Method, device and equipment for navigating diagnosis function based on function path and medium

Abstract

The application relates to the technical field of diagnostic function navigation based on a function path, and discloses a diagnostic function navigation method and device based on a function path, equipment and a medium, wherein the method comprises the following steps: constructing a vector library by previously splitting each engineering system data into function category sub-data and generating a structured abstract vector, so that a recall operation is performed on the pre-built abstract library; controlling the context length by screening a first preset number of recall results to input a prompt word template; and generating an interactive option for user confirmation and execution. The application has the beneficial effects of reducing the data overhead and irrelevant information interference of a large language model, improving the analysis accuracy, shortening the manual screening path, and improving the operation efficiency.

Description

Technical Field

[0001] This application relates to the field of diagnostic function navigation technology based on function paths, and in particular to a diagnostic function navigation method, apparatus, device and medium based on function paths. Background Technology

[0002] Vehicle repair and diagnosis rely on the rapid location and accurate understanding of functional paths within heterogeneous texts such as repair manuals and fault guides to guide technicians in performing corresponding diagnostic operations. Existing solutions introduce vector retrieval technology, which uses an embedding model to map text to a semantic space for similarity matching, and then injects the matching results into a large model to assist in reasoning and decision-making.

[0003] However, existing systems typically merge all functional entries into a single vector library and perform a global traversal retrieval, resulting in high response latency. After retrieval, a large number of unfiltered candidate results are directly injected into the large model, leading to excessively long context, increased overhead, and irrelevant information interfering with the focus on key semantics, resulting in low interaction efficiency. Summary of the Invention

[0004] Therefore, it is necessary to propose a method, device, equipment and medium for diagnostic function navigation based on function path to address the existing problem of diagnostic function navigation based on function path.

[0005] A diagnostic function navigation method based on function paths, the method comprising: Based on the query information input by the user, a recall operation is performed on the preset structured summary vector library, and a first preset number of structured summary vectors are selected as the recall results. The recall results are input into a preset prompt word template to form contextual input information; The context input information is input into a preset large language model to generate at least one functional path corresponding to the query information; Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before the step of performing a recall operation on a preset structured summary vector library based on the query information input by the user, and selecting a first preset number of structured summary vectors as the recall results, the method further includes: Acquire system data from each engineering system, and then split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; A structured summary vector is generated based on the functional category sub-data to construct a structured summary vector library.

[0006] Further, the step of generating structured summary vectors corresponding to the functional categories based on the functional category sub-data to construct a structured summary vector library includes: Keyword extraction is performed on the functional description text of each functional category sub-data. Based on the extracted keywords, the sub-data of each functional category are hierarchically classified according to a preset classification system to obtain the corresponding classification results; Based on the classification results, semantic compression is performed on the functional category sub-data to obtain a structured summary; Each structured summary is encoded into a fixed-dimensional structured summary vector to build a structured summary vector library.

[0007] Further, the step of performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results includes: After receiving the query information input by the user, the text character length and keyword density of the query information are obtained; wherein, the keyword density is the ratio of the number of words in the query information that belong to the preset domain dictionary to the total number of words in the query information; The recall parameters in the recall operation are set according to the text character length and keyword density. Based on the recall parameters, a recall operation is performed, and a first preset number of structured summary vectors are selected as the recall results.

[0008] Further, the step of performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results includes: In each independent vector library of the structured summary vector library, a similarity calculation based on cosine similarity is performed, and the structured summary vectors in each structured summary vector library are sorted from high to low according to their cosine similarity with the query information; wherein, the structured summary vector library includes multiple independent vector libraries corresponding to different engineering systems; Select the second preset number of structured summary vectors from the independent vector libraries corresponding to each engineering system, before sorting. The structured summary vectors selected from the structured summary vector library corresponding to each engineering system are merged and globally sorted from high to low according to cosine similarity. The first preset number of structured summary vectors before global sorting are selected as the recall results.

[0009] Furthermore, before the step of inputting the recall result into a preset prompt word template to form contextual input information, the method further includes: The user-inputted query information is input into a preset intent classifier to obtain a classification result; wherein, the intent classifier classifies the query information into one of a plurality of preset intent categories, and the preset intent categories include at least search step intents, status confirmation intents, and troubleshooting intents; Based on the classification result of the intent classifier, a prompt word template structure matching the classification result is selected from a preset prompt word template library; wherein, the prompt word template library stores at least a three-segment template structure and a five-segment template structure, the three-segment template structure sequentially including a prefix guidance section, a dynamic insertion section, and a suffix inference instruction section, and the five-segment template structure sequentially including a prefix guidance section, a dynamic insertion section, a sensor data context section, a historical similar case summary section, and a suffix inference instruction section; When the classification result is a troubleshooting intent, the five-segment template structure is selected as the prompt word template; when the classification result is a search step intent or a status confirmation intent, the three-segment template structure is selected as the prompt word template.

[0010] Furthermore, after the step of generating corresponding interactive options based on at least one functional path for the user to confirm in order to execute the corresponding diagnostic function, the method further includes: After the user confirms and executes the corresponding diagnostic function, the function path executed this time and its corresponding timestamp are recorded to form a historical operation path record; The historical operation path records are stored in a preset historical operation database.

[0011] Furthermore, before the step of inputting the recall result into a preset prompt word template to form contextual input information, the method further includes: Calculate the semantic similarity between each structured summary in the recall results and each historical operation path record in the historical operation database; Sort the target historical operation path records from high to low according to semantic similarity, and select a third preset number of records. The target historical operation path record is injected into the prompt word template in plain text form, and together with the recall result, forms the context input information.

[0012] A diagnostic function navigation device based on a function path, the device comprising: The filtering module is used to perform a recall operation on a preset structured summary vector library based on the query information input by the user, and to filter a first preset number of structured summary vectors as the recall results; The generation module is used to input the recall results into a preset prompt word template to generate contextual input information; The generation module is used to input the context input information into a preset large language model and generate at least one functional path corresponding to the query information; The confirmation module is used to generate corresponding interactive options based on at least one functional path for the user to confirm in order to execute the corresponding diagnostic function; The acquisition module is used to acquire system data of various engineering systems and split the system data according to a preset structure splitting method to obtain multiple functional category sub-data; The building module is used to generate structured summary vectors for corresponding functional categories based on functional category sub-data, in order to build a structured summary vector library.

[0013] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Based on the query information input by the user, a recall operation is performed on the preset structured summary vector library, and a first preset number of structured summary vectors are selected as the recall results. The recall results are input into a preset prompt word template to form contextual input information; The context input information is input into a preset large language model to generate at least one functional path corresponding to the query information; Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before the step of performing a recall operation on a preset structured summary vector library based on the query information input by the user, and selecting a first preset number of structured summary vectors as the recall results, the method further includes: Acquire system data from each engineering system, and then split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; A structured summary vector is generated based on the functional category sub-data to construct a structured summary vector library.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Based on the query information input by the user, a recall operation is performed on the preset structured summary vector library, and a first preset number of structured summary vectors are selected as the recall results. The recall results are input into a preset prompt word template to form contextual input information; The context input information is input into a preset large language model to generate at least one functional path corresponding to the query information; Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before the step of performing a recall operation on a preset structured summary vector library based on the query information input by the user, and selecting a first preset number of structured summary vectors as the recall results, the method further includes: Acquire system data from each engineering system, and then split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; A structured summary vector is generated based on the functional category sub-data to construct a structured summary vector library.

[0015] The beneficial effects of this application are as follows: By pre-splitting the data of each engineering system into functional category sub-data and generating structured summary vectors to build a vector library, the recall operation is performed on the pre-built summary library, reducing the retrieval time complexity to the constant level, shortening the response latency, and requiring only incremental library building for new systems without rebuilding the full index, thus reducing expansion costs; by filtering the input prompt word templates of the first preset number of recall results to control the context length, the data overhead of the large language model and the interference of irrelevant information are reduced, improving the parsing accuracy; by generating interactive options for users to confirm execution, the manual screening path is shortened, and the operation efficiency is improved. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] in: Figure 1 This is an application environment diagram of a diagnostic function navigation method based on function paths in one embodiment; Figure 2 This is a flowchart of a diagnostic function navigation method based on a function path in one embodiment; Figure 3 This is a structural block diagram of a diagnostic function navigation device based on a functional path in one embodiment; Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] Figure 1 This is a diagram illustrating a diagnostic function navigation application environment based on a function path in one embodiment. (Refer to...) Figure 1 This function path-based diagnostic function navigation method is applied to a function path-based diagnostic function navigation system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 receives query information input by the user, and the server 120 generates at least one function path corresponding to the query information.

[0020] like Figure 2 As shown, in one embodiment, a diagnostic function navigation method based on function paths is provided. This method can be applied to both terminals and servers; this embodiment uses a terminal application as an example. The diagnostic function navigation method based on function paths specifically includes the following steps: S1: Based on the query information input by the user, perform a recall operation on the preset structured summary vector library, and select a first preset number of structured summary vectors as the recall results; S2: Input the recall results into a preset prompt word template to form contextual input information; S3: Input the context input information into a preset large language model to generate at least one functional path corresponding to the query information; S4: Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before step S1, which involves performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results, the method further includes: S001: Obtain system data for each engineering system, and split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; S002: Generate structured summary vectors for the corresponding functional categories based on the functional category sub-data to build a structured summary vector library.

[0021] As described in step S1 above, a recall operation is performed on a preset structured summary vector library based on the user-input query information, selecting a first preset number of structured summary vectors as the recall results. The structured summary vector library refers to a collection of multiple structured summary vectors stored in physically isolated independent vector libraries according to their respective engineering systems. Each independent vector library corresponds to an engineering system (such as HVAC systems, electrical systems, and water supply and drainage systems). Based on the user-input query information, an efficient vector similarity recall operation is performed on the preset structured summary vector library, selecting a first preset number (preferably 5) of structured summary vectors most relevant to the user's query semantics from a massive number of engineering paths as the basic candidate set for subsequent processing. Each summary in the structured summary vector library is encoded into a fixed-dimensional semantic vector using an embedding model (a vectorization model, such as BGE-M3, a general multilingual text embedding model). The recall operation is based on cosine similarity calculation, enabling the mapping of natural language queries and engineering paths to a unified semantic space for mathematical comparison. This overcomes the technical limitation of traditional keyword matching, where synonyms cannot be recalled. Furthermore, this recall operation relies on multiple independent vector libraries physically isolated by engineering system (HVAC, water supply and drainage, electrical, etc.), employing a parallel execution strategy to reduce the overall retrieval complexity from O(N) to O(N / S), thus improving response speed. A structured summary refers to a standardized text fragment generated after keyword extraction, hierarchical classification, and semantic compression of functional category sub-data. Its format is "execution + component + core action" or "detection + component + status," with a length not exceeding 64 tokens (the basic unit for text processing in large language models). A "structured summary vector" refers to this standardized text fragment encoded into a fixed-dimensional semantic vector using an embedding model. An engineering system refers to a physically independent and functionally complete equipment system, including but not limited to heating, ventilation, and air conditioning (HVAC) systems, electrical systems, water supply and drainage systems, power systems, and chassis systems. Each engineering system has its own independent maintenance manual, troubleshooting guide, and other technical documents.

[0022] As described in step S2 above, the recall result is input into a preset prompt word template to form contextual input information. The text content and its metadata (including source system identifier, path encoding, etc.) corresponding to the structured summary vector recalled in step S1 are structured and dynamically assembled according to the preset prompt word template to form lightweight contextual input information that can be directly used by large language models for semantic parsing and path matching. The core technical features of this step are threefold: First, instead of simply concatenating the recalled text and feeding it into the model, a pre-defined template structure assigns a clear semantic role to each recalled summary—including a candidate list sorted in descending order of similarity, a source system identifier added before each summary (such as [HVAC], [Electrical]), and suffix instructions that constrain the output format, ensuring that the large model performs accurate inference within a clear instruction framework; Second, the total number of injected tokens is strictly controlled. Since the first preset quantity K=5 and each structured summary is limited to no more than 64 tokens, the overall number of context tokens is usually no more than 300, i.e., controlled within 10% of the total number of tokens in the entire directory, achieving a compression of over 90% compared to the traditional approach of inputting 2000+ tokens; Third, the system incorporates an intent classifier (based on a finely tuned BERT (Bidirectional Encoder Representations from...)). The Transformers model can adaptively select the template structure based on the query intent—a five-segment template (including sensor data context and historical case summary segments) is used when the intent is identified as "troubleshooting," and a three-segment template is used to minimize overhead when the intent is identified as "find steps" or "confirm status."

[0023] As described in step S3 above, the context input information is input into a preset large language model to generate at least one functional path corresponding to the query information. A functional path refers to the navigation path required from the user-input query information to the execution of the corresponding diagnostic function, including one or more sequentially arranged diagnostic operation steps. Each functional path has a unique code (path code) for unique identification and traceability within the system. The lightweight context input information constructed in step S2 is then fed into the preset large language model. Leveraging the powerful semantic understanding, logical reasoning, and instruction compliance capabilities of the large model, at least one functional path matching the user's query information is accurately generated from the recalled structured summary candidates. The large model does not rely on its own general knowledge to guess or fabricate. Instead, it strictly uses the structured summary candidate list injected in step S2 and auxiliary context (sensor data, historical cases) to perform limited and controllable comparison and selection reasoning. Unlike conventional schemes that directly allow the large model to retrieve from the entire knowledge base, this step compresses the candidate set from tens of thousands to a maximum of 5. The large model only needs to complete path matching judgment within a very small candidate space, reducing the amount of reasoning computation and shortening the generation delay. At the same time, it avoids the illusion problem caused by the overload of irrelevant information. In order to ensure the controllability and consistency of the output, the generation process is constrained by the content of the suffix reasoning instruction section in step S2 (such as "Please analyze possible causes according to the causal chain and recommend the optimal detection order"). The large model is required to output structured results in a specified format (such as "Path code: xxx, Reason: xxx") to facilitate subsequent steps of parsing and interactive presentation. For example, after receiving the complete context constructed in step S2, the large language model performs internal reasoning: the user query "poor air conditioning cooling effect" and "detect refrigerant pressure" (similarity 0.92) have the highest semantic match. Combined with the sensor data showing a high-pressure side pressure of only 0.8 MPa (significantly lower than the normal value of 1.2-1.6 MPa), and referring to the historical record that the user successfully located the leak point two days ago by "detecting refrigerant pressure," the model comprehensively judges that the most matching path is "detect refrigerant pressure." The model follows the format constraints of the suffix instruction and outputs a structured result: "Functional path code: HVAC-REFRIG-PRESS-TEST; Recommendation reason: The high-pressure side pressure is significantly low, which is highly consistent with your recent successfully handled refrigerant leak case. It is recommended to prioritize this detection." A total of one functional path is output. If configured to allow multiple solutions, the model can also output a second candidate path, "check expansion valve opening," with the supplementary explanation "if the refrigerant pressure is normal, this path is an alternative," providing more options for subsequent interactions.

[0024] As described in step S4 above, corresponding interactive options are generated based on at least one functional path for the user to confirm in order to execute the corresponding diagnostic function. A diagnostic function refers to an operation performed on a device or component in an engineering system, including inspection, testing, adjustment, calibration, troubleshooting, and replacement. Each diagnostic function corresponds to a functional path. One or more functional paths generated in step S3 are transformed into graphical or textual interactive options (such as buttons, cards, or lists) and presented to the user for confirmation. After the user selects and confirms, the corresponding diagnostic function is triggered, thus completing a full functional loop from semantic retrieval to physical execution. In this invention, a functional path refers to the navigation path required from the user-input query information to the execution of the corresponding diagnostic function, including one or more sequentially arranged diagnostic operation steps. Instead of simply listing path options, it integrates a "personalized matching solution generation unit" that calculates a confidence score for each candidate path based on multiple dimensions—including vector similarity score, contextual matching entropy value, historical adoption rate, and operational risk level from functional attribute tags. Based on this, it generates personalized recommendation tags with explanations (e.g., "Recommended solution, confidence 0.91, highly matches similar faults you recently handled and does not require disassembling high-voltage components"), enabling users to make safe choices with full knowledge. It also supports user-specified preference constraints (e.g., "prioritize short-duration operations" or "avoid steps requiring special tools"), which are fed back to the prompt enhancement layer for secondary reasoning, achieving configurable personalized path matching. In terms of interactive presentation, it adaptively adjusts to different terminal forms—displayed as rich interactive cards on tablets or PCs, and as voice broadcasts with simple options on smart glasses or voice terminals.

[0025] As described in step S001 above, system data for each engineering system is acquired, and the system data is then split according to a preset structural splitting method to obtain multiple functional category sub-data. The original system data for each engineering system is acquired, and these unstructured long texts are logically segmented and reorganized according to the preset structural splitting method to obtain multiple fine-grained functional category sub-data. System-level physical isolation strictly distinguishes the data sources of different professional systems such as HVAC, water supply and drainage, and electrical systems, and splits them independently to ensure that each functional category sub-data carries a clear system affiliation label from its inception. The preset structural splitting method is not a simple text truncation or word-by-word segmentation, but rather a logical aggregation based on a domain knowledge-driven "three-level or four-level classification system" (such as "system → subsystem → component → operation type"), ensuring that each split sub-data corresponds to a minimum executable atomic functional unit, rather than redundant paragraph fragments. To achieve accurate segmentation, specialized terminology dictionaries for each discipline (such as the HVAC terminology database containing "fan coil unit" and "fresh air handling unit", and the electrical terminology database containing "relay" and "contactor") and chapter structure templates (such as identifying "1, 2, 3" level headings as segmentation boundaries) are loaded. Combined with regular expression matching and heading level parsing, each chapter of the original manual is accurately mapped to the corresponding classification system.

[0026] As described in step S002 above, structured summary vectors corresponding to the functional category are generated based on the functional category sub-data to construct a structured summary vector library. The functional category sub-data generated in S001 is further transformed into structured summary vectors with high information density, unified format, and semantic computability. These vectors are then stored in physically isolated vector libraries according to system affiliation, ultimately forming a structured summary vector library for efficient retrieval in step S1. "Structured summary generation" uses a three-step process of keyword extraction, hierarchical classification, and semantic compression to compress the original lengthy functional description into a standard summary of no more than 64 tokens, retaining the core verb-object structure (such as "detecting the clutch relay engagement status of the compressor") and removing non-core modifying information such as environmental conditions and safety warnings, achieving high information density expression. "Physically isolated storage" uses the system tags assigned in S001 to write the summary vectors of different systems into different partitions or collections of the distributed vector database, achieving complete physical isolation and providing architectural support for reducing O(N) to O(N / S) during subsequent parallel retrieval. Each summary vector is also stored with bidirectional mapping metadata—including the original path encoding and the original document full-text link, for example, stored in JSON-LD (JavaScript Object Notation for Linked Data) format—ensuring that the vector retrieval results can be losslessly restored to readable complete path information.

[0027] In one embodiment, the step S002 of generating a structured summary vector corresponding to the functional category according to the functional category sub-data to construct a structured summary vector library includes: S0021: performing keyword extraction on the function description text of each functional category sub-data; S0022: based on the extracted keywords, performing hierarchical classification on each of the functional category sub-data according to a preset classification system to obtain a corresponding classification result; S0023: performing semantic compression on the functional category sub-data based on the classification result to obtain a structured summary; S0024: encoding each structured summary into a structured summary vector with a fixed dimension, so as to construct a structured summary vector library.

[0028] As described in the above step S0021, keyword extraction is performed on the function description text of each functional category sub-data; this step is a pre-filter for semantic compression, and its technical purpose is to extract keyword elements representing "core action" and "action object" from the lengthy and noise-mixed function description text, so as to provide structured input for subsequent hierarchical classification. Since vehicle maintenance texts have an extremely high term density (such as "duty cycle", "solenoid valve", "pressure relief"), general extractors tend to incorrectly segment professional compound words (for example, segmenting "exhaust gas recirculation valve" into "exhaust gas / recirculation"), resulting in loss of key semantics. By loading a pre-constructed vehicle engineering term dictionary (covering powertrain, chassis, electrical and other subsystems) and a fault mode library, combined with part-of-speech tagging and dependency syntax analysis, verbs (representing operation intentions, such as "calibrate", "evacuate", "turn on") and nouns (representing components or parameters, such as "crankshaft position sensor", "coolant temperature") are preferentially locked, irrelevant stop words such as "de", "le", "ranhou" are effectively excluded, ensuring that the extraction result can truly reflect the operation essence of the entry.

[0029] As described in the above step S0022, based on the extracted keywords, each of the functional category sub-data is hierarchically classified according to a preset classification system to obtain a corresponding classification result. The unordered keywords are substituted into the unique "four-level classification system" of the present application (that is: system → subsystem → component → operation type) to realize structured mapping. This classification process does not rely on manual labeling one by one, but adopts a hybrid strategy based on a rule engine and a lightweight classification model: firstly, a first-level rough screening is performed by using the system feature words in the keywords (for example, "compressor" tends to be classified into the "HVAC / refrigeration subsystem"); then, the operation type words (such as "calibrate", "demarcate") are combined to further subdivide to the fourth level. This hierarchical anchoring enables the subsequently generated vectors to naturally carry the "affiliated system" and "operation nature" labels, which not only provides a basis for physical isolation during parallel retrieval, but also provides a data source for the "source system identifier" (such as [Electrical]) in the large model prompt.

[0030] As described in step S0023 above, semantic compression is performed on the functional category sub-data based on the classification results to obtain a structured summary. "High information density and high fidelity compression" requires both reducing the number of words to decrease the overhead of large model tokens and fully preserving the three elements of "subject (who)," "action (what to do)," and "object (to what)," without causing semantic collapse. This step abandons the randomness of traditional generative summarization and adopts a "template-based slot filling" strategy, that is, pre-setting slots for [subject / system], [component], [core action], and [supplementary attribute], and forcibly injecting the results of S0021 and S0022 into fixed sentence patterns (such as "execution + component + action" or "detection + component + state").

[0031] As described in step S0024 above, each structured summary is encoded into a fixed-dimensional structured summary vector to construct a structured summary vector library. Encoding can be performed using an embedding model (such as BGE-M3 or M3E embedding model (M3EEmbedding)), because these models can capture deep semantic relationships in the summary text through deep Transformers (e.g., mapping "draining coolant" and "replacing antifreeze" to similar vector spaces, even if the words do not completely overlap). The encoded output dimension is fixed (e.g., 1024 dimensions), ensuring a uniform vector size across the entire library. During library construction, based on the first-level system tags in S0022, the vectors are written into different partitions or collections of the distributed vector database, and the path encoding and original document link are carried in the metadata.

[0032] In one embodiment, step S1, which involves performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results, includes: S101: After receiving the query information input by the user, obtain the text character length and keyword density of the query information; wherein, the keyword density is the ratio of the number of words in the query information that belong to the preset domain dictionary to the total number of words in the query information; S102: Set the recall parameters in the recall operation according to the text character length and keyword density; S103: Perform a recall operation based on the recall parameters and select a first preset number of structured summary vectors as the recall results.

[0033] As described in step S101 above, after receiving the query information input by the user, the text character length and keyword density of the query information are obtained. The keyword density is the ratio of the number of words in the query information belonging to a preset domain dictionary to the total number of words in the query information. This step belongs to the pre-processing signal acquisition and feature engineering stage. Its technical purpose is to transform the user's original natural language query into two quantifiable statistical indicators—text character length (L) and keyword density (D)—providing a numerical basis for the dynamic adjustment of subsequent retrieval thresholds. This density value strongly depends on the preset domain dictionary (e.g., covering vehicle engineering terms such as "thermostat," "duty cycle," and "pressure relief valve"), and its calculation method is the ratio of "the number of words in the query information belonging to the preset domain dictionary" to "the total number of words in the query information." This design enables the system to distinguish between "short sentences full of technical terms" (e.g., "EGR valve stuck," with a high D value) and "long sentences mostly containing colloquialisms" (e.g., "how to deal with that thing in front of that car," with a low D value), thereby avoiding misjudgments caused by a single length indicator and laying the foundation for subsequent precise policy implementation.

[0034] As described in step S102 above, the recall parameters in the recall operation are set according to the text character length and keyword density. The quantized features extracted in S101 are mapped to specific retrieval strategy parameters. In existing fixed-threshold retrieval, the cosine similarity threshold (e.g., fixed at 0.75) is a static constant and cannot cope with drastic fluctuations in query semantic density. This step dynamically adjusts the recall parameters based on the value ranges of L and D using a built-in rule engine or lightweight decision model. The core parameter is the cosine similarity recall threshold (θ). In addition, the K value of Top-K or the number of parallel recall threads can also be adjusted in conjunction. Its adjustment logic follows the engineering principle of "low threshold to ensure recall and high threshold to improve accuracy": when the query is very short and the term density is low (semantic ambiguity), the system determines that the user cannot provide enough constraint information. At this time, the threshold should be lowered to expand the candidate coverage. It is better to recall more weakly relevant items for secondary screening by the large model than to miss the unique correct answer. When the query is long and the term density is high (semantic clarity), the system determines that the user's intent is clear. At this time, the threshold should be raised to filter noise and only send high-confidence candidates to the prompt words to ensure that the large model is not interfered with by generalization information.

[0035] As described in step S103 above, a recall operation is performed based on the recall parameters, and a first preset number of structured summary vectors are selected as the recall results. The recall parameters determined in S102 are applied to the actual vector retrieval process, and the first preset number of structured summary vectors are finally output as the recall results. Specifically, the user query text is encoded into a query vector using the same embedding model (such as BGE-M3), and then the similarity search interface of the vector database is called, with the dynamic threshold θ and K value set in S102 passed in. This search process is executed in parallel on all physically isolated system vector libraries. After each library returns its local Top-K results, they are then uniformly merged and sorted globally, and finally only the summary vector IDs and their metadata with similarity scores ≥ θ and serial numbers in the top K positions are retained. Transforming the complex feature analysis and parameter decisions in the early stages into actual retrieval actions ensures the dynamic adaptability of the retrieval strategy. This allows the system to tolerantly recall multiple possibilities such as "lighting switch", "water valve switch", and "power switch" when faced with extremely simple fuzzy queries like "which switch", while sharply focusing on long and precise queries like "troubleshooting steps for main relay sticking fault in power battery pack", returning only highly relevant entries with a score >0.85.

[0036] In one embodiment, step S1, which involves performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results, includes: S111: In each independent vector library of the structured summary vector library, a similarity calculation based on cosine similarity is performed, and the structured summary vectors in each structured summary vector library are sorted from high to low according to their cosine similarity with the query information; wherein, the structured summary vector library includes multiple independent vector libraries corresponding to different engineering systems; S112: Select the second preset number of structured summary vectors from the independent vector libraries corresponding to each engineering system, before sorting; S113: Merge the structured summary vectors selected from the structured summary vector library corresponding to each engineering system, and sort them globally from high to low according to cosine similarity. Select the first preset number of structured summary vectors before global sorting as the recall result.

[0037] As described in step S111 above, in each independent vector library of the structured summary vector library, cosine similarity calculation is performed, and the structured summary vectors in each structured summary vector library are sorted from high to low according to their cosine similarity with the query information. The structured summary vector library includes multiple independent vector libraries corresponding to different engineering systems. The user query vector is simultaneously distributed to multiple physically isolated system-level vector libraries, and within each library, full or index-accelerated calculation based on cosine similarity is performed independently, with independent sorting from high to low within each library. Separate execution embodies parallelism, meaning that the computation threads of each library do not interfere with each other and proceed synchronously. From an architectural perspective, the total latency is compressed from the sum of the times consumed by each library to the maximum value of the times consumed by each library (Max-Time). Cosine similarity, as a metric, effectively characterizes the degree of convergence in semantic direction by calculating the cosine value of the angle between the query vector and each summary vector in the library in high-dimensional space (range [-1,1]), which is better than the dependence of Euclidean distance on the vector magnitude. "Separate sorting" means that each system library will generate an independent sequence list that only represents the relative relevance within its own system. This local sorting result is not mixed across systems, providing a data foundation for subsequent steps of truncation according to system quotas.

[0038] As described in step S112 above, a second preset number of structured summary vectors are selected from the independent vector libraries corresponding to each engineering system. From each independently sorted vector library, the top-ranked structured summary vectors are truncated according to a unified quota (i.e., the second preset number, denoted as M) to form the "local elite candidate set" for each system. A quota M (M must be greater than the final global K, for example, M=10 or 15) is introduced as a "buffer pool"; the value of M is strictly set less than the total number of entries in each library. By pruning in advance, the computational complexity of subsequent global merging is drastically reduced from O(N_total) to O(S×M) (S is the number of systems), further compressing the computational load of global sorting; this truncating operation is a pure logical data extraction that does not change the underlying vector index structure, so the execution speed is extremely fast and almost no additional I / O latency is generated.

[0039] As described in step S113 above, the structured summary vectors selected from the structured summary vector libraries corresponding to each engineering system are merged and globally sorted from high to low according to cosine similarity. The first preset number of structured summary vectors before global sorting are selected as the recall results. The local elite candidates extracted from each independent vector library in S112 are merged into a unified global candidate pool and globally re-sorted across systems. Finally, the first preset number (denoted as K, preferably K=5) of structured summary vectors are selected as the final recall results of the entire S1 step. Although it ensures that each system is representative, there may be systematic biases in the similarity score scale between different systems (e.g., the overall similarity of the electrical library is generally high), so it is not possible to simply compare across systems based on the original scores. This global sorting result provides a clear "descending similarity sorted" input order for the subsequent prompt word injection, enabling the large model to prioritize the most relevant candidate paths when reading prompt words.

[0040] In one embodiment, before step S2, which involves inputting the recall result into a preset prompt word template to form contextual input information, the method further includes: S121: Input the query information input by the user into a preset intent classifier to obtain a classification result; wherein, the intent classifier classifies the query information into one of a plurality of preset intent categories, and the preset intent categories include at least search step intent, confirmation status intent, and troubleshooting intent; S122: Based on the classification result of the intent classifier, select a prompt word template structure that matches the classification result from a preset prompt word template library; wherein, the prompt word template library stores at least a three-segment template structure and a five-segment template structure, the three-segment template structure sequentially including a prefix guidance section, a dynamic insertion section, and a suffix inference instruction section, and the five-segment template structure sequentially including a prefix guidance section, a dynamic insertion section, a sensor data context section, a historical similar case summary section, and a suffix inference instruction section; S123: When the classification result is a troubleshooting intent, the five-segment template structure is selected as the prompt word template; when the classification result is a search step intent or a status confirmation intent, the three-segment template structure is selected as the prompt word template.

[0041] As described in step S121 above, the user-input query information is input into a preset intent classifier to obtain a classification result. The intent classifier categorizes the query information into one of several preset intent categories, which include at least three types: search steps, status confirmation, and troubleshooting. The user-inputted natural language query is mapped to predefined intent category labels, providing a basis for subsequent template selection. The classifier is preferably a small model (approximately 110M parameters) fine-tuned based on a pre-trained BERT architecture. Compared to directly using a large model for classification, the classifier is trained using supervised learning, and the labeled dataset comes from real-world text such as historical repair work orders and technical Q&A logs. The pre-defined intent categories include at least three types: First, "search steps" (e.g., "how to change brake fluid"), characterized by procedural words such as "how," "steps," and "methods" in the query; second, "confirmation status" (e.g., "what does it mean when the ABS light is on"), characterized by judgmental words such as "whether," "is it normal," and "status" in the query; and third, "troubleshooting" (e.g., "how to check if the air conditioner is not cooling"), characterized by negative conditional words such as "no," "abnormal," "fault," and "failure" in the query. The category label with the highest probability output by the classifier is the classification result, which directly affects the subsequent prompt word assembly strategy.

[0042] As described in step S122 above, based on the classification result of the intent classifier, a prompt word template structure matching the classification result is selected from a preset prompt word template library. The prompt word template library stores at least three-segment and five-segment template structures. The three-segment template structure sequentially includes a prefix guidance segment, a dynamic insertion segment, and a suffix inference instruction segment. The five-segment template structure sequentially includes a prefix guidance segment, a dynamic insertion segment, a sensor data context segment, a historical similar case summary segment, and a suffix inference instruction segment. Based on the intent tag output in S121, a matching template structure is selected from the preset prompt word template library to determine which information segments the final prompt word contains. The three-segment template includes ① a prefix guidance section (used to set the model's role and task objectives), ② a dynamic insertion section (used to inject the Top-K summary list for recall), and ③ a suffix inference instruction section (used to constrain the output format and inference direction). This structure is suitable for routine operations with clear linear processes and simple status confirmations. The five-segment template adds ④ a sensor data context section (used to inject real-time parameters of the current vehicle operation, such as "the water temperature sensor reading is consistently higher than 98°C") and ⑤ a historical similar case summary section (used to inject successful path records of handling similar problems in the past). This structure is suitable for complex fault diagnosis scenarios that require comprehensive judgment of multiple factors.

[0043] As described in step S123 above, when the classification result is a troubleshooting intent, the five-segment template structure is selected as the prompt word template; when the classification result is a search step intent or a status confirmation intent, the three-segment template structure is selected as the prompt word template. Explicit conditional branching logic clarifies the mapping relationship between different intent categories and template structures, providing deterministic execution rules for the selection action in S122. The specific conditions are as follows: When the classification result is "troubleshooting intent", the system forcibly selects a five-segment template structure. This is because troubleshooting tasks typically involve complex logical chains with multiple causes and a single effect, requiring the injection of real-time sensor parameters (such as temperature, pressure, and voltage readings) and historical similar cases (such as "the last similar phenomenon was caused by relay sticking") to assist the large model in causal reasoning and priority ranking. When the classification result is "find step intent" or "confirm status intent", the system selects a three-segment template structure. This is because these two types of tasks are usually linear and highly deterministic processes (such as "changing steps" being a fixed sequence, and "status confirmation" being a binary judgment). Injecting additional information would dilute key semantics and increase unnecessary token overhead and inference latency. This conditional selection logic is an important manifestation of the "lightweight" design principle of this application—allocating differentiated computational resources to tasks of different complexities, rather than applying the same heavy inference process to all queries, thereby achieving Pareto optimality in overall system performance.

[0044] In one embodiment, after step S4, which generates corresponding interactive options based on at least one functional path for the user to confirm in order to execute the corresponding diagnostic function, the method further includes: S501: After the user confirms and executes the corresponding diagnostic function, record the function path executed this time and its corresponding timestamp to form a historical operation path record; S502: Store the historical operation path records in a preset historical operation database.

[0045] As described in step S501 above, after the user confirms and executes the corresponding diagnostic function, the executed function path and its corresponding timestamp are recorded to form a historical operation path record. When the user confirms and executes a diagnostic function path, the core information of the operation—including the executed path identifier (such as path code or path name) and the timestamp of the execution time—is automatically extracted and combined into a structured historical operation path record. The record fields include at least two core dimensions: "function path identifier" and "timestamp." The timestamp is usually accurate to the second (e.g., 2026-07-08 14:32:18) and is used for subsequent calculation of time decay weights. In addition, as a preferred option, the execution result status (e.g., "execution successful" or "execution failed") can also be recorded to provide a basis for the effect evaluation in the subsequent historical case summary.

[0046] As described in step S502 above, the historical operation path record is stored in a preset historical operation database. The instantaneous record object generated in S501 is written into the preset historical operation database, making it a long-term usable persistent data asset. "Independent storage" means that the historical operation database is physically and logically completely separated from the structured summary vector library constructed in S002—the former stores behavioral logs (time-series data), while the latter stores semantic vectors (high-dimensional mathematical space data). Regarding storage media selection, this application prefers memory-based caching databases such as Redis (Remote Dictionary Server, an in-memory database) and Memcached (a distributed memory object caching system), or distributed key-value storage systems, rather than relational databases, to support extremely high write throughput (without complex transactions) and ordered set operations based on timestamps or scores, facilitating fast sorting and Top-N truncation by time decay weights in subsequent steps. Simultaneously, each record can have its lifespan set, for example, automatically discarding old data older than 30 days to control storage costs.

[0047] In one embodiment, before step S2, which involves inputting the recall result into a preset prompt word template to form contextual input information, the method further includes: S131: Calculate the semantic similarity between each structured summary in the recall result and each historical operation path record in the historical operation database; S132: Sort by semantic similarity from high to low, and select a third preset number of target historical operation path records; S133: The target historical operation path record is injected into the prompt word template in plain text form, and together with the recall result, it forms the context input information.

[0048] As described in step S131 above, the semantic similarity between each structured summary in the recall result and each historical operation path record in the historical operation database is calculated. The recall result (Top-K structured summary vector set) obtained in step S1 is compared one by one with each historical operation path record stored in the historical operation database. By calculating the semantic similarity between the two, the relevance of each historical record to the current user query context is quantitatively evaluated.

[0049] As described in step S132 above, the historical operation path records are sorted from high to low semantic similarity, and a third preset number of target historical operation path records are selected. The historical operation path records that have been calculated and marked with similarity scores (all higher than the threshold of 0.72) in S131 are sorted in descending order of similarity from high to low, and the top three preset number (denoted as P, preferably 2 records) of records are selected as the "target historical operation path records" for the final injection of prompt words.

[0050] As described in step S133 above, the target historical operation path record is injected into the prompt word template in plain text form, forming the context input information together with the recall result. The target historical operation path record selected in S132 is inserted into the designated section of the prompt word template in plain text natural language form, forming a complete context input information together with the recall result of S1, for the large model to perform comprehensive reasoning. The injection position of "injecting in plain text form" depends on the template structure selected in S122: if it is a three-segment template, the historical record is usually appended before or after the suffix reasoning instruction; if it is a five-segment template, it is filled into the "historical similar case summary section". Each injected historical record can be appended with execution time information, enabling the large model to perceive the time decay factor (recent experience has higher weight).

[0051] Reference Figure 3 This application provides a diagnostic function navigation device based on a functional path, the device comprising: The filtering module 902 is used to perform a recall operation on a preset structured summary vector library based on the query information input by the user, and to filter a first preset number of structured summary vectors as the recall results; The generation module 904 is used to input the recall result into a preset prompt word template to generate contextual input information; The generation module 906 is used to input the context input information into a preset large language model and generate at least one functional path corresponding to the query information. The confirmation module 908 is used to generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; The acquisition module 910 is used to acquire system data of each engineering system and split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; Module 912 is used to generate structured summary vectors for corresponding functional categories based on functional category sub-data, in order to build a structured summary vector library.

[0052] In one embodiment, the construction module 912 includes: The extraction submodule is used to extract keywords from the functional description text of each functional category subdata. The classification submodule is used to classify the sub-data of each functional category according to a preset classification system based on the extracted keywords, and obtain the corresponding classification results. The compression submodule is used to perform semantic compression on the functional category subdata based on the classification results to obtain a structured summary; The construction submodule is used to encode each structured summary into a fixed-dimensional structured summary vector to build a structured summary vector library.

[0053] In one embodiment, the filtering module 902 includes: The receiving submodule is used to obtain the text character length and keyword density of the query information after receiving the query information input by the user; wherein, the keyword density is the ratio of the number of words in the query information that belong to the preset domain dictionary to the total number of words in the query information; The setting submodule is used to set the recall parameters in the recall operation based on the text character length and keyword density. The filtering submodule is used to perform a recall operation based on the recall parameters and filter a first preset number of structured summary vectors as the recall results.

[0054] In one embodiment, the filtering module 902 includes: The first sorting submodule is used to perform cosine similarity calculations in each independent vector library of the structured summary vector library, and sort the structured summary vectors in each structured summary vector library from high to low according to their cosine similarity with the query information; wherein, the structured summary vector library includes multiple independent vector libraries corresponding to different engineering systems; The selection submodule is used to select the second preset number of structured summary vectors from the independent vector libraries corresponding to each engineering system, based on the order of sorting. The second sorting submodule is used to merge the structured summary vectors selected from the structured summary vector library corresponding to each engineering system, and sort them globally from high to low according to cosine similarity, and select the first preset number of structured summary vectors before global sorting as the recall result.

[0055] In one embodiment, the diagnostic function navigation device based on the function path further includes: The input module is used to input the query information input by the user into a preset intent classifier to obtain a classification result; wherein, the intent classifier classifies the query information into one of a plurality of preset intent categories, and the preset intent categories include at least search step intents, status confirmation intents, and troubleshooting intents; The first selection module is used to select a prompt word template structure that matches the classification result from a preset prompt word template library based on the classification result of the intent classifier; wherein, the prompt word template library stores at least a three-segment template structure and a five-segment template structure, wherein the three-segment template structure sequentially includes a prefix guidance section, a dynamic insertion section, and a suffix inference instruction section, and the five-segment template structure sequentially includes a prefix guidance section, a dynamic insertion section, a sensor data context section, a historical similar case summary section, and a suffix inference instruction section; The second selection module is used to select the five-segment template structure as the prompt word template when the classification result is a troubleshooting intent; and to select the three-segment template structure as the prompt word template when the classification result is a search step intent or a status confirmation intent.

[0056] In one embodiment, the diagnostic function navigation device based on the function path further includes: The recording module is used to record the function path executed this time and its corresponding timestamp after the user confirms and executes the corresponding diagnostic function, forming a historical operation path record; The storage module is used to store the historical operation path records into a preset historical operation database.

[0057] In one embodiment, the diagnostic function navigation device based on the function path further includes: The calculation module is used to calculate the semantic similarity between each structured summary in the recall result and each historical operation path record in the historical operation database; The third selection module is used to sort the target historical operation path records from high to low according to semantic similarity and select a third preset number of records. The injection module is used to inject the target historical operation path record into the prompt word template in plain text form, which together with the recall result forms the context input information.

[0058] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a function path-based diagnostic function navigation method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a function path-based diagnostic function navigation method. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0059] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Based on the query information input by the user, a recall operation is performed on the preset structured summary vector library, and a first preset number of structured summary vectors are selected as the recall results. The recall results are input into a preset prompt word template to form contextual input information; The context input information is input into a preset large language model to generate at least one functional path corresponding to the query information; Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before the step of performing a recall operation on a preset structured summary vector library based on the query information input by the user, and selecting a first preset number of structured summary vectors as the recall results, the method further includes: Acquire system data from each engineering system, and then split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; A structured summary vector is generated based on the functional category sub-data to construct a structured summary vector library.

[0060] By pre-splitting the data of each engineering system into functional category sub-data and generating structured summary vectors to build a vector library, the recall operation is performed on the pre-built summary library, reducing the retrieval time complexity to constant level, shortening the response latency, and requiring only incremental library building for new systems without rebuilding the full index, thus reducing expansion costs; by filtering the input prompt word templates of the first preset number of recall results to control the context length, the data overhead of the large language model and the interference of irrelevant information are reduced, improving parsing accuracy; by generating interactive options for user confirmation, the manual screening path is shortened, and the operation efficiency is improved.

[0061] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps: Based on the query information input by the user, a recall operation is performed on the preset structured summary vector library, and a first preset number of structured summary vectors are selected as the recall results. The recall results are input into a preset prompt word template to form contextual input information; The context input information is input into a preset large language model to generate at least one functional path corresponding to the query information; Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before the step of performing a recall operation on a preset structured summary vector library based on the query information input by the user, and selecting a first preset number of structured summary vectors as the recall results, the method further includes: Acquire system data from each engineering system, and then split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; A structured summary vector is generated based on the functional category sub-data to construct a structured summary vector library.

[0062] By pre-splitting the data of each engineering system into functional category sub-data and generating structured summary vectors to build a vector library, the recall operation is performed on the pre-built summary library, reducing the retrieval time complexity to constant level, shortening the response latency, and requiring only incremental library building for new systems without rebuilding the full index, thus reducing expansion costs; by filtering the input prompt word templates of the first preset number of recall results to control the context length, the data overhead of the large language model and the interference of irrelevant information are reduced, improving parsing accuracy; by generating interactive options for user confirmation, the manual screening path is shortened, and the operation efficiency is improved.

[0063] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0064] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A diagnostic function navigation method based on functional paths, characterized in that, The method includes: Based on the query information input by the user, a recall operation is performed on the preset structured summary vector library, and a first preset number of structured summary vectors are selected as the recall results. The recall results are input into a preset prompt word template to form contextual input information; The context input information is input into a preset large language model to generate at least one functional path corresponding to the query information; Generate corresponding interactive options based on at least one functional path for the user to confirm, so as to execute the corresponding diagnostic function; Before the step of performing a recall operation on a preset structured summary vector library based on the query information input by the user, and selecting a first preset number of structured summary vectors as the recall results, the method further includes: Acquire system data from each engineering system, and then split each system data according to a preset structure splitting method to obtain multiple functional category sub-data; A structured summary vector is generated based on the functional category sub-data to construct a structured summary vector library.

2. The diagnostic function navigation method based on functional paths according to claim 1, characterized in that, The step of generating structured summary vectors corresponding to functional categories based on functional category sub-data to construct a structured summary vector library includes: Keyword extraction is performed on the functional description text of each functional category sub-data. Based on the extracted keywords, the sub-data of each functional category are hierarchically classified according to a preset classification system to obtain the corresponding classification results; Based on the classification results, semantic compression is performed on the functional category sub-data to obtain a structured summary; Each structured summary is encoded into a fixed-dimensional structured summary vector to build a structured summary vector library.

3. The diagnostic function navigation method based on functional paths according to claim 1, characterized in that, The step of performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results includes: After receiving the query information input by the user, the text character length and keyword density of the query information are obtained; wherein, the keyword density is the ratio of the number of words in the query information that belong to the preset domain dictionary to the total number of words in the query information; The recall parameters in the recall operation are set according to the text character length and keyword density. Based on the recall parameters, a recall operation is performed, and a first preset number of structured summary vectors are selected as the recall results.

4. The diagnostic function navigation method based on functional paths according to claim 1, characterized in that, The step of performing a recall operation on a preset structured summary vector library based on user-input query information and selecting a first preset number of structured summary vectors as the recall results includes: In each independent vector library of the structured summary vector library, a similarity calculation based on cosine similarity is performed, and the structured summary vectors in each structured summary vector library are sorted from high to low according to their cosine similarity with the query information; wherein, the structured summary vector library includes multiple independent vector libraries corresponding to different engineering systems; Select the second preset number of structured summary vectors from the independent vector libraries corresponding to each engineering system, before sorting. The structured summary vectors selected from the structured summary vector library corresponding to each engineering system are merged and globally sorted from high to low according to cosine similarity. The first preset number of structured summary vectors before global sorting are selected as the recall results.

5. The diagnostic function navigation method based on functional paths according to claim 1, characterized in that, Before the step of inputting the recall result into a preset prompt word template to form contextual input information, the method further includes: The user-inputted query information is input into a preset intent classifier to obtain a classification result; wherein, the intent classifier classifies the query information into one of a plurality of preset intent categories, and the preset intent categories include at least search step intents, status confirmation intents, and troubleshooting intents; Based on the classification result of the intent classifier, a prompt word template structure matching the classification result is selected from a preset prompt word template library; wherein, the prompt word template library stores at least a three-segment template structure and a five-segment template structure, the three-segment template structure sequentially including a prefix guidance section, a dynamic insertion section, and a suffix inference instruction section, and the five-segment template structure sequentially including a prefix guidance section, a dynamic insertion section, a sensor data context section, a historical similar case summary section, and a suffix inference instruction section; When the classification result is a troubleshooting intent, the five-segment template structure is selected as the prompt word template; when the classification result is a search step intent or a status confirmation intent, the three-segment template structure is selected as the prompt word template.

6. The diagnostic function navigation method based on functional paths according to claim 1, characterized in that, After the step of generating corresponding interactive options based on at least one functional path for the user to confirm in order to execute the corresponding diagnostic function, the method further includes: After the user confirms and executes the corresponding diagnostic function, the function path executed this time and its corresponding timestamp are recorded to form a historical operation path record; The historical operation path records are stored in a preset historical operation database.

7. The diagnostic function navigation method based on functional paths according to claim 6, characterized in that, Before the step of inputting the recall result into a preset prompt word template to form contextual input information, the method further includes: Calculate the semantic similarity between each structured summary in the recall results and each historical operation path record in the historical operation database; Sort the target historical operation path records from high to low according to semantic similarity, and select a third preset number of records. The target historical operation path record is injected into the prompt word template in plain text form, and together with the recall result, forms the context input information.

8. A diagnostic function navigation device based on functional paths, characterized in that, The device includes: The filtering module is used to perform a recall operation on a preset structured summary vector library based on the query information input by the user, and to filter a first preset number of structured summary vectors as the recall results; The generation module is used to input the recall results into a preset prompt word template to generate contextual input information; The generation module is used to input the context input information into a preset large language model and generate at least one functional path corresponding to the query information; The confirmation module is used to generate corresponding interactive options based on at least one functional path for the user to confirm in order to execute the corresponding diagnostic function; The acquisition module is used to acquire system data of various engineering systems and split the system data according to a preset structure splitting method to obtain multiple functional category sub-data; The building module is used to generate structured summary vectors for corresponding functional categories based on functional category sub-data, in order to build a structured summary vector library.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the functional path-based diagnostic function navigation method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the functional path-based diagnostic function navigation method as described in any one of claims 1 to 7.

Images