2. Local Environment Deployment: Build a CPU + lightweight GPU hybrid local deployment environment, which does not require high-end computing power equipment, is compatible with ordinary maintenance workstation configurations, and adopts lightweight deployment frameworks such as Ollam and TextGenerationWebUI to simplify the deployment process and achieve one-click start and offline operation.

[0106] 3. Model Quantization Compression: The selected model is quantized and compressed using INT4 / INT8. Without significantly reducing the inference accuracy, the model's GPU memory and RAM usage are greatly reduced, improving local running speed and ensuring no lag and a response latency of less than 3 seconds on ordinary hardware, meeting the real-time requirements of on-site maintenance.

[0107] IV. Model-specific fine-tuning and prompt project optimization

[0108] The large language model inference module used in this invention is based on Llama3-8B. The overall network structure consists of a main backbone network, a feature extraction layer, a multi-head attention layer, a feedforward network layer, and an output layer stacked sequentially. The main backbone network is composed of multiple Transformer encoders connected in series, responsible for capturing the global contextual semantic information of the text sequence. The feature extraction layer, embedded at the bottom of the backbone network, is composed of convolutional and fully connected layers, used to map word embeddings of the input text into high-level semantic features. The multi-head attention layer is connected after the feature extraction layer, constructing multiple attention heads in parallel to learn semantic dependencies in different dimensions, and outputting enhanced feature representations through weighted fusion. The feedforward network layer follows the attention layer, further refining features through two fully connected nonlinear transformations. The output layer decodes the final features and outputs structured diagnostic results. The overall network uses residual connections and layer normalization to achieve stable training and feature transfer of the deep network. Specific operations are as follows:

[0109] 1. Fine-tuning Method Selection: A dual strategy of LoRA (Lower-Rank Adaptation) fine-tuning and transfer learning is adopted. Based on a small model pre-trained in the medical device field, the small model base is constructed using a local offline lightweight Llama3-8B model. Transfer learning is used to freeze the main backbone network parameters of the base, with a freezing ratio of no less than 95%, only opening the feature extraction layer and the multi-head attention layer. Common knowledge from the general medical device maintenance domain is transferred to complete the basic cognitive adaptation for the maintenance scenario. Subsequently, in the multi-head attention layer, a LoRA low-rank matrix with a rank of 8–64 is connected in parallel next to the projection matrix of each attention head, forming a structure where the sovereign heavy matrix and the low-rank increment matrix are superimposed in parallel. During the training phase, all backbone weights W of the original model are frozen, and only the low-rank matrices A and B are updated. Small-sample targeted fine-tuning is performed using 343 high-pressure injector fault data points to achieve accurate injection of fault diagnosis scenario knowledge. This saves computational resources and avoids the problems of model overfitting and poor generalization caused by small samples, significantly improving the efficiency of utilizing limited data to complete the targeted adaptation for high-pressure injector fault diagnosis scenarios. Recent research based on the TransQwen model in transformer fault diagnosis has confirmed that the parameter-efficient fine-tuning method can also generate high-quality maintenance strategies under small sample conditions.

[0110] 2. Fine-tuning Dataset Construction: Based on 343 precisely labeled maintenance data, a highly standardized instruction fine-tuning dataset was constructed. The fine-tuning dataset was built using a database of high-pressure injector failure cases. The format of each instruction was strictly standardized as "User: Fault Phenomenon + Equipment Model + Fault Code; Assistant: Fault Root Cause Analysis + Step-by-Step Maintenance Steps + Required Tools + Precautions + Risk Warning". All ambiguous expressions were eliminated, and sentence structure was optimized for each sample to improve the model's learning efficiency on small sample data.

[0111] 3. Prompt Engineering Optimization: A domain-specific prompt word template is designed, with additional small-sample constraint logic. This logic, integrated into the prompt word template, constrains the reasoning behavior of the large language model's reasoning module. It limits the module to reasoning based on the existing knowledge base and a limited set of cases. The existing knowledge base is the fine-tuning dataset, and the limited cases are high-pressure injector-related repair samples from the vector database, prohibiting excessive divergence. Simultaneously, the fault information input format is optimized, supporting multiple input methods including text description, fault code input, and module selection. This lowers the barrier to entry for engineers and adapts to the precise reasoning needs of the small-sample model. The output format remains consistent with the fine-tuning dataset, achieving precise constraints on the output logic of the large language model's reasoning module.

[0112] 4. By leveraging RAG (Retrieval Enhancement Generation) technology, the model reasoning is deeply integrated with the vector knowledge base. Specifically, this is achieved by: establishing a data interaction channel between the large language model reasoning module and the vector database; integrating RAG technology into the reasoning process of the large language model reasoning module, with a pre-defined retrieval trigger mechanism; automatically triggering a vector database retrieval instruction when the large language model reasoning module receives fault input and enters the diagnostic result generation stage; the large language model reasoning module first preprocesses and vectorizes the fault information input by the user, then initiates a similar case retrieval from the vector database, prioritizing the retrieval of all 343 samples in the knowledge base before model output, recalling relevant repair samples; based on the retrieved similar cases and its learned fault diagnosis knowledge, combined with reasoning to generate answers, relying on precise matching from the knowledge base to compensate for the insufficient accuracy of the model's small-sample training, preventing the model from fabricating repair plans and fault causes, ensuring the safety and accuracy of repair guidance, and achieving deep linkage and adaptation between the two.

[0113] Table 5 shows the data on the hallucination suppression effect of the model.

[0114]

[0115] Introducing RAG and a knowledge base can effectively reduce model illusions and root cause errors, while significantly improving the accuracy of repair solutions, thus having significant engineering value for actual repair tasks.

[0116] V. Experimental and Test Results

[0117] 1. Data Statistics and Analysis Methods

[0118] To address the characteristics of the small sample experimental data, SPSS and Excel were used for refined statistical analysis. Quantitative data employed descriptive statistics such as accuracy, matching degree, time consumption, t-value, and p-value. Due to the small sample size, paired t-tests were used instead of independent samples tests for the experimental and control groups. Non-parametric tests were employed to assist in validation, avoiding the normality bias problem associated with small sample data and ensuring robust and reliable test results. The study focused on analyzing the differences in repair efficiency and diagnostic accuracy before and after AI assistance. Furthermore, fault type sub-categorization statistics were performed on the 100 samples to explore the assistance effect under different fault scenarios. Qualitative data were analyzed using content analysis, compiling engineer feedback and field test opinions to extract system optimization directions. Targeted knowledge base supplementation was conducted to address the shortcomings of the small sample model, ensuring the research results are objective, scientific, and conform to the academic norms of small sample research.

[0119] 2. Quality Control Measures

[0120] Data Quality: For the small sample of 343 cases, a dual-person, dual-round precision annotation + full-staff verification mechanism was implemented. Two senior high-pressure injector maintenance engineers independently completed the annotation of all data. Any inconsistent data were discussed and resolved by a special team. The final data annotation accuracy rate was required to reach 100%, so as to prevent a single annotation error in a small sample from affecting the overall research results. The extreme data quality made up for the shortcomings of insufficient sample size.

[0121] Model quality: Multiple rounds of model validation are conducted using k-fold cross-validation (k=5). Each validation set comprehensively covers all types of faults. The model only enters the testing phase after all validation set metrics meet the standards. Model outputs are reviewed case-by-case by engineers before use in experiments and testing, strictly controlling the generalization error of the model after small-sample fine-tuning. The average loss of k-fold cross-validation is... The calculation is shown in formula (1).

[0122]

[0123] in, Let be the loss value of the model on the validation set in the i-th round of cross-validation.

[0124] Experimental quality: A paired control design was adopted to strictly control irrelevant variables such as maintenance engineer qualifications, fault samples, equipment environment, and test procedures. The experimental standards were consistent throughout the process. To address the deviation problem that is prone to occur with small samples, repeated testing was added, and core fault cases were repeatedly verified to minimize experimental errors.

[0125] Model quality: Multiple rounds of validation are set up. The model will only enter the testing phase after the validation set accuracy meets the standard. The model output can only be used for experiments and testing after being reviewed by engineers.

[0126] Experimental quality: Experimental variables were controlled, and the qualifications of the maintenance engineers, fault samples, and equipment environment of the two groups were kept consistent to avoid irrelevant variables from interfering with the experimental results.

[0127] 3. Ethical and Compliance Safeguards

[0128] We strictly adhere to laws and regulations related to medical data security. All data is stored locally and processed offline throughout the entire process, without being uploaded to any cloud servers, thus eliminating the risk of medical data leakage. The model output is only used as a maintenance assistance reference and does not replace the final judgment of professional engineers. We clearly define the AI-assisted positioning to avoid misleading maintenance operations. We select open-source and commercially viable models to avoid intellectual property risks.

[0129] 4. Experimental Design and Effect Verification

[0130] 4.1 Controlled Experimental Design

[0131] A combination of controlled experiments and field testing was used to evaluate system performance from multiple dimensions and verify the effectiveness of local small-scale model-assisted maintenance.

[0132] Considering the small sample size of 343 cases, the experimental design was adjusted to a small-sample paired control experiment, eliminating the large-sample random grouping method. 100 valid fault samples were selected as the experimental basis, employing a combination of self-control and cross-control: In the first stage, 100 faults were troubleshooted using traditional manual experience, and relevant data were recorded; in the second stage, a local AI model assisted in troubleshooting the same 100 faults, operated by equally qualified maintenance engineers. This approach minimized sample differences and avoided experimental errors due to insufficient sample size, focusing on comparing the core maintenance indicators of the two groups. Furthermore, stratified statistics were performed for difficult and common faults, with 80 common faults and 20 difficult faults, to separately analyze the auxiliary effects of different fault types, improving the reliability of the experimental results.

[0133] 4.2 Experimental Conditions

[0134] (1) Hardware environment

[0135] Computer: CPU Intel Core i7-12700H, RAM 16GB DDR4, Graphics Card NVIDIA RTX 4060 6GB

[0136] Storage: 512GB NVMe SSD for local offline storage of models, knowledge base, and data.

[0137] Test equipment: Clinically used high-pressure injectors (mainstream models such as MEDRAD and EOLIC). The experimental environment fully considered the potential impact of special working conditions such as high field strength and magnetic field in the radiology department on the operation and failure performance of the equipment.

[0138] Network environment: The entire process is conducted offline to ensure the security of medical data.

[0139] (2) Software environment

[0140] Operating System: Windows 10 64-bit Professional Edition

[0141] Model Deployment: Olama, Text Generation Website

[0142] Development frameworks: PyTorch, Transformers, FAISS, Chroma

[0143] Data processing: Python 3.9, Excel, SPSS 26.0

[0144] Model and Quantization: Qwen2-7B, INT4 quantization, LoRa fine-tuning

[0145] Knowledge base: JSON structure library + BERT vectorization + RAG search enhancement

[0146] (3) Experimental personnel and control conditions

[0147] Maintenance Engineers: 2 people, 3-5 years of experience in high-pressure injector repair, both teams must have the same qualifications.

[0148] Fault Samples: 100 real high-pressure injector faults (80 common cases, 20 difficult cases)

[0149] Controlled variables: Fault samples, equipment models, maintenance procedures, and evaluation criteria are all completely standardized.

[0150] Evaluation method: Double-blind comparison, cross-comparison between traditional repair and AI-assisted repair.

[0151] (4) Evaluation criteria and tools

[0152] Evaluation metrics: diagnostic accuracy, root cause localization rate, solution matching degree, repair time, repeat repair rate, response time, and engineer satisfaction.

[0153] Statistical methods: paired t-test, p < 0.05 was considered statistically significant;

[0154] Recording tools: stopwatch, maintenance record sheet, satisfaction rating sheet.

[0155] (5) Evaluation indicator system

[0156] Technical performance indicators: fault diagnosis accuracy, fault root cause location accuracy, repair scheme matching degree, model response time, and knowledge base retrieval accuracy. The mathematical expression of the accuracy is shown in formula (3):

[0157]

[0158] Practical efficiency indicators: average repair time per fault and repeat repair rate.

[0159] Security and compliance metrics: Local data storage compliance, model output error rate, and accuracy of maintenance operation risk warnings.

[0160] User experience metrics: maintenance engineer satisfaction rating, system ease of operation, and user-friendliness of the interface.

[0161] (6) Cooperative reasoning using RAG and Prompt

[0162] A RAG retrieval-enhanced generative reasoning mechanism is constructed. Semantic retrieval of the user-input fault query text is performed using a vector database to obtain the Top-K relevant repair cases and knowledge entries (K ranges from 3 to 5) with the highest matching degree to the fault problem. The retrieved structured knowledge, such as fault root causes, troubleshooting steps, and reference cases, is integrated with the user's question to construct optimized prompts, which are then input into a large language model. The large language model performs joint reasoning based on the retrieved knowledge and the user's question, outputting standardized and highly accurate fault diagnosis conclusions and repair guidance solutions. This achieves precise question answering under domain knowledge constraints, avoiding model-induced illusions.

[0163] The code example is shown below.

[0164] "fault_category":"Abnormal stress",

[0165] "possible_causes":[

[0166] {"cause":“Poor contact in the pressure sensor connection wire”,"confidence":"high”","basis":“Case #032 Perfect Match”},

[0167] {"cause":"Leakage at the connection between the syringe and the syringe","confidence":"medium","basis":"Partial match of Case #045"}

[0168] {"cause":"pressure sensor module aging","confidence":"Iow","basis":"based on knowledge reasoning"}],

[0169] "troubleshooting_steps":[

[0170] 1. Check the pressure sensor connection cable: unplug and replug the connector, and observe whether the pressure value recovers.

[0171] 2. Check the syringe seal: Replace with a new seal, reinstall, and test.

[0172] 3. Use a multimeter to test the sensor output signal: the normal range should be 0.5-4.5V; otherwise, replace the sensor.

[0173] "reference_cases":["#032","#045"],

[0174] "need_verification":false}

[0175] Table 6 shows the statistical results of the comparative experiment between traditional repair and system-assisted repair.

[0176]

[0177] A bar chart comparing traditional repair and system-assisted repair, such as... Figure 4 As shown.

[0178] The comparative experimental results of traditional maintenance and system-assisted maintenance show that this application is significantly superior to the traditional model in all indicators (P<0.01 or P<0.001). The fault diagnosis accuracy rate of system-assisted maintenance reaches 94.67%, the root cause location accuracy rate is 92.00%, and the maintenance plan matching degree is 91.33%, which is far higher than the 72.33%, 68.00%, and 65.67% of traditional maintenance, respectively. The average time for a single fault is reduced from 48.62 minutes to 16.45 minutes, the average number of troubleshooting steps is reduced from 9.2 steps to 3.8 steps, and the repeat maintenance rate is reduced from 18.00% to 3.00%. Engineer satisfaction score is increased from 5.2 to 9.1, and the average response time of the model is only 2.46 seconds, which meets the needs of rapid on-site diagnosis.

[0179] The accuracy rates of common faults (80 cases) and difficult faults (20 cases) are compared in Table 7.

[0180] Table 7 compares the accuracy rates for common faults and difficult faults.

[0181]

[0182] Among 80 common faults, the accuracy rate of traditional repair was 78.75%, while the accuracy rate of system-assisted repair was 97.50%, an improvement of 18.75%. Among 20 difficult faults, the accuracy rate of traditional repair was only 45.00%, while the accuracy rate of system-assisted repair reached 80.00%, an improvement of 35.00%. This indicates that the system has a more significant effect on improving the accuracy of hidden and complex difficult faults.

[0183] The model retrieval and performance indicators are shown in Table 8.

[0184] Table 8 shows the model retrieval and performance indicators.

[0185]

[0186] The knowledge base retrieval accuracy is 95.33%. After INT4 quantization, the Qwen2-7B model has a video memory usage of 3.8GB, a CPU inference speed of 18.6 tokens / s, and a model hallucination rate of only 1.33%. It can run stably on ordinary desktop computers and repair workstations.

[0187] like Figure 5 As shown, the model training loss function curves are displayed. The model training loss and validation loss curves decrease steadily overall, with the training loss decreasing from 2.85 to 0.37 and the validation loss decreasing from 3.12 to 0.85. There are no obvious oscillations or overfitting phenomena, indicating that the small sample LoRA fine-tuning strategy is effective and the model has good convergence.

[0188] VI. Results

[0189] like Figure 6 As shown, a method for intelligent diagnosis of high-pressure injector faults, applied to the intelligent diagnosis system for high-pressure injector faults of this application, includes the following steps:

[0190] Step 1: Collect real repair cases, repair logs, fault codes, and expert repair knowledge of different brands of high-pressure injectors to obtain raw fault data.

[0191] Step 2: Perform text cleaning, information anonymization, and structural standardization preprocessing on the original fault data, divide the equipment functional modules and corresponding fault types, and construct the fault case database.

[0192] Step 3: Convert the natural language content in the fault case database into a JSON key-value pair data structure, use the BERT pre-trained language model to complete the text vectorization process, build a vector index based on FAISS, and construct the vector database.

[0193] Step 4: Select the Llama3-8B local offline lightweight small model as the base model, and use a combination of few-shot learning, LoRA lightweight fine-tuning and transfer learning to complete the model training and optimization, and combine prompt word engineering to complete the instruction adaptation.

[0194] Step 5: Obtain the fault description information input by the user through the human-computer interaction interface, and perform preprocessing operations on the input text.

[0195] Step 6: Convert the fault description text into a data vector using Embedding, perform vector similarity retrieval in the vector database, and recall a preset number of similar historical fault cases.

[0196] Step 7: Combine system commands, retrieved similar cases, and user troubleshooting issues to construct a complete set of prompts.

[0197] Step 8: Input the prompt words into the large language model inference module for inference calculation, and output the fault diagnosis results, which include fault cause analysis, troubleshooting steps suggestions, and reference case information.

[0198] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A high-pressure syringe failure intelligent diagnosis system, characterized in that, It includes a fault case database, a vector database, a large language model inference module, and a human-computer interaction interface; The fault case database is based on several real repair cases, repair logs, repair codes and expert repair knowledge of multi-brand high-pressure injectors, and is formed through text cleaning, classification and structural standardization preprocessing. The vector database uses JSON format to convert the natural language content of the fault case database into a programmable key-value pair data structure, and generates a vector index through FAISS. The large language model inference module adopts the technical architecture of retrieval enhancement generation RAG and large language model LLM, and both it and the human-computer interaction interface are based on the local offline lightweight small model of Llama3-8B. The diagnostic process of the system is as follows: After the user inputs a fault description through the human-computer interaction interface, the system preprocesses the input text, converts the fault phenomenon language into data vectors through Embedding, performs vector similarity retrieval in the vector database, recalls the top K similar cases, constructs prompt words containing system instructions, retrieval cases and user questions, and drives the large language model to generate diagnostic results containing fault cause analysis, troubleshooting steps suggestions and reference case links.

2. The intelligent fault diagnosis system for high-pressure injectors according to claim 1, characterized in that, The fault case database is divided into fault modules and fault types. The fault modules include contrast agent injection module, pressure control module, pipeline system module, motor drive module, mainboard control module, and component matching module. The fault types of the contrast agent injection module include unstable injection flow rate / fluctuation between fast and slow, deviation in injection dosage accuracy, and injection interruption / inability to start injection. The fault types of the pressure control module include excessive injection pressure triggering an alarm, insufficient pressure / failure to build up pressure, and abnormal pressure deviation in the pressure chamber. The fault types of the pipeline system module include pipeline leakage / contrast agent leakage, pipeline blockage / tangling, and air detection error / air intake in the pipeline; The fault types of the motor drive module include motor not turning / power output failure, potentiometer error / abnormal motor speed, and motor overheating / protection mechanism triggering. The fault types of the motherboard control module include power-on failure / system failure to start, irregular fault code errors, and communication interruption with imaging equipment. The fault types of the component supporting modules include syringe installation detection failure and sensor failure / data acquisition abnormality.

3. The intelligent fault diagnosis system for high-pressure injectors according to claim 1, characterized in that, The preprocessing of the fault case database includes data cleaning, data labeling, and structured processing; The data cleaning process is used to remove missing key information, incorrectly labeled data, and duplicate or redundant data; correct inconsistent data formats; desensitize sensitive information; and retain core fields such as fault symptoms, fault causes, repair steps, and precautions. The data annotation and structuring process are used to annotate historical fault data with fault type, fault phenomenon, fault root cause, maintenance steps and precautions, ensuring that each piece of data is complete and unambiguous in classification.

4. The intelligent fault diagnosis system for high-pressure injectors according to claim 1, characterized in that, The process of building the vector database includes: using a BERT pre-trained language model to vectorize the JSON-formatted structured knowledge base for the high-pressure injector fault domain. The specific steps are as follows: The input fault symptoms, fault causes, repair steps and precautions are segmented into words, generating a token sequence, and category and separator tags are added; The token sequence is input into the BERT encoder, which learns contextual information through a multi-head self-attention mechanism and outputs the hidden layer features at each position. The output features corresponding to the classification labels are taken as the semantic sentence vector of the entire text. L2 normalization is performed on the semantic sentence vector.

5. The intelligent fault diagnosis system for high-pressure injectors according to claim 4, characterized in that, The construction of the vector database also includes the following processes: Combine the fault symptoms, fault causes, repair steps, and precautions into a standard input text; The BERT pre-trained language model is used to perform batch inference on all maintenance cases to generate a fault vector set in a unified format. The fault vector and its corresponding metadata are stored in the FAISS vector database, and a cosine similarity index is established to support efficient approximate nearest neighbor search. When a user inputs new fault information, the system vectorizes the query text and performs a similarity match with vectors in the database, returning the most relevant historical cases.

6. The intelligent fault diagnosis system for high-pressure injectors according to claim 1, characterized in that, The large language model inference module is based on a small model built on the local offline lightweight model of Llama3-8B and pre-trained in the field of medical equipment. The module freezes the main backbone network parameters of the base through transfer learning technology, with a freezing ratio of no less than 95%, and only opens the feature extraction layer and attention layer. It transfers common knowledge in the field of general medical equipment maintenance to complete the basic cognitive adaptation of medical equipment maintenance scenarios. A LoRA matrix with rank values ​​of 8-64 is connected to the attention layer of the base. A fine-tuning dataset is constructed using a database of high-pressure injector failure cases. The weight parameters of the LoRA matrix are trained based on the fine-tuning dataset to complete the directional adaptation of high-pressure injector failure diagnosis scenarios.

7. The intelligent fault diagnosis system for high-pressure injectors according to claim 6, characterized in that, The large language model inference module achieves model illusion suppression through the synergistic effect of cue word engineering optimization and retrieval enhancement generation RAG technology, as detailed below: A dedicated prompt word template and small sample constraint logic for the field of high-pressure injector fault diagnosis are constructed. The small sample constraint logic is a technical logic integrated into the prompt word template to constrain the reasoning behavior of the large language model reasoning module. The technical logic limits the large language model reasoning module to conduct reasoning based on the fine-tuned dataset and a limited number of cases. The limited cases are high-pressure injector-related maintenance samples in the vector database. The fault information input mode has been optimized to support three input methods: text description, fault code input, and module selection. Its output format is consistent with the output format of the fine-tuning dataset, so as to achieve precise constraints on the output logic of the large language model inference module. A Retrieval Enhancement Generation (RAG) technology is employed to achieve deep integration between the large language model inference module and the vector database. Specifically, this is implemented by: establishing a data interaction channel between the large language model inference module and the vector database; integrating the RAG technology into the inference process of the large language model inference module, and pre-setting a retrieval trigger mechanism so that when the large language model inference module receives fault input and enters the diagnostic result generation stage, it automatically triggers a vector database retrieval command; the large language model inference module first preprocesses and vectorizes the fault information input by the user, then initiates a similar case retrieval from the vector database to recall relevant maintenance samples; based on the retrieved similar cases and its learned fault diagnosis knowledge, it completes the inference and generation of diagnostic results, achieving deep linkage and adaptation between the two.

8. The intelligent fault diagnosis system for high-pressure injectors according to claim 1, characterized in that, The system adopts a hybrid local deployment environment with a central processing unit and a lightweight graphics processor, and uses the Ollam and TextGenerationWebUI lightweight deployment framework to complete the deployment. Perform INT4 or INT8 quantization compression on the local offline lightweight small model of Llama3-8B.

9. The intelligent fault diagnosis system for high-pressure injectors according to claim 1, characterized in that, The system divides the preprocessed sample data into training, validation, and test sets in a 6:2:2 ratio, and uses leave-one-out cross-validation and k-fold cross-validation to verify the model performance.

10. A high-pressure injector failure intelligent diagnosis method, characterized in that, The intelligent diagnostic system for high-pressure injector faults according to any one of claims 1 to 9 comprises the following steps: Step 1: Collect real-world repair cases, repair logs, fault codes, and expert repair knowledge from different brands of high-pressure injectors to obtain raw fault data; Step 2: Perform text cleaning, information anonymization, and structural standardization preprocessing on the original fault data, divide the equipment functional modules and corresponding fault types, and construct the fault case database; Step 3: Convert the natural language content in the fault case database into a JSON key-value pair data structure, use the BERT pre-trained language model to complete the text vectorization process, build a vector index based on FAISS, and construct the vector database. Step 4: Select the Llama3-8B local offline lightweight small model as the base model, and use a combination of few-shot learning, LoRA lightweight fine-tuning and transfer learning to complete the model training and optimization, and combine prompt word engineering to complete the instruction adaptation. Step 5: Obtain the fault description information input by the user through the human-computer interaction interface, and perform preprocessing operations on the input text; Step 6: Convert the fault description text into a data vector using Embedding, perform vector similarity retrieval in the vector database, and recall a preset number of similar historical fault cases; Step 7: Combine system commands, retrieved similar cases, and user troubleshooting issues to construct complete prompt keywords; Step 8: Input the prompt words into the large language model inference module for inference calculation, and output the fault diagnosis results, which include fault cause analysis, troubleshooting steps suggestions, and reference case information.

Images