A method and system for restricted fact derivation suppression and corpus reconstruction for power large language model training

Abstract

The application discloses a restricted fact derivation inhibition and corpus reconstruction method and system for power large language model training. Training tasks and candidate power training corpora are obtained, fact association structures are constructed, task-required fact relationships and task-required fact paths are determined; restricted facts are determined according to corpus usage constraints, non-restricted facts are derived step by step based on power fact derivation rules, and restricted fact derivation paths are formed; the overlapping relationship between the restricted fact derivation paths and the task-required fact structures is analyzed, target adjustment elements are determined according to the influence of candidate adjustment on the restricted fact derivation and the task maintenance state; fact abstraction, replacement, identification generalization, parameter intervalization, association relationship disconnection or fact removal are performed on the candidate corpora, the task maintenance state is verified again after reconstruction, and the restricted fact derivation path is searched, and the controlled training corpus is formed when the task maintenance condition is met and the restricted fact cannot be derived. The application considers task knowledge maintenance and restricted fact derivation inhibition.

Description

Technical Field

[0001] This invention belongs to the field of power artificial intelligence training data processing and power data security technology. Specifically, it relates to a method and system for restricted fact inference suppression and corpus reconstruction for power large language model training. It is applicable to power large language model training, power professional model training corpus processing, power artificial intelligence training data security control, and controlled use scenarios of training corpus in intelligent applications of the power industry. Background Technology

[0002] As large language model technology is increasingly applied in scenarios such as power dispatching, equipment operation and maintenance, fault analysis, operation mode analysis, customer service, and decision support, the power industry has gradually accumulated and formed various types of training corpora, including operating procedures, equipment data, dispatching records, operating events, fault handling records, and professional question-and-answer data. Compared with general domain training data, the factual elements in power training corpora usually have strong professional relevance, and different factual elements may be related through electrical topology, equipment state transitions, operating events, operation sequences, and changes in operating parameters.

[0003] To mitigate the risk of sensitive information exposure during training data usage, existing training data security measures typically employ methods such as sensitive field identification, data anonymization, identifier anonymization, access control, permission management, encrypted storage, or controlled execution environments to identify and protect sensitive content within the training data. These methods can reduce the risk of sensitive fields being directly accessed or used without authorization.

[0004] However, in power training corpora, some factual elements, when used individually, do not constitute restricted facts, but there may be definite power industry-related connections between multiple factual elements. For example, facts such as equipment voltage level, protection action type, circuit breaker status change, equipment topology connection method, and operating parameter change may not directly determine specific restricted power operation facts when used individually. However, when specific electrical topology relationships, state transition relationships, operation timing relationships, or operating parameter response relationships are met, the combination of multiple unrestricted facts may progressively narrow down the range of candidate electrical objects or operating states, and ultimately determine a specific electrical object, a specific operating state, a specific operating parameter, or other restricted power facts.

[0005] Therefore, simply deleting, replacing, anonymizing, or generalizing a single sensitive field or factual element makes it difficult to identify the derivation process of restricted facts formed by multiple unrestricted facts and their relationships. Even if the training corpus does not contain fields or text that directly describe restricted facts, it is still possible to redetermine restricted power facts through electrical topology relationships, equipment state transition relationships, event relationships, operational timing relationships, and operating parameter relationships among multiple facts.

[0006] On the other hand, directly deleting all factual elements related to the restricted facts to block the derivation of the aforementioned restricted facts may disrupt the factual relationship structure required for the power large-scale language model to complete a specific training task. For example, in equipment status analysis, protection action analysis, or fault handling training tasks, the relationships between equipment status, protection actions, operation timing, and parameter responses may themselves be professional knowledge required to complete the training task. Extensive deletion of relevant facts may cause the training corpus to fail to maintain the factual relationships required for the target training task, reducing the effectiveness of the training corpus for the target task.

[0007] Therefore, a training corpus processing scheme is needed for training large-scale power language models. This scheme should be able to identify the derivation process of multiple unrestricted power facts gradually forming restricted facts through power professional relationships. While maintaining the facts and relationships required for the target training task, the scheme should make targeted adjustments to the derivation paths that form restricted facts. This will ensure that the reconstructed training corpus maintains the factual relationship structure required for the target training task, while reducing the possibility that restricted power facts can be re-determined through the combination of multiple unrestricted facts. Summary of the Invention

[0008] (a) Technical problems to be solved The purpose of this invention is to provide a method and system for restricted fact derivation suppression and corpus reconstruction for training large-scale power language models, in order to solve the problems of difficulty in identifying the formation of restricted facts from multiple unrestricted power facts through power professional correlation during the security processing of existing training corpora, and the problem that directly deleting relevant facts can easily destroy the fact association structure required by the target training task.

[0009] This invention establishes a factual association structure for candidate power training corpora, determines the necessary factual relationships and paths for the specific power large language model training task, and identifies the derivation path from multiple unrestricted facts to restricted facts based on power fact derivation rules. Furthermore, it determines candidate adjustment elements based on the overlap between the restricted fact derivation path and the necessary factual relationships for the task, selects target adjustment elements while maintaining the factual association structure required for the target training task, and reconstructs the training corpora to suppress the derivation process of multiple unrestricted power facts combining to form restricted facts.

[0010] (II) Technical Solution To achieve the above objectives, this invention provides a method for restricted fact inference suppression and corpus reconstruction for training large-scale power language models, comprising the following steps.

[0011] S1, Training Task Analysis Obtain the power big language model training task to be executed and the candidate power training corpus, parse the power big language model training task, and determine the target power fact type and target association type corresponding to the power big language model training task.

[0012] The power big language model training task is used to characterize the target capabilities to be obtained by training or optimizing the power big language model using candidate power training corpora, including at least one of equipment status analysis, protection action analysis, fault handling analysis, operation mode analysis, operation behavior analysis, and operation parameter correlation analysis.

[0013] The target power fact type is used to characterize the category of power facts that need to be learned to complete the corresponding training task, including at least one of electrical object facts, equipment state facts, operating event facts, operational behavior facts, and operating parameter facts; the target association type is used to characterize the association between facts that need to be maintained to complete the corresponding training task, including at least one of electrical topology relationships, equipment state transition relationships, event association relationships, operational timing relationships, and operating parameter association relationships.

[0014] S2, Constructing the Factual Relationship Structure Extract factual elements representing electrical objects, equipment status, operational events, operational behaviors, operating parameters, and time series information from candidate power training corpora. Each factual element includes at least a fact type, a fact value, and the corresponding electrical object, and may also include time information, source identifier, and fact identifier.

[0015] In one implementation, the fact element Fi is represented as: Fi = {Typei, Valuei, Objecti, Timei} Where Typei represents the fact type, Valuei represents the fact value, Objecti represents the electrical object corresponding to the fact, and Timei represents the time information corresponding to the fact.

[0016] Establish the fact association structure corresponding to the candidate power training corpus based on at least one of the following: electrical topology relationship, equipment state transition relationship, event association relationship, operation timing relationship, and operating parameter association relationship among fact elements.

[0017] For example, for protection action facts Fa, circuit breaker state change facts Fb, and active power change facts Fc in the candidate power training corpus, when Fa and Fb correspond to the same electrical object or have a preset association between protection equipment and switching equipment, and Fb and Fc occur within a preset time window, an event-state association between Fa and Fb and an operation-parameter response association between Fb and Fc are established respectively.

[0018] S3, Determine the necessary factual relationships and necessary factual paths for the task. Based on the target power fact type and target relation type, the fact elements that need to be retained to complete the current power big language model training task and the relations between fact elements are determined in the fact relation structure, forming the set of fact relations required for the task.

[0019] The set of factual relationships required for the task is represented as follows: T = {T1, T2, …, Tn} Each Ti represents a factual element or relation that needs to be retained to complete the current training task.

[0020] Based on the connections between fact elements and relationships in the set of factual relationships required for the task, at least one factual path required for the task is formed: Q = {Q1, Q2, …, Qr} Among them, the required fact path Qh for each task is used to represent a set of continuous or semantically constrained fact association structures that need to be maintained to complete the current training task.

[0021] For example, in a training task that associates protection actions with changes in equipment status, the training objective requires the model to learn the association between protection actions, corresponding circuit breaker status changes, and changes in equipment operating status. The protection action facts, circuit breaker status facts, equipment status facts, and their temporal and object associations required to form the above association are determined as the necessary factual relationships for the task, and the corresponding necessary factual paths for the task are formed.

[0022] S4, Identify the limited facts and perform multi-level fact deduction. Restricted facts are determined based on the corpus usage constraints corresponding to the current power large-scale language model training task. These restricted facts are power facts that, under the current training task or the current corpus usage scope, need to be limited from being directly obtained from candidate power training corpora or derived through the combination of multiple factual elements.

[0023] The restricted facts include at least one of the following: the identity of a specific electrical object, the actual topological location of a specific electrical object, a specific operating mode, a specific fault handling strategy, restricted operating parameters, and the association between a specific device and a critical electrical object.

[0024] Obtain a pre-established set of rules for deriving electrical facts: R = {R1, R2, …, Rm} Each power fact derivation rule includes at least a rule identifier, input fact type, correlation conditions between input facts, rule application conditions, and output fact type.

[0025] In one embodiment, the power fact derivation rule Rj is expressed as: Rj = {IDj, Inputj, Conditionj, Applyj, Outputj} Wherein IDj represents the rule identifier, Inputj represents the input fact type, Conditionj represents the association condition that needs to be satisfied between the input facts, Applyj represents the rule application condition, and Outputj represents the output fact type generated when the corresponding conditions are satisfied.

[0026] The association condition includes at least one of an electrical topological relationship, a device state transition relationship, an event association relationship, an operation sequence relationship, and an operating parameter association relationship; the rule application condition is used to limit the electrical object, operating state, time range or operating parameter condition corresponding to the fact elements participating in the corresponding power fact derivation rule.

[0027] For example, the device state transition rule can be expressed as: Device state = closed position + operation behavior = opening operation → subsequent device state = open position The operation sequence rule can be expressed as: Protection action occurs at t1 + circuit breaker state change occurs at t2 + 0<t2-t1≤ΔT → there is an action response relationship between the protection action and the circuit breaker state change Where ΔT is a time window preset according to the device type, protection type or training task.

[0028] Taking the unrestricted fact in the fact association structure as the starting fact, adding the starting fact to the current derivable fact set, and performing matching according to the input fact type corresponding to the power fact derivation rule.

[0029] When multiple fact elements in the current derivable fact set satisfy the input fact type required by a certain power fact derivation rule, and the multiple fact elements satisfy the corresponding electrical topological relationship, device state transition relationship, event association relationship, operation sequence relationship or operating parameter association relationship, the corresponding power fact derivation rule is executed to generate an output fact.

[0030] Adding the output fact as an intermediate derivation fact to the current derivable fact set, so that the intermediate derivation fact can jointly participate in the subsequent power fact derivation rule matching with other fact elements in the current derivable fact set, thereby forming a multi-level fact derivation process.

[0031] For example: {Fa,Fb} →R1 G1 {G1,Fc} →R2 G2 {G2,Fd} →R3 G3 Where G1 and G2 represent intermediate derivation facts generated step by step, and G3 represents subsequent output facts; when G3 satisfies the preset determination conditions corresponding to the restricted facts, the corresponding restricted facts Fs are determined to be derivable.

[0032] When the output facts obtained through step-by-step derivation meet the preset determination conditions corresponding to the restricted facts, the corresponding restricted facts are determined to be derivable.

[0033] The preset determination conditions include at least one of the following: the output fact uniquely corresponds to the restricted fact; the number of candidate electrical objects determined based on the output fact is not greater than a preset number threshold; the operating parameters corresponding to the output fact fall within the preset restricted parameter range corresponding to the restricted fact; and the output fact satisfies the preset power knowledge rules corresponding to the restricted fact.

[0034] When no new intermediate derivation facts can be generated during the derivation process, the current derivation state is the same as the derivation state that has been processed, or the derivation level reaches the preset maximum derivation depth Lmax, the corresponding derivation branch is terminated.

[0035] For restricted facts that meet the preset conditions, record the fact elements, relationships, and power fact derivation rules corresponding to each level of derivation from the initial fact to the restricted fact, forming a restricted fact derivation path.

[0036] A restricted factual derivation path is represented as: Pj = {Fstart, R1, Fmid1, R2, Fmid2, …, Fs} When a rule for deriving an electricity fact corresponds to multiple input facts, the multiple input facts are collectively recorded as the set of preceding facts for the corresponding output fact.

[0037] For the same constrained fact, when different starting facts, different intermediate derived facts, different relationships, or different power fact derivation rules can make it meet the preset conditions, the corresponding constrained fact derivation paths are recorded respectively, forming: Ps = {P1, P2, …, Pm} It continues to search for other valid derivation paths that meet the preset search criteria.

[0038] S5, Identify candidate adjustment elements and target adjustment elements. By comparing the factual elements and relationships in each restricted factual derivation path with the set of factual relationships required for the task, the overlap between the restricted factual derivation path and the set of factual relationships required for the task can be determined.

[0039] The overlapping relationship is used to characterize whether the factual elements or relationships used in the restricted fact derivation process also belong to the factual elements or relationships that need to be retained to complete the current training task.

[0040] Candidate adjustment elements are determined based on the overlapping relationships. Fact elements or relationships that are located on the restricted fact derivation path but do not belong to the set of fact relationships required for the task are considered as candidate adjustment elements; fact elements or relationships that belong to both the restricted fact derivation path and the set of fact relationships required for the task can also be considered as candidate adjustment elements if the type of fact or relationship required for the target training task can be maintained through adjustment.

[0041] For each candidate adjustment element Ci, at least one preset adjustment method Ai corresponding to the candidate adjustment element is determined, and the preset adjustment is simulated and executed on the fact association structure corresponding to the candidate power training corpus to obtain the simulated adjusted fact association structure.

[0042] The preset adjustment methods include at least one of the following: fact abstraction, fact replacement, identifier generalization, running parameter range, disconnection of association, and fact removal.

[0043] The restricted fact derivation path judgment and task-necessary fact retention judgment are re-executed on the simulated and adjusted fact association structure to determine the impact of the preset adjustment on the restricted fact derivation process and the fact structure required for the target training task.

[0044] For candidate adjustment element Ci, re-evaluate whether each restricted fact derivation path in the original restricted fact derivation path set still satisfies the corresponding power fact derivation rules and preset determination conditions.

[0045] definition: Bij = 1 indicates that after performing the preset adjustment on candidate adjustment element Ci, the j-th restricted fact derivation path Pj becomes invalid; otherwise: The derivation path blocking number BRi corresponding to the candidate adjustment element Ci when Bij = 0 is: BRi = ΣBij For each restricted fact Fsk, when all valid restricted fact derivation paths corresponding to the restricted fact Fsk become invalid after the preset adjustment is performed on the candidate adjustment element Ci, the restricted fact Fsk is updated from a derivable state to an underivable state.

[0046] Definition: Gik = 1 indicates that after adjusting Ci, all valid derivation paths corresponding to the restricted fact Fsk become invalid, causing Fsk to change from a derivable state to an underivable state; otherwise: The restricted fact blocking number FRi corresponding to candidate adjustment element Ci when Gik = 0 is: FRi = ΣGik BRi is used to characterize the number of valid derivation paths of restricted facts that are rendered invalid by the candidate adjustment, and FRi is used to characterize the number of restricted facts that are changed from a derivable state to an underivable state by the candidate adjustment. For the same restricted fact, even if some derivation paths have been rendered invalid, as long as there is still at least one valid derivation path that meets the preset determination conditions, the restricted fact will not be included in the FRi of the corresponding candidate adjustment element.

[0047] The simulation-adjusted fact association structure is used to redetermine whether each task-required fact or association in the set of task-required fact relationships still satisfies the target power fact type and target association type corresponding to the current training task.

[0048] Definition: Lik = 1 indicates that after performing the preset adjustment on candidate adjustment element Ci, the necessary fact or relationship Tk for the k-th task becomes invalid; otherwise: The task-required fact relation loss TLi corresponding to candidate adjustment element Ci when Lik = 0 is: TLi = ΣLik When the specific fact value is changed after the candidate adjustment element is processed, but the processed fact can still express the fact type and relationship required by the current training task, the corresponding task-required fact or relationship is not determined as invalid.

[0049] For example, if the "differential protection action of main transformer No. 1 in 220kV A substation" is generalized to "differential protection action of main transformer in 220kV substation", and the current training task only requires learning the correlation between the differential protection action of the main transformer and the change of circuit breaker status, then the generalization of the specific substation and equipment identity does not necessarily cause the corresponding task-necessary factual relationship to become invalid.

[0050] Further determine whether the necessary fact paths for each task are still valid.

[0051] Definition: Hih = 1 indicates that after performing the preset adjustment on the candidate adjustment element Ci, the required fact path Qh for the h-th task becomes invalid; otherwise: Hih = 0. The required fact path loss TPi for the candidate adjustment element Ci is: TPi = ΣHih When the fact elements or relationships required to constitute a necessary fact path for a task cannot continue to satisfy the target fact type, target relationship, or connection condition corresponding to the necessary fact path for the task after adjustment, the necessary fact path for the task is determined to be invalid.

[0052] Based on the preset task retention conditions, candidate adjustment elements that do not meet the task retention requirements are first eliminated.

[0053] The preset task retention conditions are determined based on the training task, including requiring the core task essential fact paths to remain valid, requiring pre-specified task essential facts and relationships to remain valid, or for training tasks with multiple equivalent task essential fact paths, requiring at least one of the following to remain valid: a preset number or a preset proportion of valid task essential fact paths.

[0054] Among the candidate adjustment elements that meet the preset task preservation conditions, priority is given to the candidate adjustment elements with a larger restricted fact blocking number FRi; when FRi is the same, priority is given to the candidate adjustment elements with a larger derivation path blocking number BRi; when both FRi and BRi are the same, priority is given to the candidate adjustment elements with a smaller task-necessary fact path loss number TPi; when FRi, BRi and TPi are all the same, the candidate adjustment element with a smaller task-necessary fact relationship loss number TLi is selected as the target adjustment element.

[0055] When multiple candidate adjustment elements have the same FRi, BRi, TPi, and TLi, the target adjustment element can be further determined based on at least one of the following secondary selection conditions: the granularity of factual information retained after adjustment, the number of training samples that need to be modified, the number of factual associations involved in the adjustment, and the degree of change to the original semantic expression of the candidate power training corpus.

[0056] S6, Corpus Reconstruction and Closed-Loop Verification After identifying at least one target adjustment element, perform at least one of the following processing steps on the target adjustment element: fact abstraction, fact replacement, identifier generalization, operating parameter range, relational disconnection, and fact removal. Then, reconstruct the candidate power training corpus based on the processed fact elements and relational relationships.

[0057] After the corpus is reconstructed, the factual association structure is updated based on the reconstructed factual elements and relationships. The retention status of the task-necessary factual relationships and task-necessary factual paths is re-determined, and the restricted factual derivation path is searched again according to step S4.

[0058] When the reconstructed factual association structure simultaneously satisfies the following conditions, the reconstructed candidate power training corpus is determined as the controlled training corpus corresponding to the current power large language model training task: the task-necessary factual relations and task-necessary factual paths satisfy the preset task preservation conditions; there is no effective restricted fact derivation path that can make any restricted fact satisfy the corresponding preset determination conditions.

[0059] If, after processing the current target adjustment element, there are still restricted fact derivation paths that meet the preset determination conditions, the candidate adjustment elements corresponding to the remaining restricted fact derivation paths are re-determined based on the reconstructed fact association structure, and the candidate adjustment impact judgment and target adjustment element determination are re-executed.

[0060] Each iteration selects a next target adjustment element that can invalidate at least one currently valid restricted fact derivation path. The process iteratively executes target adjustment element determination, candidate power training corpus reconstruction, and restricted fact derivation path re-search until no valid restricted fact derivation path satisfying the preset determination conditions exists.

[0061] When there are still valid restricted fact derivation paths, but no candidate adjustment element that can invalidate at least one currently valid restricted fact derivation path while satisfying the preset task preservation conditions, stop generating controlled training corpus corresponding to the current power big language model training task.

[0062] For the successfully generated controlled training corpus, establish the correlation between the controlled training corpus and the corresponding power big language model training task, the original candidate power training corpus, the restricted facts, the restricted fact derivation path, the target adjustment elements and adjustment methods, and form a safe use traceability relationship.

[0063] (III) Beneficial Effects Compared with the prior art, the present invention has at least the following beneficial effects.

[0064] First, the present invention establishes a fact association structure based on electrical topology relationships, equipment state transition relationships, event association relationships, operation timing relationships, and operating parameter association relationships. It also uses power fact derivation rules to deduce multiple unrestricted facts step by step. This allows the invention to identify the derivation process of re-determining restricted power facts by combining multiple fact elements in the training corpus that do not directly determine restricted facts when used alone.

[0065] Second, the present invention allows intermediate derivation facts generated by the power fact derivation rules to continue to serve as input for subsequent power fact derivation rules, and records the derivation relationships between the initial facts, intermediate derivation facts, power fact derivation rules, and restricted facts; for cases where there are multiple valid derivation paths for the same restricted fact, they are recorded and searched separately, thereby reducing the possibility that restricted facts can still be re-determined through other paths after only cutting off a single derivation path.

[0066] Third, this invention establishes restricted fact derivation paths and task-necessary fact relationships and task-necessary fact paths respectively. By analyzing the overlapping relationship between the two types of structures, candidate adjustment elements are determined, and the impact of candidate adjustments on the restricted fact derivation process and the task-necessary fact structure is judged respectively. Thus, while suppressing restricted fact derivation, the fact association structure required by the target training task is maintained.

[0067] Fourth, the present invention characterizes the impact of candidate adjustment elements on the derivation process of restricted facts and the knowledge structure of the training task by deriving the path blocking number BRi, the restricted fact blocking number FRi, the task-necessary fact relation loss number TLi, and the task-necessary fact path loss number TPi, respectively, and determines the target adjustment element based on the said impact, thereby reducing unnecessary adjustments to fact elements that are irrelevant to the target training task.

[0068] Fifth, after processing the target adjustment elements, the present invention re-determines the retention status of the necessary facts and the necessary fact paths for the task based on the reconstructed fact association structure, and re-executes the search for restricted fact derivation paths. Controlled training corpus is output only when the preset task retention conditions are met and there are no valid restricted fact derivation paths that meet the preset determination conditions, thereby forming a closed-loop verification process after corpus reconstruction.

[0069] Sixth, for fact elements that are simultaneously located in the restricted fact derivation path and the task-necessary fact structure, the present invention can adopt processing methods such as fact abstraction, fact replacement, identifier generalization or operation parameter range-based processing while maintaining the target fact type and target correlation, instead of having to delete the corresponding facts, thereby reducing the impact of large-scale fact deletion on the integrity of professional knowledge expression in power training corpus. Attached Figure Description

[0070] Figure 1 This is a flowchart of the overall process of the restricted fact inference suppression and corpus reconstruction method for training large-scale power language models according to the present invention.

[0071] Figure 2 This is a schematic diagram of the fact association structure of the candidate power training corpus of the present invention.

[0072] Figure 3 This is a schematic diagram of the restricted fact multi-level derivation based on the power fact derivation rules of the present invention.

[0073] Figure 4 This is a schematic diagram illustrating the overlapping relationship between the restricted fact derivation path and the task-required fact path of the present invention.

[0074] Figure 5 This is a flowchart of the process for judging the influence of candidate adjustment elements and determining the target adjustment elements in this invention.

[0075] Figure 6 This is a schematic diagram illustrating the closed-loop verification of corpus reconstruction and restricted fact deduction in this invention. Detailed Implementation The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can adaptively set the specific fact types, power fact derivation rules, preset determination conditions, task retention conditions, and corpus adjustment methods.

[0076] Example 1: Training Task on Analysis of Main Transformer Protection Actions

[0077] This embodiment takes the power large language model training task for the analysis of main transformer protection actions as an example to explain the establishment of fact association structure in candidate power training corpus, determination of necessary facts for the task, multi-level derivation of restricted facts, determination of candidate adjustment elements, corpus reconstruction, and post-reconstruction verification process.

[0078] In this embodiment, the candidate power training corpus is the candidate training data of an existing power large language model, which includes at least equipment type, equipment voltage level, protection action, circuit breaker status, topology connection method, operating parameter changes, and corresponding timing information. This embodiment does not consider the process of generating candidate training corpus from raw data of the power monitoring system as a necessary step.

[0079] The current training task is to analyze the main transformer protection actions, with the goal of enabling the power large language model to learn the correlation between the main transformer protection actions, circuit breaker state changes, and operating parameter responses.

[0080] Based on the training task, the target power fact type is determined to include at least equipment type, protection action type, circuit breaker status, and changes in operating parameters; the target correlation type is determined to include at least the event correlation between protection actions and changes in circuit breaker status, the operational timing relationship between protection actions and changes in circuit breaker status, and the parameter response relationship between changes in circuit breaker status and changes in operating parameters.

[0081] Based on the aforementioned target power fact types and target relationship types, a set T of necessary fact relationships for the task is formed, and task-necessary fact paths Q1 and Q2 are formed. For example: Q1: Main transformer differential protection operates → Circuit breaker changes from closed to open position Q2: Main transformer differential protection trips → Circuit breaker status changes → Active power decreases The following factual elements were extracted from the candidate electricity training corpus: F1: Equipment voltage level = 220kV F2: Equipment Type = Main Transformer F3: Protection Action Type = Main Transformer Differential Protection Action F4: The status of the associated circuit breaker changes from closed to open. F5: Device topology connection method = T1 F6: Within 1.8 seconds after the protection action, the active power decreased from 186.4MW to 21.7MW. Wherein, T1 represents the specific topological connection characteristics that can further distinguish candidate electrical objects. The electrical object identifiers, topological connection characteristics, operating parameter values, and time parameters used in this embodiment are all used to illustrate the technical processing of the present invention and do not correspond to specific actual power grid operating objects or restricted operating data.

[0082] A factual association structure is established based on the electrical object associations, equipment topology relationships, timing relationships between protection actions and circuit breaker state changes, and parameter response relationships after circuit breaker action among the aforementioned factual elements.

[0083] Based on the corpus usage constraints corresponding to the current training task, the restricted facts Fs in this embodiment are determined to be the specific electrical object identity corresponding to a specific main transformer.

[0084] This embodiment uses the unique determination of candidate electrical objects as an example of a preset determination condition for constrained facts. That is, when the number of candidate electrical objects determined based on the current deducible facts is not greater than 1, the identity of the corresponding specific electrical object is determined to be in a deducible state. The number mentioned is only used for illustration in this embodiment. In other implementations, other number thresholds can be set or other preset determination conditions can be adopted based on the constraints of corpus usage.

[0085] In this embodiment, a joint constraint rule R1 for device attributes and protection actions is pre-configured. R1 takes the device voltage level, device type, protection action type, and corresponding circuit breaker state change as input. When the protection action and the circuit breaker correspond to the same main transformer device or satisfy the preset protection device-switching device association relationship, and the circuit breaker state change occurs within a preset time range after the protection action, the first candidate electrical object set G1 is output.

[0086] Configure topology constraint rule R2. R2 takes the first candidate electrical object set G1 and the specific topology connection feature T1 as input, performs topology matching between the candidate electrical objects in G1 and T1, and outputs the second candidate electrical object set G2.

[0087] Configure the operating parameter response constraint rule R3. R3 takes the second candidate electrical object set G2 and the changes in operating parameters after the protection action as input. When the parameter change occurs within a preset time window and the direction and range of parameter change satisfy the preset operating response relationship of the corresponding electrical object, it outputs the third candidate electrical object set G3.

[0088] Match F1, F2, F3, and F4 based on R1: F1 + F2 + F3 + F4 →R1 G1 get: G1={A, B, C} Since the number of candidate electrical objects is 3, which does not meet the preset determination conditions of this embodiment, the limited facts are still in an uncertain state.

[0089] The intermediate derivation fact G1 is continued as input to R2 and matched with F5: G1 + F5 → R2 G2 get: G2={A, C} The number of candidate electrical objects is 2, which still does not meet the preset determination conditions of this embodiment.

[0090] Continue to use the intermediate derivation fact G2 as input to R3, and match it with F6: G2 + F6 → R3 G3 get: G3={A} At this point, the candidate electrical object is uniquely determined, satisfying the preset determination conditions, and therefore the restricted fact Fs is in a deducible state.

[0091] The system records the initial facts, intermediate derivation facts, power fact derivation rules, and constrained facts used in the above derivation process, forming a constrained fact derivation path: P1={F1,F2,F3,F4,R1,G1,F5,R2,G2,F6,R3,G3,Fs} The constrained fact derivation path P1 is compared with the task-required fact relation set T and the task-required fact paths Q1 and Q2. Among them, the main transformer differential protection action represented by F3, the circuit breaker state change represented by F4, and the active power change represented by F6 belong to the fact structure that the current training task needs to learn; the topology identification features contained in F5 that can further distinguish specific candidate electrical objects do not belong to the specific object identity knowledge that the current training task must retain.

[0092] Based on the aforementioned overlap relationship, multiple candidate adjustment schemes are determined, including: C1: Remove F3; C2: Remove F4; C3: Perform topological fact abstraction on F5; C4: Range the runtime parameters of F6.

[0093] Perform simulation adjustments on C1 to C4 respectively, and re-execute the restricted fact derivation and task-necessary fact path judgment based on the fact association structure after simulation adjustment.

[0094] In one calculation result of this embodiment: Candidate Adjustment BRi FRi TPi TLi C1 1 1 2 2 C2 1 1 2 1 C3 1 1 0 0 C4 1 1 0 0 The above values ​​are only used to illustrate the target adjustment element determination process of this embodiment and do not constitute fixed parameters or fixed thresholds. When FRi, BRi, TPi, and TLi of C3 and C4 are all the same, the target adjustment elements are further determined according to the secondary selection conditions.

[0095] Assuming that C3 only needs to abstract a specific topological recognition fact, while C4 needs to intervalize the running parameters in multiple training samples, then C3 is selected as the target adjustment element based on the number of training samples to be modified and the degree of change to the original semantic expression.

[0096] F5 is subjected to factual abstraction. Before adjustment, F5 represents the topological connectivity feature T1 that can distinguish specific candidate objects in A, B, and C; after adjustment, F5' represents the higher-level topological connectivity feature T1' that satisfies the device category relationship required for the current training task but can no longer further distinguish A and C.

[0097] The candidate power training corpus is reconstructed based on the adjusted F5', and the factual association structure is updated.

[0098] Re-execute R1 through R3 based on the reconstructed factual association structure. R1 will still yield: G1 = {A, B, C}.

[0099] After using F5' for subsequent matching, since F5' no longer contains specific topological connection features for further distinguishing candidate electrical objects A and C, it is impossible to establish a unique object association between the operating parameter response represented by F6 and candidate electrical object A based on the topological connection features. Therefore, the object association condition corresponding to R3 cannot further distinguish candidate electrical objects A and C. Further combining with F6, we obtain: G3'={A, C}.

[0100] at this time: |G3'|=2 The preset determination conditions of this embodiment are not met, so Fs changes from a derivable state to a non-derivable state.

[0101] The necessary factual relationships and paths for the task are reassessed based on the reconstructed factual correlation structure. Since the facts of the main transformer differential protection action, the circuit breaker state change, the timing relationship between them, the active power decrease, and the target correlation between protection actions, state changes, and parameter responses still exist, both Q1 and Q2 remain valid, corresponding to: TP3=0 and TL3=0 The system determines that the reconstructed factual association structure satisfies the preset task retention conditions, and there is no effective restricted factual derivation path that can make Fs satisfy the preset determination conditions. Therefore, the reconstructed training corpus adjusted by C3 is determined as the controlled training corpus corresponding to the current main transformer protection action analysis training task.

[0102] Example 2: Multiple valid derivation paths exist for the same constrained fact In another implementation, there are three valid derivation paths for the same restricted fact Fs: Ps={P1,P2,P3} Candidate adjustment element C1 can disable P1 and P2, but cannot disable P3.

[0103] After adjusting C1: BR1=2 Since P3 can still ensure that Fs satisfies the preset determination conditions, therefore: FR1=0 At this point, Fs is not determined to be a non-derivable state.

[0104] Based on the adjusted fact association structure, the system re-searches for the remaining valid and limited fact derivation paths, and obtains: Premain={P3} And the candidate adjustment elements were redefined for P3.

[0105] When candidate adjustment element C2 meets the preset task retention conditions, and adjusting C2 can invalidate P3, C2 is determined as the next target adjustment element, forming: C*={C1,C2} After adjusting C2, search for restricted fact derivation paths again. When P1, P2, and P3 all fail and there are no other valid restricted fact derivation paths that meet the preset search conditions, update Fs from a derivable state to a non-derivable state.

[0106] Example 3: Task preservation and restricted fact blocking cannot be satisfied simultaneously. In another implementation, all adjustable fact elements and relationships contained in a certain restricted fact derivation path P4 are simultaneously part of the core task-essential fact path Qcore, and performing an available adjustment on any fact element or relationship will cause Qcore to fail.

[0107] At this point, none of the candidate adjustment schemes meet the preset task retention conditions, and the system cannot determine the target adjustment element that can both disable P4 and retain Qcore.

[0108] In this case, stop generating the controlled training corpus corresponding to the current training task, and record the current training task identifier, the identifier of the restricted fact derivation path that has not been cut off, the identifier of the core task necessary fact path that makes it impossible to adjust, and the corresponding restricted fact identifier.

[0109] VI. System Implementation This invention also provides a restricted fact inference suppression and corpus reconstruction system for training large language models for power systems, including a task parsing module, a fact association construction module, a task fact determination module, a restricted fact inference module, an adjustment element determination module, a corpus reconstruction module, and a tracing module.

[0110] The task parsing module is used to obtain the power big language model training task to be executed and the candidate power training corpus, and to determine the target power fact type and target association type based on the training task.

[0111] In one implementation, the training task is characterized by a task identifier, task type, target capability description, target fact type, and target relationship type. For a pre-configured training task, the corresponding target fact type and target relationship type are read through the task identifier; for a newly created training task, the corresponding target fact type and target relationship type are determined through the task configuration interface.

[0112] The fact association construction module is used to extract fact elements corresponding to electrical objects, equipment states, operating events, operating behaviors, operating parameters and timing information from the existing candidate power training corpus, and to establish a fact association structure based on at least one of the following: electrical topology relationship, equipment state transition relationship, event association relationship, operation timing relationship and operating parameter association relationship between fact elements.

[0113] The task fact determination module is used to determine the fact elements and relationships that need to be retained to complete the current training task from the fact association structure based on the target power fact type and target association type, forming a set of task-necessary fact relationships, and forming at least one task-necessary fact path based on the connection relationship between the fact elements and relationships in the set of task-necessary fact relationships.

[0114] When there are multiple combinations of facts that can express the target knowledge for the same training task, multiple task-necessary fact paths can be formed, and the core task-necessary fact paths and non-core task-necessary fact paths can be determined according to the training task configuration.

[0115] The restricted fact derivation module is used to obtain restricted facts and pre-established power fact derivation rules corresponding to the current training task, and to perform step-by-step fact derivation starting from unrestricted facts in the fact association structure.

[0116] During each level of derivation, the restricted fact derivation module determines candidate input facts based on the input fact type corresponding to the power fact derivation rules, and judges whether the candidate input facts satisfy the corresponding electrical topology relationship, equipment state transition relationship, event association relationship, operation timing relationship, or operating parameter association relationship; when the corresponding conditions are met, output facts are generated and added to the current derivable fact set.

[0117] When the output fact meets the preset determination conditions corresponding to the restricted fact, the starting fact, intermediate derivation facts, correlations and power fact derivation rules participating in each level of derivation are recorded to form the corresponding restricted fact derivation path; for the same restricted fact, other valid derivation paths that meet the preset search conditions are searched.

[0118] The adjustment element determination module is used to compare the restricted fact derivation path with the set of fact relationships required for the task, determine the overlap relationship between the two, and determine the candidate adjustment elements and corresponding preset adjustment methods based on the overlap relationship.

[0119] For each candidate adjustment element, the adjustment element determination module simulates the execution of the corresponding preset adjustment on the fact association structure, and determines the derivation path blocking number BRi, the restricted fact blocking number FRi, the task-necessary fact relationship loss number TLi, and the task-necessary fact path loss number TPi respectively based on the simulated adjusted fact association structure.

[0120] The adjustment element determination module first eliminates candidate adjustment elements that do not meet the task retention requirements based on the preset task retention conditions, and then determines at least one target adjustment element from the remaining candidate adjustment elements according to FRi, BRi, TPi and TLi in sequence.

[0121] The corpus reconstruction module is used to perform at least one of the following processes on the candidate power training corpus according to the determined target adjustment elements and corresponding adjustment methods: fact abstraction, fact replacement, label generalization, operating parameter range, relational disconnection and fact removal, to form the reconstructed power training corpus.

[0122] After completing a corpus reconstruction, the corpus reconstruction module updates the corresponding fact association structure, redetermines the preservation status of the necessary fact relationships and necessary fact paths for the task, and re-searches for effective restricted fact derivation paths.

[0123] When the reconstructed power training corpus meets the preset task preservation conditions and there are no valid restricted fact derivation paths that meet the preset determination conditions, the reconstructed power training corpus will be determined as controlled training corpus.

[0124] When a valid restricted fact derivation path still exists, the next target adjustment element is redefined. Each redefined target adjustment element invalidates at least one currently valid restricted fact derivation path, until no valid restricted fact derivation path exists, or no candidate adjustment element can simultaneously satisfy the task preservation condition and the restricted fact derivation path blocking requirement.

[0125] The tracing module is used to establish the relationship between the controlled training corpus and the power big language model training task, the original candidate power training corpus, restricted facts, restricted fact derivation paths, target adjustment elements, and corresponding adjustment methods.

[0126] In one implementation, the following security usage traceability relationship is established: Controlled_Corpus_ID → Task_ID → Original_Corpus_ID → Restricted_Fact_ID → Derivation_Path_ID → Adjustment_Element_ID → Adjustment_Method Through the aforementioned tracing relationship, it is possible to determine the reasons for changes in factual elements in the controlled training corpus, the corresponding training task, the restricted facts targeted, and the derivation path of the restricted facts before adjustment.

[0127] VII. Implementation Methods of Electronic Devices and Computer Programs The system described in this invention can be deployed on a power artificial intelligence training data processing server, a power large language model training platform, a data security processing platform, or a computing environment consisting of multiple computing nodes.

[0128] In one embodiment, the electronic device includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the aforementioned restricted fact inference suppression and corpus reconstruction method for training large-scale power language models.

[0129] The memory is used to store at least one of the following: candidate power training corpus, fact association structure, power fact derivation rules, task-required fact relationship set, task-required fact path, restricted fact derivation path, candidate adjustment elements, target adjustment elements, and controlled training corpus.

[0130] The method described in this invention can also be implemented by a computer program, which, when run on a processor, causes the processor to execute the steps of the above method; the computer program can be stored in a computer-readable storage medium.

[0131] Those skilled in the art will understand that all or part of the steps in the above embodiments can be completed by computer program instructions to related hardware, and each functional module can be set in the same computing device or different computing nodes according to actual deployment requirements, and each module can interact with each other through data interfaces.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any equivalent substitutions, modifications, or combinations made within the scope of the inventive concept of the present invention, based on the disclosed technical solutions and inventive concepts, shall fall within the scope of protection of the present invention.

Claims

1. A method for constrained fact-based inference suppression and corpus reconstruction for training large-scale language models in the power industry, characterized in that, Includes the following steps: S1, obtain the power big language model training task to be executed and the candidate power training corpus, parse the power big language model training task, and determine the target power fact type and target association type corresponding to the power big language model training task; S2, extract factual elements representing electrical objects, equipment states, operating events, operational behaviors, operating parameters, and timing information from the candidate power training corpus, and establish a factual association structure corresponding to the candidate power training corpus based on at least one of the electrical topology relationship, equipment state transition relationship, event association relationship, operation timing relationship, and operating parameter association relationship among the factual elements; S3, based on the target power fact type and the target association type, determine the fact elements and the associations between the fact elements that need to be retained to complete the power big language model training task in the fact association structure, form a set of task-required fact relationships, and form at least one task-required fact path based on the connection relationship between the fact elements and the associations in the set of task-required fact relationships. S4, determine the restricted facts based on the corpus usage constraints corresponding to the power big language model training task, and obtain the pre-established power fact inference rules, which include the input fact type, the association conditions between the input facts, and the output fact type; Starting with the unrestricted facts in the aforementioned fact association structure, the fact elements and their relationships are matched step by step according to the power fact derivation rules. When the fact types of multiple fact elements and at least one of the electrical topology relationship, equipment state transition relationship, event relationship, operation timing relationship, and operating parameter relationship among the multiple fact elements satisfy the input conditions of the corresponding power fact derivation rules, the output fact is determined according to the power fact derivation rules, and the output fact is used as the input fact for subsequent power fact derivation rules to continue the derivation. When the output fact obtained by step-by-step derivation satisfies the preset determination conditions corresponding to the restricted fact, the fact elements, relationships, and corresponding power fact derivation rules from the starting fact to the restricted fact are recorded to form a restricted fact derivation path. S5, compare the fact elements and relationships in the restricted fact derivation path with the set of task-essential fact relationships, determine the overlap relationship between the restricted fact derivation path and the set of task-essential fact relationships, and determine candidate adjustment elements based on the overlap relationship; for each candidate adjustment element, determine the restricted fact derivation path that becomes invalid after the candidate adjustment element is preset adjusted, and the restricted fact that changes from a derivable state to an underivable state, and determine the task-essential facts or relationships that become invalid after the preset adjustment, as well as the invalid task-essential fact paths; Candidate adjustment elements that do not meet the preset task retention conditions are excluded, and among the remaining candidate adjustment elements, at least one target adjustment element is determined based on the derivation path of the failed restricted facts and the restricted facts that change from a derivable state to an underivable state. S6, perform at least one of the following processing steps on the at least one target adjustment element: fact abstraction, fact replacement, identifier generalization, operating parameter range, relational disconnection, and fact removal, and reconstruct the candidate power training corpus; Based on the reconstructed fact association structure, the retention status of the necessary fact relationships and the necessary fact paths for the task is re-determined, and the restricted fact derivation path is searched again according to step S4. When the required factual relations and the required factual paths of the task satisfy the preset task preservation condition and there is no restricted factual derivation path that satisfies the preset determination condition, the reconstructed power training corpus is determined as the controlled training corpus corresponding to the power big language model training task. When there are still restricted fact derivation paths that satisfy the preset determination conditions, candidate adjustment elements and at least one target adjustment element are re-determined for the remaining restricted fact derivation paths, and the candidate power training corpus is reconstructed until the task-necessary fact relations and the task-necessary fact paths satisfy the preset task preservation conditions and there are no restricted fact derivation paths that satisfy the preset determination conditions.

2. The method according to claim 1, characterized in that, The power fact derivation rules include rule identifiers, input fact types, correlation conditions between input facts, rule application conditions, and output fact types. The association conditions include at least one of electrical topology relationships, equipment state transition relationships, event association relationships, operation timing relationships, and operating parameter association relationships; The applicable conditions of the rules are used to limit the electrical objects, operating states, time ranges or operating parameter conditions corresponding to the factual elements participating in the corresponding power fact derivation rules.

3. The method according to claim 2, characterized in that, According to the aforementioned power fact derivation rules, factual elements and their relationships are matched step by step, including: For the current set of derivable facts, identify multiple fact elements that satisfy the input fact type and association conditions of any power fact derivation rule, generate intermediate derivable facts according to the corresponding power fact derivation rule, and record the derivation relationship between the multiple fact elements, the corresponding power fact derivation rule, and the intermediate derivable facts; The intermediate derivation facts are added to the current derivable fact set, and the intermediate derivation facts and other fact elements in the current derivable fact set are used as inputs for subsequent power fact derivation rules until the preset derivation termination condition is met.

4. The method according to claim 3, characterized in that, The preset determination conditions include at least one of the following: The output fact uniquely corresponds to the restricted fact; The number of candidate electrical objects determined based on the output facts is no greater than a preset number threshold; The operating parameters corresponding to the output fact fall within the preset restricted parameter range corresponding to the restricted fact; The output facts satisfy the preset power knowledge rules corresponding to the restricted facts.

5. The method according to claim 3, characterized in that, The restricted fact derivation path includes at least the starting fact, intermediate derivation facts, restricted facts, and the corresponding power fact derivation rules between adjacent facts; When a rule for deriving an electricity fact corresponds to multiple input facts, the multiple input facts are collectively recorded as the set of preceding facts for the corresponding output fact; For the same restricted fact, record the restricted fact derivation paths that satisfy the preset determination conditions and have different fact elements, relationships or power fact derivation rules.

6. The method according to claim 1, characterized in that, For each candidate adjustment element, determine the blocking effect of the preset adjustment on the restricted fact derivation path, including: Perform corresponding preset adjustments on candidate adjustment elements, and re-determine whether each restricted fact derivation path satisfies the corresponding power fact derivation rule and the preset determination condition based on the adjusted fact association structure; The number of restricted fact derivation paths that fail due to the preset adjustment is counted to obtain the derivation path blocking number BRi for the corresponding candidate adjustment element; For each restricted fact, when all the derivation paths of its corresponding restricted fact fail, the restricted fact is determined to change from a derivable state to a non-derivable state, and the number of restricted facts that change from a derivable state to a non-derivable state is counted to obtain the restricted fact blocking number FRi.

7. The method according to claim 6, characterized in that, Determining the impact of the preset adjustment on the set of factual relationships and the path of factual relationships required for the task for each candidate adjustment element includes: Based on the adjusted fact association structure, determine the task-required facts or associations that no longer satisfy the target power fact type or target association type due to the preset adjustment, and count the number of failed task-required facts or associations to obtain the task-required fact relationship loss number TLi; Identify the required fact paths of the task that have failed due to the preset adjustment, and count the number of failed required fact paths of the task to obtain the required fact path loss number TPi.

8. The method according to claim 7, characterized in that, Determining the at least one target adjustment element includes: Candidate adjustment elements that do not meet the preset task retention conditions are excluded from the task-essential fact path loss number TPi. Among the remaining candidate adjustment elements, priority is given to those with a larger restricted fact blocking number (FRi). When multiple candidate adjustment elements have the same number of restricted facts blocking FRi, the candidate adjustment element with the larger number of derivation path blocking BRi is selected first. When multiple candidate adjustment elements have the same number of restricted fact blocking FRi and number of derivation path blocking BRi, the candidate adjustment element with the smaller number of task-required fact path loss TPi is preferred. When multiple candidate adjustment elements have the same number of restricted facts blocking FRi, number of derivation path blocking BRi, and number of task-necessary fact path loss TPi, the candidate adjustment element with the smaller number of task-necessary fact relationship loss TLi is selected as the target adjustment element.

9. The method according to claim 8, characterized in that, When there are still restricted fact deduction paths that satisfy the preset determination conditions after processing the currently determined target adjustment elements, the candidate adjustment elements corresponding to the remaining restricted fact deduction paths are re-determined based on the reconstructed fact association structure, and the next target adjustment element is determined from the candidate adjustment elements that can invalidate at least one currently valid restricted fact deduction path in the manner described in claim 8. The process involves iteratively determining the target adjustment elements, reconstructing candidate power training corpora, and re-searching for restricted fact derivation paths until no restricted fact derivation path that meets the preset determination conditions is found. When there are no candidate adjustment elements that can invalidate at least one currently valid restricted fact derivation path while satisfying the preset task preservation conditions, stop generating controlled training corpus corresponding to the current power big language model training task.

10. A constrained fact-based inference suppression and corpus reconstruction system for training large-scale language models in the power industry, characterized in that, include: The task parsing module is used to obtain the power big language model training task to be executed and the candidate power training corpus, parse the power big language model training task, and determine the target power fact type and target association type corresponding to the power big language model training task. The fact association construction module is used to extract fact elements representing electrical objects, equipment status, operating events, operating behaviors, operating parameters and time series information from the candidate power training corpus, and to establish the fact association structure corresponding to the candidate power training corpus based on the association relationship between the fact elements. The task fact determination module is used to determine the set of necessary task fact relationships based on the target power fact type and the target association relationship type, and to form at least one necessary task fact path based on the connection relationship of fact elements and association relationships in the set of necessary task fact relationships; The restricted fact deduction module is used to deduce the fact elements and their relationships in the fact association structure step by step according to the power fact deduction rules, and to form a restricted fact deduction path when the output fact obtained by step-by-step deduction satisfies the preset determination conditions corresponding to the restricted fact. The adjustment element determination module is used to determine the overlap between the restricted fact derivation path and the set of task-essential fact relationships, and to determine at least one target adjustment element based on the changes in the restricted fact derivation path, restricted facts, task-essential facts or relationships, and task-essential fact paths after the candidate adjustment element is adjusted. The corpus reconstruction module is used to perform at least one of the following processes on the at least one target adjustment element: fact abstraction, fact replacement, identifier generalization, running parameter range, relational disconnection, and fact removal to reconstruct the candidate power training corpus. Based on the reconstructed fact relational structure, it redetermines the retention status of the task-required fact relations and the task-required fact paths and researches restricted fact derivation paths. When the task-required fact relations and the task-required fact paths meet the preset task retention conditions and there are no restricted fact derivation paths that meet the preset determination conditions, a controlled training corpus is formed. The traceability module is used to establish the association between the controlled training corpus and the corresponding power big language model training task, candidate power training corpus, restricted facts, restricted fact derivation path, and at least one target adjustment element, forming a traceability relationship for the safe use of the controlled training corpus.

Images