Industrial vertical data fine adjustment method and device based on large model

Abstract

The invention provides an industry vertical data fine tuning method and device based on a large model, and the method comprises the steps: obtaining an operation order text data flow, carrying out the preliminary structural processing of each operation order text, extracting an operation action, an execution object and a constraint condition, and forming text data carrying an initial label; determining an operation process directed acyclic graph and an emergency operation process network according to the fine tuning corpus of the text classification model, and taking the operation process directed acyclic graph and the emergency operation process network as structured training samples for fine tuning of the text classification model; determining a mixed training batch of fine tuning of the text classification model according to the structured training sample of fine tuning of the text classification model, and setting a dynamic balance threshold value of the standard operation sample and the emergency operation sample; and after fine tuning of the text classification model is completed and deployment is carried out, if the operation ticket text to be examined belongs to emergency operation, parallel or conditional branch logic reasoning processing is carried out based on the structure of the emergency operation process network, and a structured examination report is output.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and apparatus for fine-tuning industry-specific vertical data based on a large model. Background Technology

[0002] In high-risk industries such as power and chemicals, intelligent review of operation permits is a core element in ensuring production safety and operational standardization. Automated review systems effectively prevent operational errors caused by human negligence and are crucial for maintaining stable system operation. They not only improve work efficiency but also build a critical technical defense.

[0003] Current large-scale model-based review methods, in pursuit of model generalization ability, tend to train all types of operation ticket texts, whether routine planned operations or emergency repair operations, using a unified data source. This approach ignores the distinctly different execution logics underlying the texts in different operational scenarios—one being a rigorous linear process, the other a flexible non-linear decision-making process—attempting to understand two inherently contradictory instruction paradigms with a single unified model. The core problem lies in the fact that standard operation tickets emphasize a strict sequence of steps. For example, in power grid operations, one must first "disconnect switch A," then "verify the absence of power," and finally "connect the grounding wire." Any reversal of this order can lead to serious accidents. However, emergency repair operation tickets, for rapid response, often include parallel or variable-order instructions such as "simultaneously dispatching personnel to handle faults at points A and B" or "isolating equipment D while restoring power to line C." As the model learns this flexible emergency logic during training, its ability to judge the insurmountable "sequential" dependencies inherent in standard operations is weakened.

[0004] As the amount of emergency scenario data increases, the model may mistakenly assume that certain strict steps in standard operations are also adjustable, leading to critical misjudgments in scenarios where precise sequential review is most needed.

[0005] Therefore, how to enable the model to learn the flexibility and parallelism of emergency operations without compromising its ability to review the strict sequence of standard operating procedures, and effectively isolate and manage the mutual interference between these two conflicting operational logics during the training process, has become a key issue in improving the reliability of intelligent review of operation tickets in complex industrial scenarios. Summary of the Invention

[0006] This invention provides a method for fine-tuning industry-specific vertical data based on a large model, the method comprising: Obtain the operation ticket text data stream, perform preliminary structured processing on each operation ticket text, extract the operation action, execution object and constraints, and form text data with initial labels; Obtain industry-specific operational ticket corpus, construct a text classification model for operational ticket review, identify the operational sentence structure, specific keywords, and context of the text data carrying initial labels through the text classification model, analyze to obtain the business type, and classify the text data into the standard operational sample pool or the emergency operational sample pool according to the business type. Semantic analysis is performed on the sample texts in the standard operation sample pool and the emergency operation sample pool respectively to identify the conjunctions and action steps that represent a strict sequential order, as well as the text fragments that describe parallel execution, conditional branches or optional order, to form the fine-tuning corpus of the text classification model. Based on the fine-tuning corpus of the text classification model, determine the directed acyclic graph of the operation process and the emergency operation process network, and use the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model. After the text classification model is fine-tuned and deployed, if the operation ticket text to be reviewed belongs to an emergency operation, then parallel or conditional branch logic reasoning processing is performed based on the structure of the emergency operation process network to output a structured review report.

[0007] Furthermore, the process of acquiring the operation ticket text data stream involves performing preliminary structured processing on each operation ticket text, extracting the operation action, execution object, and constraints, and forming text data carrying initial tags, including: The standard sentence template in the operation ticket is matched using regular expressions to identify key verbs, and the device name after the key verb is extracted as the execution object to obtain the pairing relationship between the action and the object. Based on the pairing relationship between the action and the object, the syntactic component annotation method is used to identify adverbial and attributive components, and to extract the relationship between the constraints and operation steps corresponding to the temporal and conditional identifiers. For each set of operation instructions in the aforementioned association, a triplet structure containing action type, device number, and constraint identifier is constructed. Based on the correspondence rules between action type and standard operation and emergency operation, an initial business label is assigned to each triplet.

[0008] Furthermore, the process of acquiring industry-specific operational ticket corpora, constructing a text classification model for operational ticket review, identifying operational sentence structures, specific keywords, and contextual information in the text data carrying initial labels using the text classification model, analyzing to obtain the business type, and classifying the text data into a standard operational sample pool or an emergency operational sample pool based on the business type includes: We acquired a historical corpus of operation tickets from the power industry, and used sequence labeling to tag the operation verbs, equipment names, and temporal words in the corpus with part-of-speech tags. We then constructed a sentence template library that includes verb-equipment conditional structures and temporal verb-object structures. Extract the inverse document frequency feature value of the word frequency, calculate the distribution weight of the keywords in different business types, generate the context vector through word embedding method, train the text classifier with support vector machine, and the text classifier outputs the classification decision function of standard operation and emergency operation; According to the classification decision function, the text data carrying the initial label is matched with sentence patterns, and the similarity score between the text and each sentence pattern template is calculated. The word frequency inverse document frequency feature value, keyword distribution weight and context vector are input into the classification decision function to obtain the business type confidence score. The business type is determined based on the confidence score of the business type and then assigned to the standard operation sample pool or the emergency operation sample pool.

[0009] Furthermore, the semantic parsing of sample texts in the standard operation sample pool and the emergency operation sample pool is performed respectively to identify conjunctions and action steps representing strict sequential order, as well as text fragments describing parallel execution, conditional branches, or optional sequences, forming a fine-tuning corpus for the text classification model, including: For the text in the standard operation sample pool, a dependency parser is used to analyze the dependency relationship, identify the conjunctions and the action verbs they modify, extract the action step pairs associated with the conjunctions, and construct a directed graph structure containing action nodes and execution order edges. For the text in the emergency operation sample pool, parallel identifiers and condition words are identified by keyword matching, and parallel execution logic segments and conditional branch paths are extracted. By adding sequence labels to semantic fragments with strict sequence constraints and adding parallel or branch labels to logical fragments executed in parallel, the fragments are differentiated and then merged.

[0010] Furthermore, the step of determining the directed acyclic graph of the operation process and the emergency operation process network based on the fine-tuning corpus of the text classification model, and using the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model, includes: Based on the sequential constraint semantic fragments in the fine-tuning corpus, the action nodes and execution order relationships are extracted, and the topological sorting algorithm is used to detect loops to determine the directed acyclic graph of the operation process. Based on the parallel execution and conditional branching segments in the fine-tuned corpus, concurrent nodes and branch nodes are identified to form the emergency operation process network; By structurally encoding the directed acyclic graph of the operation process and the emergency operation process network, the node attributes are encoded into a node attribute matrix, and the connection relationships and weights are encoded into an adjacency relationship matrix, thus generating structured training samples for fine-tuning the text classification model.

[0011] Furthermore, the method also includes: Based on the category labels of the structured training samples, the number of standard operation samples and emergency operation samples is counted, the initial ratio of the two types of samples is calculated, and the samples are grouped according to the rule that each batch contains a predetermined number of samples to determine the mixed training batch. The dynamic balance threshold is set using the ratio of the two types of samples in the mixed training batch.

[0012] Furthermore, the method also includes: During the fine-tuning of the text classification model, the proportion of emergency operation process network samples in the current mixed training batch is counted. If it exceeds the dynamic equilibrium threshold, the error enhancement and recognition mechanism is activated. The action execution sequence in the directed acyclic graph samples of the operation process is extracted, and the sequence error type is determined by comparing it with the standard sequence template. Error category labels and severity markers are added to the position of the error action node.

[0013] Furthermore, if the operation ticket text to be reviewed belongs to an emergency operation, then based on the structure of the emergency operation process network, parallel or conditional branch logic reasoning is performed to output a structured review report, including: The text of the operation ticket to be reviewed is classified. If it is determined to be an emergency operation type, the emergency operation process network structure is loaded, and concurrent nodes and branch nodes are traversed to obtain the operation execution path. Based on the operation execution path, the path rationality is verified and the condition satisfaction is checked. The review results are organized into a structured review report that includes operation compliance, execution order, and risk level.

[0014] The present invention also provides an industry vertical data fine-tuning device based on a large model, the device comprising: The data generation unit is used to acquire the operation ticket text data stream, perform preliminary structuring processing on each operation ticket text, extract the operation action, execution object and constraints, and form text data carrying initial labels. The data classification unit is used to acquire industry-specific operational ticket corpus, construct a text classification model for operational ticket review, identify the operational sentence structure, specific keywords, and context of the text data carrying initial labels through the text classification model, analyze to obtain the business type, and classify the text data into the standard operational sample pool or the emergency operational sample pool according to the business type. The corpus formation unit is used to perform semantic parsing on the sample texts in the standard operation sample pool and the emergency operation sample pool, respectively, to identify the conjunctions and action steps that represent a strict sequential order, as well as the text fragments that describe parallel execution, conditional branches or optional order, and to form a fine-tuning corpus for the text classification model. The sample setting unit is used to determine the directed acyclic graph of the operation process and the emergency operation process network based on the fine-tuning corpus of the text classification model, and to use the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model. The report output unit is used to, after the text classification model has been fine-tuned and deployed, if the operation ticket text to be reviewed belongs to an emergency operation, perform parallel or conditional branch logic reasoning based on the structure of the emergency operation process network, and output a structured review report.

[0015] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a method for fine-tuning industry-specific vertical data based on a large model. It addresses business scenario problems in processing operation ticket text data within industry verticals, namely, how to efficiently distinguish between standard and emergency operations, and accurately analyze the strict sequence and complex logical relationships of operation processes to generate structured review reports. This invention performs preliminary structuring processing on the operation ticket text, extracts key information, and constructs a text classification model, classifying the data into a standard operation sample pool and an emergency operation sample pool. Subsequently, semantic parsing identifies connectives and logical relationships in the process, generating a directed acyclic graph and an emergency operation process network as training samples. The model is then fine-tuned using a dynamic balancing threshold to ensure the review model efficiently identifies and labels errors in the step sequence. Finally, for emergency operations, parallel and conditional branching logical reasoning is implemented to output accurate review reports. This invention significantly improves the intelligence level and process optimization capabilities of operation ticket review, providing important guarantees for the safety and efficiency of industry operations. Attached Figure Description

[0016] Figure 1 This is a flowchart of an industry-specific data fine-tuning method based on a large model, according to the present invention.

[0017] Figure 2 This is a schematic diagram of an industry vertical data fine-tuning method based on a large model according to the present invention.

[0018] Figure 3 This is another schematic diagram of an industry vertical data fine-tuning method based on a large model according to the present invention.

[0019] Figure 4 This is a schematic diagram of the structure of an industry vertical data fine-tuning device based on a large model according to the present invention. Detailed Implementation

[0020] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0021] like Figures 1-3 This embodiment of a method for fine-tuning industry-specific data based on a large model may specifically include: Step S101: Obtain the operation ticket text data stream, and perform preliminary structuring processing on each operation ticket text to extract the operation action, execution object and constraint conditions, forming text data carrying initial labels.

[0022] After acquiring the original operation ticket text data stream, regular expressions are used to match the standard sentence templates in the operation ticket, identify key verbs, and extract the device names following the key verbs as execution objects, thus obtaining the pairing relationship between actions and objects. Based on the pairing relationship between actions and objects, syntactic component annotation methods are used to identify adverbial and attributive components in the operation instructions. If a temporal identifier word exists in the sentence, it is determined to be a sequential constraint condition; if a conditional identifier word exists, it is determined to be a trigger constraint condition. The association relationship between the constraint condition and the corresponding operation step is extracted. For each set of operation instructions in the association relationship, a triplet structure containing action type, device number, and constraint identifier is constructed. According to the correspondence rules between action type and standard operation and emergency operation, an initial business label is assigned to each triplet, resulting in structured text data carrying the initial label.

[0023] Specifically, in one implementation, for the text data stream of operation tickets in the power industry, the system first performs sentence matching using a regular expression template library. The template library predefines standard sentence structures for power operations, such as the format "verb + equipment name + equipment number". When the system recognizes an operation command such as "disconnect 220kV bus tie switch 2201", the regular expression extracts "disconnect" as the operation verb and "220kV bus tie switch 2201" as the execution object, forming a pairing relationship between action and object.

[0024] Specifically, syntactic component labeling employs dependency parsing technology, identifying the grammatical function of each component by constructing a dependency tree of the sentence. When analyzing instructions like "first disconnect the high-voltage side switch of main transformer No. 2, then open the isolating switch," the system labels "first" as a temporal adverb, forming a sequential constraint relationship with the "disconnect" action. If a conditional sentence like "when the system frequency is below 49.8Hz, cut off non-critical loads" is detected, the system labels "when the system frequency is below 49.8Hz" as a conditional adverb, establishing a triggering constraint. These constraints and their corresponding operation steps are stored in an association matrix. Each element in the matrix records the constraint type, constraint content, and association strength value. The association strength value is calculated based on the temporal order or urgency of the constraint, sourced from a preset rule base, and calculated using the formula S=W. (1-D), where S is the strength value, W is the weight (ranging from 0 to 1), and D is the delay factor (ranging from 0 to 1). Specifically, the constraint weight W is extracted from a preset rule base according to the constraint type: time-mandated constraints (such as "must first") correspond to W=1.0, time-suggested constraints (such as "should first") correspond to W=0.7, and condition-triggered constraints (such as "when...") correspond to W=0.8. The delay tolerance factor D reflects the flexibility of the operation interval: D=0 for non-delayable safety-related operations, D=0.3 for operations that allow short time intervals, and D=0.6 for operations that can be flexibly arranged. The larger the correlation strength value, the more urgent the constraint, and the higher the execution priority of the corresponding operation. This matrix is ​​subsequently used to evaluate the execution priority in operation ticket classification. This refined constraint identification mechanism ensures the accurate grasp of the execution logic under different operation scenarios, laying the foundation for subsequent operation ticket classification.

[0025] It should be noted that the extraction of constraints is not limited to explicit timing words and condition words; the system also analyzes implicit logical relationships. For example, in "after verifying that there is no power, connect the grounding wire," although there is no explicit timing word, the word "after" implies a strict sequential requirement.

[0026] In one embodiment, the construction of the triplet structure involves multi-dimensional parsing of each operation instruction. The system establishes an encoding system for action types, such as encoding "disconnect" as 01 and "close" as 02. The device number is directly extracted from the numerical identifier in the original text. Constraint identifiers are assigned values ​​based on the aforementioned recognition results: sequential constraints are marked as S, triggered constraints as T, and no constraints as N.

[0027] Preferably, the allocation of business tags follows a predefined mapping rule table. When the action type in the triple is "disconnect" or "pull apart" and the constraint identifier is S, the system determines that the operation belongs to the standard power outage operation procedure and assigns the "Standard Operation - Power Outage" tag. If the action type contains keywords such as "emergency" or "immediate" and the constraint identifier is T, then the "Emergency Operation - Fault Handling" tag is assigned.

[0028] For example, when processing substation switching operation tickets, the system parses the instruction "open 10kV feeder switch 1021", constructs a triple {03,1021,N}, determines it as a routine power distribution operation according to the mapping rules, assigns the corresponding initial business tag, and finally outputs structured text data with the tag for use by subsequent processing modules.

[0029] Step S102: Obtain industry-specific operation ticket corpus, construct a text classification model for operation ticket review, identify the operation sentence structure, specific keywords and context of the text data carrying initial labels through the text classification model, analyze to obtain the business type, and classify the text data according to the business type, including standard operation sample pool or emergency operation sample pool.

[0030] A historical corpus of power industry operation tickets is acquired. Sequence labeling is used to tag the operation verbs, equipment names, and temporal words in the corpus with part-of-speech tags. Based on the tagging results, a sentence template library containing verb-equipment conditional structures and temporal verb-object structures is constructed, resulting in training corpus with sentence tags. For the training corpus, inverse document frequency (IVF) features are extracted, and the distribution weights of the keywords "emergency," "immediate," and "fault" in different business types are calculated. Contextual vectors are generated using word embedding methods, and a support vector machine (SVM) is used to train a text classifier. The text classifier outputs a classification decision function for standard operations and emergency operations. Based on the classification decision function, sentence matching is performed on the text data carrying initial labels, and the similarity score between the text and each sentence template is calculated. The IVF features, keyword distribution weights, and contextual vectors are input into the classification decision function to obtain a business type confidence score. If the confidence score exceeds a preset threshold, the business type is determined to be a standard operation and assigned to the standard operation sample pool; otherwise, it is determined to be an emergency operation and assigned to the emergency operation sample pool.

[0031] Specifically, in one implementation, the construction of an electricity operation ticket corpus involves the systematic processing of historical operation ticket texts. The sequence labeling method employs a conditional random field labeling tool to assign labels to each lexical unit in the corpus.

[0032] Specifically, the sentence template library is constructed based on statistical analysis of the annotation results. The system traverses all annotated corpora and extracts sentence structures whose frequency exceeds a preset threshold. Each template records its frequency of occurrence in standard and emergency operations, forming a template weight matrix. Through annotation analysis of 5,000 historical operation tickets, the system extracts 128 common sentence templates, forming the structured foundation of the training corpus.

[0033] It should be noted that the calculation of the Term Frequency-Inverse Document Frequency (TF-IDF) feature involves statistics in two dimensions. The TF value is calculated as TF = n / N, where n is the number of times a word appears in a single operation ticket text, and N is the total number of words in that text. The IDF value is calculated as IDF = log(D / d), where D is the total number of documents, and d is the number of documents containing the word. The TF-IDF feature value is calculated as: TF-IDF = TF × IDF. For example, for the keyword "urgent," if it appears 3 times in a certain operation ticket with a total of 100 words, then TF = 3 / 100 = 0.03; if 50 out of a total of 1000 operation tickets contain "urgent," then IDF = log(1000 / 50) ≈ 2.996. Multiplying the two results in the TF-IDF feature value of this word: 0.03 × 2.996 ≈ 0.0899. The system calculates the TF-IDF value for all keywords to form a feature vector.

[0034] Preferably, the word embedding method adopts the Skip-gram architecture of Word2Vec, mapping each word to a 300-dimensional continuous vector space. During training, the window size is set to 5, considering the contextual relationships of the five words before and after the target word. For the operation command "immediately disconnect the faulty line and transfer the load," the context of the word "fault" includes "immediately," "disconnect," "line," "and," and "transfer." Through neural network training, semantically similar words such as "fault," "accident," and "abnormal" are close in distance in the vector space. The context vector is obtained by averaging the word vectors of all words in the sentence, reflecting the semantic features of the entire operation command. This vectorized representation captures the potential semantic relationships between words, providing rich semantic information for subsequent classification.

[0035] In one embodiment, the training process of the support vector machine classifier employs a radial basis function kernel. The training samples include 3000 standard operation texts and 2000 emergency operation texts, each text represented as a high-dimensional vector fusing TF-IDF features and word embedding features. The kernel function parameter γ is set to 0.01, and the penalty parameter C is set to 1.0. The classifier learns to distinguish between standard and emergency operations by finding the maximum margin hyperplane. The output form of the classification decision function is f(x) = sign(Σαi·yi·K(xi,x) + b), where αi is the Lagrange multiplier, yi is the training sample label, K is the kernel function, b is the bias term, and sign is the sign function (positive value output +1, negative value output -1).

[0036] For example, in the actual classification process, the system performs multi-dimensional analysis on the text to be classified. Sentence matching is achieved by calculating the edit distance between the text and each template; the smaller the distance, the higher the similarity. When the input text is "Discover equipment malfunction and immediately shut down for repair," the system identifies the emergency operation feature word "immediately," with a TF-IDF weight of 0.125. Simultaneously, the text matches the emergency operation sentence template "discover + malfunction + immediately + action" with a score of 0.85. The contextual vector generated by word embedding shows that the text is closer to the emergency operation cluster center in the semantic space. Furthermore, feature fusion employs a weighted combination strategy. The TF-IDF feature weight is set to 0.3, the word embedding feature weight is set to 0.4, and the sentence matching score weight is set to 0.3. For the example text above, the TF-IDF feature value is 0.125, the normalized word embedding similarity is 0.72, the sentence matching score is 0.85, and the fused feature vector is input into the classification decision function, yielding a confidence score of 0.687.

[0037] Understandably, the threshold determination mechanism employs a dynamic adjustment strategy. The system initially sets the threshold to 0.5, and then calculates the classification accuracy by testing the validation set. If the proportion of standard operations misclassified as emergency operations exceeds 5%, the threshold is increased; conversely, if the proportion of emergency operations misclassified as standard operations exceeds 5%, the threshold is decreased. After multiple rounds of adjustments, the system determined that a threshold of 0.6 achieved a balanced classification performance.

[0038] For example, when processing the operation ticket "220kV busbar undervoltage, immediately start the backup power supply, and simultaneously notify the dispatcher," the system identifies emergency characteristic words such as "undervoltage," "immediately," and "simultaneously," calculating a confidence score of 0.42, which is below the threshold of 0.6. Therefore, it is assigned to the emergency operation sample pool. However, for standard operations such as "following the operating procedures, first open the isolating switch, then disconnect the circuit breaker," the confidence score reaches 0.78, and it is assigned to the standard operation sample pool.

[0039] In one possible implementation, a dynamic maintenance mechanism for the sample pool ensures classification accuracy. The system periodically assesses the quality of the sample pool, manually sampling to identify misclassified samples and promptly adjusting classifier parameters. This closed-loop optimization mechanism allows the system to maintain high classification accuracy even when processing new types of operation ticket texts, laying a reliable data foundation for subsequent operational process review.

[0040] Step S103: Semantic parsing is performed on the sample texts in the standard operation sample pool and the emergency operation sample pool respectively to identify the conjunctions and action steps that represent a strict sequential order, as well as the text fragments that describe parallel execution, conditional branches or optional order, to form the fine-tuning corpus for the text classification model.

[0041] For text in the standard operation sample pool, a dependency parser is used to analyze the dependencies between sentence components, identify conjunctions and the action verbs they modify, extract the action step pairs associated with the conjunctions, and construct a directed graph structure containing action nodes and execution order edges to obtain semantic fragments with strict order constraints. For text in the emergency operation sample pool, parallel identifiers are identified through keyword matching, and multiple action groups connected by the parallel identifiers are extracted. If the text contains conditional words, conditional branch paths and their triggering conditions are extracted to obtain parallel execution logic fragments. Differentiated annotation is performed by adding order labels to the strictly order-constrained semantic fragments and parallel or branch labels to the parallel execution logic fragments. The two types of annotated semantic fragments are then merged to form a text classification model fine-tuning corpus containing execution mode identifiers.

[0042] Specifically, in one implementation, a dependency parser identifies dependencies between actions by analyzing the syntactic structure of the operation ticket text. The parser constructs a dependency tree, where each word is a node and the grammatical relationships between words are edges. When processing standard operation text such as "first open the isolating switch, then disconnect the circuit breaker," the parser identifies "first" as an adverbial modifier of "open" and "then" as an adverbial modifier of "disconnect," thereby determining the execution order of the two actions.

[0043] Specifically, the construction process of a directed graph involves two stages: node creation and edge connection. Each action step is a node in the graph, and the node attributes include action type, operation object, and execution condition. When a connector is identified, the system establishes a directed edge between the corresponding action nodes, and the direction of the edge indicates the order of execution.

[0044] For example, the "open the disconnect switch" node points to the "disconnect the circuit breaker" node, forming a strict execution chain. The directed graph is stored using an adjacency list, and each node maintains a list of its successor nodes to ensure the accurate expression of the order relationship. By performing a depth-first traversal of the graph, the system extracts the complete operation execution path, forming a semantic fragment of the order constraint for standard operations.

[0045] It should be noted that the emergency operation sample pool uses a different identification mechanism. The system uses regular expressions to match parallel identifiers such as "simultaneously" and "while...while" to extract action groups that are executed in parallel. When the text contains conditional statements such as "if the equipment temperature exceeds 80 degrees, start the cooling system; otherwise, continue monitoring," the system identifies the conditional trigger word "if" and the branch identifier words "then" and "otherwise," constructs a conditional branch structure, and records the triggering conditions and the corresponding execution paths.

[0046] Preferably, the differentiated annotation adopts a unified labeling system. Sequential constraint semantic fragments are labeled as "SEQ" type, with each action appended with a sequence number to indicate the execution order; parallel execution fragments are labeled as "PAR" type, with parallel actions in the same group using the same group number; conditional branch fragments are labeled as "COND" type, recording the conditional expression and branch path.

[0047] In one embodiment, the labeled semantic fragments are formatted to form a fine-tuned corpus. The system converts each fragment into a structured data format, containing the original text, execution mode labels, action sequences, and constraint relationships.

[0048] This structured representation preserves the original semantic information while highlighting the feature differences between different execution modes, providing high-quality training data for fine-tuning text classification models.

[0049] Step S104: Determine the directed acyclic graph of the operation process and the emergency operation process network based on the fine-tuning corpus of the text classification model, and use the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model.

[0050] Based on the sequential constraint semantic fragments in the fine-tuning corpus, action nodes and execution order relationships are extracted, and a node set and an edge set are constructed. A topological sorting algorithm is used to detect the existence of cycles; if no cycles are found, a directed acyclic graph (DAG) for the operation flow is determined. For the DAG, the weight value of each edge is calculated based on the temporal mandatory and reversible attributes between actions; the higher the temporal mandatory attribute, the greater the weight. A depth-first traversal algorithm is used to obtain all feasible paths from the starting node to the ending node, resulting in the execution path set for the standard operation. Based on the parallel execution and conditional branch fragments in the fine-tuning corpus, action groups that can be executed simultaneously are identified as concurrent nodes, and condition judgment points are identified as branch nodes. The concurrent nodes allow multiple outgoing edges to be activated simultaneously, while the branch nodes select a single outgoing edge based on the condition, forming an emergency operation flow network. By performing structured encoding on the execution path set and the emergency operation flow network, the action type, execution condition, dependency relationship, and path sequence of the nodes are encoded into a node attribute matrix, and the connection relationship and weight between nodes are encoded into an adjacency relationship matrix, generating structured training samples for fine-tuning the text classification model.

[0051] Specifically, in one implementation, the construction of the directed acyclic graph (DAG) begins with fine-tuning the sequential constraint semantic fragments in the corpus. The system first parses each semantic fragment, extracting action node information, including action identifier, operation object, and execution condition. Each action node is assigned a unique number, such as N1 for "disconnecting the 220kV switch" and N2 for "verification confirmation." When a semantic fragment indicates that N1 must be executed before N2, the system creates a directed edge from N1 to N2 in the node set. A topology sorting algorithm detects cycles by maintaining the in-degree value of each node; if all nodes can be visited during the sorting process and the in-degree drops to zero, the graph structure is confirmed to be a DAG.

[0052] Specifically, topological sorting is implemented using Kahn's algorithm. The system initializes a queue, adding all nodes with an in-degree of zero; these nodes represent the initial actions in the workflow. Nodes are removed from the queue and added to the sorting result, while the in-degree of all their successors is decremented by one. If the in-degree of a successor node becomes zero, it is added to the queue. When the queue is empty, if the sorting result contains all nodes, the graph is acyclic; otherwise, a cycle exists, and the system marks conflicting action sequences and prompts for correction. This cycle detection mechanism prevents logical contradictions in the workflow, such as circular dependencies like "Operation A must precede B, B must precede C, and C must precede A."

[0053] It should be noted that the edge weights are calculated based on attribute evaluation across multiple dimensions. Temporal mandatory order is determined by analyzing the semantic strength of the connecting words: "must first" corresponds to a weight of 1.0, "should first" to 0.8, and "suggests first" to 0.6. The reversibility attribute examines whether an operation is undone; irreversible operations such as "disconnecting the main transformer" increase the weight of preceding edges by 0.2. Safety relevance reflects the risk level that violating the order may cause; operations involving personal safety increase the weight of edges by an additional 0.3. The final weight is the weighted sum of the three attribute values, with weight coefficients of 0.5, 0.3, and 0.2, respectively.

[0054] Preferably, depth-first traversal employs a recursive approach for path exploration. Starting from the initial node, the system maintains a path stack to record the current traversal path. When a node is visited, it is pushed onto the stack, and all its unvisited adjacent nodes are explored. When the final node is reached, the sequence of nodes in the stack constitutes a complete execution path. During backtracking, the top node is popped from the stack, and other branches are explored. In this way, the system can find all feasible paths from the start to the end, forming a set of execution paths for standard operations.

[0055] In one embodiment, the construction of the emergency operation process network involves the identification of special node types. Concurrent nodes are identified by recognizing keywords such as "simultaneously" and "separately." When the text description is "simultaneously start the backup power supply and cut off non-critical loads," the system creates a concurrent node P1, from which two edges point to the "start backup power supply" and "cut off non-critical loads" action nodes respectively. The characteristic of the concurrent node is that all its outgoing edges can be activated simultaneously, representing that multiple operations can be executed in parallel. Branch nodes are identified by "if...then...otherwise" statements, such as "if the bus voltage is normal, continue operation; otherwise, transfer the load." The system creates a branch node B1, setting the conditional judgment logic "bus voltage >= rated value." If the condition is met, activate the edge pointing to "Continue running"; otherwise, activate the edge pointing to "Transfer load".

[0056] For example, the structured encoding process converts the graph structure into a matrix representation suitable for machine learning. The node attribute matrix is ​​N×M dimensional, where N is the total number of nodes and M is the attribute dimension. Attributes include five dimensions: action type encoding, device type encoding, voltage level, operation priority, and safety level. For example, the node "disconnect 220kV switch" is encoded as [1,3,220,2,3], where 1 represents a disconnection action, 3 represents a switch device, 220 represents the voltage level, 2 represents medium priority, and 3 represents a high safety level. Furthermore, the adjacency matrix uses a weighted sparse matrix representation. Matrix element A[i,j] represents the edge weight from node i to node j, with a value of 0 when there is no edge connection. For a directed acyclic graph, after numbering nodes in topological order, this matrix is ​​an upper triangular matrix. The adjacency matrix of the emergency operation procedure network uses an extended encoding method: positive numbers represent edge weights executed in normal order, -1 represents concurrent relationships (allowing simultaneous execution), -2 represents conditional branch relationships (selecting execution based on conditions), and 0 represents no connection. These special markers distinguish different connection types, enabling the adjacency matrix to fully express the complex execution logic of emergency operations.

[0057] Understandably, the encoded structured training samples consist of three parts: a node attribute matrix providing node feature information, an adjacency relation matrix describing the topology, and execution mode labels distinguishing between standard and emergency operations. This multi-dimensional structured representation preserves the complete semantic information and topological features of the operation process.

[0058] For example, when handling substation switching operations, the system constructs a directed acyclic graph containing 12 nodes and 18 edges. Through structured encoding, a 12×5 node attribute matrix and a 12×12 adjacency matrix are formed. When handling emergency fault handling procedures, the constructed emergency network contains 8 regular nodes, 2 concurrent nodes, and 1 branch node. The encoded matrix accurately expresses the complex logic of parallel and conditional execution through special labels.

[0059] In one possible implementation, these structured training samples are directly input into the GCN (Graph Convolutional Network) layer of a text classification model. The input is a graph representation with operation steps as nodes and dependencies as edges, and the output is the classification result of the operation ticket. First, a graph structure is constructed: each node embeds a vector representation of the operation description, and edges capture sequential or conditional dependencies. Then, through a message-passing mechanism, nodes aggregate neighbor information to learn dependencies and execution logic. Standard operation samples train the model to identify strict sequential constraints, while emergency operation samples train the model to understand flexible execution patterns. This graph-based training method enables the model to accurately distinguish different types of operation tickets, providing a reliable classification foundation for subsequent intelligent review.

[0060] Step S105: Determine the mixed training batch for fine-tuning the text classification model based on the structured training samples for fine-tuning the text classification model, and set the dynamic balance threshold between standard operating samples and emergency operating samples.

[0061] Based on the category labels of the structured training samples, the number of standard operation samples and emergency operation samples is counted, and the initial ratio of the two types of samples is calculated. The samples are then grouped according to the rule that each batch contains a predetermined number of samples to determine the mixed training batch. Using the ratio of the two types of samples in the mixed training batch, a dynamic balance threshold is set based on the proportion of standard operation samples to the total samples. When the proportion of standard operation samples falls below the threshold, the sampling weights are adjusted to obtain the batch configuration for fine-tuning the text classification model.

[0062] Specifically, in one implementation, the determination of mixed training batches is based on statistical analysis of sample categories. The system scans the category labels of all structured training samples, resulting in 3200 standard operation samples and 1800 emergency operation samples, with an initial ratio of 64:36. Each batch is set to contain 100 samples, allocated according to the initial ratio, with each batch containing 64 standard operation samples and 36 emergency operation samples, forming a total of 50 mixed training batches.

[0063] Specifically, the dynamic balancing threshold is set to take into account the stability requirements of model training. The threshold is set to 0.6, which requires that the proportion of standard operation samples be no less than 60%. When the actual proportion of standard operation samples in a batch drops to 55%, the system rebalances by adjusting the sampling weights. The sampling weight of standard operation samples is increased from 1.0 to 1.2, while the weight of emergency operation samples remains at 1.0, thus increasing the probability of standard operation samples being selected.

[0064] For example, in the 25th training batch, if there are too many emergency operation samples leading to an imbalance in the ratio, the system automatically increases the sampling probability of standard operation samples, preferentially selecting standard operation samples from the remaining sample pool to supplement the batch until the ratio recovers to above the threshold. This dynamic adjustment mechanism ensures that the model maintains a balance in learning the two types of operation tickets during training, avoiding bias towards one category.

[0065] Step S106: When fine-tuning the text classification model, evaluate whether the proportion of emergency operation procedure networks in the mixed training batch exceeds the dynamic equilibrium threshold. If it does, strengthen the identification and annotation of incorrect step order in the directed acyclic graph samples of the operation procedure.

[0066] During the fine-tuning of the text classification model, the number of emergency operation procedure network samples within the current mixed training batch is counted, and their proportion of the total batch samples is calculated to obtain the actual proportion of emergency samples. This actual proportion is compared with a preset dynamic balance threshold. If the actual proportion exceeds the threshold, an error enhancement and identification mechanism is activated. The action execution sequence in the directed acyclic graph samples of the operation procedure is extracted, and the sequence error type is determined by comparing it with a standard sequence template. For each sequence error type, the weight coefficients of the loss function for samples containing errors are adjusted from the baseline value based on the error severity. Error category labels and severity markers are added to the error action node positions to increase the model's learning intensity for this type of sequence violation pattern, resulting in the adjusted fine-tuning configuration.

[0067] Specifically, in one implementation, the dynamic monitoring mechanism of the model fine-tuning process evaluates the sample distribution of each training batch in real time. At the beginning of each training round, the system scans the 100 samples in the current batch and counts the number of emergency operation procedure network samples. When 45 emergency samples are detected, accounting for 45%, the calculated actual percentage is passed to the threshold comparison module. This module compares the actual percentage with a preset threshold (e.g., 40%). If the actual percentage exceeds the threshold, the sample adjustment mechanism is triggered.

[0068] Specifically, the activation condition of the error enhancement identification mechanism is based on a dynamic balance threshold. This threshold is set to 0.4, meaning the proportion of emergency samples should not exceed 40%. When the actual proportion exceeds 45% of the threshold, the system automatically activates the error enhancement identification mechanism. This mechanism first extracts action execution sequences from the directed acyclic graph samples of the operation flow. Each sequence contains the operation action, the execution object, and the temporal relationship. A standard sequence template predefines the standardized process of power operation, such as the fixed sequence of "power off - voltage testing - grounding wire connection." The system compares the extracted sequences with the template one by one, identifying three error types: reversed order, omitted steps, and insertion of unnecessary actions.

[0069] It should be noted that the weights of the loss function are adjusted using a tiered enhancement strategy. The baseline weight is set to 1.0. For errors involving reversed order, the weight is multiplied by 2.0; for errors involving missing critical steps, the weight is multiplied by 2.5; and for errors involving inserting unnecessary actions, the weight is multiplied by 1.5. This differentiated weight adjustment allows the model to learn different learning responses to errors of varying severity. On samples containing severe reversed order errors such as "connecting the ground wire before checking for voltage," the loss function value is amplified by 2 times, forcing the model to focus on this type of error pattern when updating the model parameters.

[0070] Preferably, the enhanced error labeling is achieved through multi-dimensional labeling. At the location of the action node where the error is identified, the system adds three layers of labeling information: an error category label to identify the error type, such as "SEQ_REVERSE" indicating a reversed sequence; a severity label using a level value of 1-5, where 5 indicates an error that may lead to a serious accident; and an association impact label to record the scope of the error's impact on subsequent operations.

[0071] For example, when processing a training sample containing the error "operating directly without voltage testing," the system adds a "MISS_CRITICAL" category label, a severity level of 5, and an association label that affects the subsequent three operation steps at the corresponding location. Through this enhanced annotation, the model increases the gradient update magnitude at these locations during backpropagation, strengthening the learning and memorization of the standard operation sequence.

[0072] In one embodiment, the adjusted fine-tuning configuration includes updated batch composition, modified loss weight matrix, and enhanced annotation information set. These configuration parameters are passed to the training module in real time to ensure that the model can maintain its understanding of the flexibility of emergency operations without losing its ability to judge the strict order of standard operations.

[0073] Step S107: After the text classification model is fine-tuned and deployed, if the operation ticket text to be reviewed belongs to emergency operation, then the parallel or conditional branch logic is used for reasoning based on the structure of the emergency operation process network to output a structured review report.

[0074] Once the finely tuned text classification model is deployed, the text of the operation ticket to be reviewed is categorized. If it is determined to be an emergency operation type, the emergency operation process network structure is loaded. Concurrent nodes refer to operation points that are allowed to execute in parallel, and branch nodes refer to operation points that branch based on conditions. The operation execution path is obtained by traversing the concurrent and branch nodes in the network. Based on the operation execution path, the rationality of parallel actions and the triggering logic of conditional branches are judged through path rationality verification and condition satisfaction checks. The review results are then organized into a structured review report in a format that includes operation compliance, execution order, and risk level.

[0075] Specifically, in one implementation, the deployed text classification model receives the text of the operation ticket to be reviewed through an interface. The model first calls the classifier to determine the text category. When the output category label is "emergency operation", the system loads the corresponding emergency operation process network structure file from the model library. The network structure includes a node information table and an edge connection table. Starting from the starting node, a breadth-first traversal algorithm is used to visit concurrent nodes and branch nodes layer by layer, recording all reachable paths to form a set of operation execution paths.

[0076] Specifically, the path rationality verification checks for resource conflicts between parallel actions, such as ensuring that two parallel operations cannot simultaneously occupy the same device. The condition satisfaction check verifies the logical completeness of the branch conditions, confirming that each condition branch has a clear trigger threshold and mutual exclusion relationship. The review results are organized according to a predefined report template, including seven fields: operation ticket number, review time, operation type, compliance determination, risk level assessment, problem description, and improvement suggestions.

[0077] For example, when processing an "emergency handling of bus fault" operation ticket, the system identifies three parallel execution paths. Verification shows that "disconnecting the faulty bus" and "starting the backup power supply" can be executed in parallel, but "restoring power supply" must be performed after the first two are completed. The final structured report output marks the operation ticket as "conditionally compliant" and the risk level as "medium".

[0078] like Figure 4 As shown, this embodiment illustrates an industry-specific data fine-tuning device based on a large model, specifically including: The data generation unit 41 is used to acquire the operation ticket text data stream, perform preliminary structuring processing on each operation ticket text, extract the operation action, execution object and constraint conditions, and form text data carrying initial labels. Data classification unit 42 is used to acquire industry-specific operation ticket corpus, construct a text classification model for operation ticket review, identify the operation sentence structure, specific keywords and context of the text data carrying initial labels through the text classification model, analyze to obtain the business type, and classify the text data into the standard operation sample pool or the emergency operation sample pool according to the business type. Corpus forming unit 43 is used to perform semantic parsing on sample texts in the standard operation sample pool and the emergency operation sample pool respectively, identify conjunctions and action steps that represent strict sequential order, as well as text fragments that describe parallel execution, conditional branches or optional order, and form fine-tuning corpus for text classification model; The sample setting unit 44 is used to determine the directed acyclic graph of the operation process and the emergency operation process network based on the fine-tuning corpus of the text classification model, and to use the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model. The report output unit 45 is used to output a structured review report after the text classification model has been fine-tuned and deployed, if the operation ticket text to be reviewed belongs to an emergency operation, perform parallel or conditional branch logic reasoning based on the structure of the emergency operation process network.

[0079] In this embodiment of the invention, for the sake of convenience and brevity, only the division of the above-described functional units and modules is used as an example. In practical applications, the above functions can be allocated to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to achieve all or part of the functions described above. Each unit and module of the device can be implemented by corresponding hardware or software units. Each unit and module can be an independent hardware or software unit, or it can be integrated into a single hardware or software unit, which is not intended to limit the invention. In addition, the specific names of each functional unit and module are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the device can be referred to the corresponding description in the foregoing method embodiments, and will not be repeated here.

[0080] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for fine-tuning industry-specific vertical data based on a large model, characterized in that, The method includes: Obtain the operation ticket text data stream, perform preliminary structured processing on each operation ticket text, extract the operation action, execution object and constraints, and form text data carrying initial labels; Obtain industry-specific operational ticket corpus, construct a text classification model for operational ticket review, identify the operational sentence structure, specific keywords, and context of the text data carrying initial labels through the text classification model, analyze to obtain the business type, and classify the text data into the standard operational sample pool or the emergency operational sample pool according to the business type. Semantic analysis is performed on the sample texts in the standard operation sample pool and the emergency operation sample pool respectively to identify the conjunctions and action steps that represent a strict sequential order, as well as the text fragments that describe parallel execution, conditional branches or optional order, to form the fine-tuning corpus of the text classification model. Based on the fine-tuning corpus of the text classification model, determine the directed acyclic graph of the operation process and the emergency operation process network, and use the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model. After the text classification model is fine-tuned and deployed, if the operation ticket text to be reviewed belongs to an emergency operation, then parallel or conditional branch logic reasoning processing is performed based on the structure of the emergency operation process network to output a structured review report.

2. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, The process of obtaining the operation ticket text data stream involves performing preliminary structured processing on each operation ticket text, extracting the operation action, execution object, and constraints, and forming text data carrying initial tags, including: The standard sentence template in the operation ticket is matched using regular expressions to identify key verbs, and the device name after the key verb is extracted as the execution object to obtain the pairing relationship between the action and the object. Based on the pairing relationship between the action and the object, the syntactic component annotation method is used to identify adverbial and attributive components, and the relationship between the constraints and operation steps corresponding to the temporal and conditional identifiers is extracted. For each set of operation instructions in the aforementioned association, a triplet structure containing action type, device number, and constraint identifier is constructed. Based on the correspondence rules between action type and standard operation and emergency operation, an initial business label is assigned to each triplet.

3. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, The process involves acquiring industry-specific operational ticket corpora, constructing a text classification model for operational ticket review, identifying operational sentence structures, specific keywords, and contextual information in the text data carrying initial labels, analyzing the business type, and classifying the text data into a standard operational sample pool or an emergency operational sample pool based on the business type. We acquired a historical corpus of operation tickets from the power industry, and used sequence labeling to tag the operation verbs, equipment names, and temporal words in the corpus with part-of-speech tags. We then constructed a sentence template library that includes verb-equipment conditional structures and temporal verb-object structures. Extract the inverse document frequency feature value of the word frequency, calculate the distribution weight of the keywords in different business types, generate the context vector through word embedding method, train the text classifier with support vector machine, and the text classifier outputs the classification decision function of standard operation and emergency operation; According to the classification decision function, the text data carrying the initial label is matched with sentence patterns, and the similarity score between the text and each sentence pattern template is calculated. The word frequency inverse document frequency feature value, keyword distribution weight and context vector are input into the classification decision function to obtain the business type confidence score. The business type is determined based on the confidence score of the business type and then assigned to the standard operation sample pool or the emergency operation sample pool.

4. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, The process involves semantic parsing of sample texts in the standard operation sample pool and the emergency operation sample pool, identifying conjunctions and action steps representing strict sequential order, as well as text fragments describing parallel execution, conditional branches, or optional sequences, to form a fine-tuning corpus for the text classification model, including: For the text in the standard operation sample pool, a dependency parser is used to analyze the dependency relationship, identify the conjunctions and the action verbs they modify, extract the action step pairs associated with the conjunctions, and construct a directed graph structure containing action nodes and execution order edges; For the text in the emergency operation sample pool, parallel identifiers and condition words are identified by keyword matching, and parallel execution logic segments and conditional branch paths are extracted. By adding sequence labels to semantic fragments with strict sequence constraints and adding parallel or branch labels to logical fragments executed in parallel, the fragments are differentiated and then merged.

5. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, The step of determining the directed acyclic graph of the operation process and the emergency operation process network based on the fine-tuning corpus of the text classification model, and using the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model, includes: Based on the sequential constraint semantic fragments in the fine-tuning corpus, the action nodes and execution order relationships are extracted, and the topological sorting algorithm is used to detect loops to determine the directed acyclic graph of the operation process. Based on the parallel execution and conditional branching segments in the fine-tuned corpus, concurrent nodes and branch nodes are identified to form the emergency operation process network; By performing structured encoding on the directed acyclic graph of the operation process and the emergency operation process network, node attributes are encoded into node attribute matrices, and connection relationships and weights are encoded into adjacency relationship matrices, thus generating structured training samples for fine-tuning the text classification model.

6. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, The method further includes: Based on the category labels of the structured training samples, the number of standard operation samples and emergency operation samples is counted, the initial ratio of the two types of samples is calculated, and the samples are grouped according to the rule that each batch contains a predetermined number of samples to determine the mixed training batch. The dynamic balance threshold is set using the ratio of the two types of samples in the mixed training batch.

7. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, The method further includes: During the fine-tuning of the text classification model, the proportion of emergency operation process network samples in the current mixed training batch is counted. If it exceeds the dynamic equilibrium threshold, the error enhancement and recognition mechanism is activated. The action execution sequence in the directed acyclic graph samples of the operation process is extracted, and the sequence error type is determined by comparing it with the standard sequence template. Error category labels and severity markers are added to the position of the error action node.

8. The method for fine-tuning industry-specific vertical data based on a large model according to claim 1, characterized in that, If the operation ticket text to be reviewed belongs to an emergency operation, then parallel or conditional branching logic reasoning processing is performed based on the structure of the emergency operation process network to output a structured review report, including: The text of the operation ticket to be reviewed is classified. If it is determined to be an emergency operation type, the emergency operation process network structure is loaded, and concurrent nodes and branch nodes are traversed to obtain the operation execution path. Based on the operation execution path, the path rationality is verified and the condition satisfaction is checked. The review results are organized into a structured review report that includes operation compliance, execution order, and risk level.

9. A device for fine-tuning industry-specific vertical data based on a large model, characterized in that, The device includes: The data generation unit is used to acquire the operation ticket text data stream, perform preliminary structuring processing on each operation ticket text, extract the operation action, execution object and constraints, and form text data carrying initial labels. The data classification unit is used to acquire industry-specific operational ticket corpus, construct a text classification model for operational ticket review, identify the operational sentence structure, specific keywords, and context of the text data carrying initial labels through the text classification model, analyze to obtain the business type, and classify the text data into the standard operational sample pool or the emergency operational sample pool according to the business type. The corpus formation unit is used to perform semantic parsing on the sample texts in the standard operation sample pool and the emergency operation sample pool, respectively, to identify the conjunctions and action steps that represent a strict sequential order, as well as the text fragments that describe parallel execution, conditional branches or optional order, and to form a fine-tuning corpus for the text classification model. The sample setting unit is used to determine the directed acyclic graph of the operation process and the emergency operation process network based on the fine-tuning corpus of the text classification model, and to use the directed acyclic graph of the operation process and the emergency operation process network as structured training samples for fine-tuning the text classification model. The report output unit is used to, after the text classification model has been fine-tuned and deployed, if the operation ticket text to be reviewed belongs to an emergency operation, perform parallel or conditional branch logic reasoning based on the structure of the emergency operation process network, and output a structured review report.

Images