Business rule intelligent conversion method and device for natural language to code conversion

Abstract

The application discloses a natural language-to-code conversion business rule intelligent conversion method and device, relates to the intelligent operation and maintenance technical field, and comprises the following steps: analyzing an operation and maintenance diagnosis rule natural language description into a semantic constraint representation, and extracting a behavior constraint representation from a labeled example set; generating a candidate data analysis code set by jointly constraining the two, and projecting the candidate data analysis code set to a related structure space to obtain a related structure decision boundary; measuring the equivalent class collapse degree of the labeled example set in the related structure space to locate a constraint blind area; synthesizing a discrimination example which simultaneously falls into a signal space identifiable neighborhood and which forms a related structure discrimination direction in the constraint blind area on the value distribution of the selected analysis data field set, and returning to flow labeling to update the constraint; measuring related structure consistency and solidifying a versioned business expert experience model unit according to the related structure consistency, so that the related structure decision boundary of the constraint sequence correlation diagnosis is updated with a small amount of labeling, and the generated code is behaviorally controllable in the related structure space.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and in particular to a method and apparatus for intelligent conversion of business rules from natural language to code. Background Technology

[0002] In intelligent operation and maintenance scenarios, transforming the diagnostic experience of business experts into executable data analysis code is a crucial step in achieving automatic assessment of equipment operating status. In the traditional development model, business experts propose diagnostic requirements, which are then understood and code written by IT personnel. Significant communication losses and conversion delays exist between these two processes. When diagnostic rules need to be changed, the entire process of requirement transmission and code writing must be repeated, resulting in a deployment cycle of several weeks for new rules. This makes it difficult to adapt to the frequent iterations of diagnostic logic in operation and maintenance scenarios.

[0003] In recent years, with the development of natural language to code conversion technology, the industry has begun to explore directly converting natural language descriptions of requirements into executable code using models. This allows business experts to participate in rule definition without relying on IT personnel. For example, Chinese patent application CN118535135A discloses a training method and code generation method for a cloud-based code generation model. This method trains the code generation model using paired samples consisting of function-level code snippets and natural language descriptions. The model is deployed on a cloud server, and the terminal sends prompts and samples to customize the target application scenario. The prompts are iteratively updated based on the processing results of the samples, thereby improving the accuracy of code generation while decoupling the online model from user-owned private data. However, this solution treats the samples provided by business experts only as feedback signals for updating prompts, assuming that the representativeness of the samples in the observable signal space is equivalent to their constraint on the generated code. However, this assumption is not valid in the context of sequence correlation diagnostic logic, which is prevalent in operation and maintenance diagnostics.

[0004] The decision-making logic for sequence correlation diagnosis occurs in the correlation structure space, which is the space spanned by the correlation strength between sequences at different lags. Meanwhile, the instinct of business experts to select and label samples occurs in the observable signal space, which is the space composed of the amplitude and shape of the sequences. The correlation operation imposes a nontrivial invariance on its input: affine scaling and partial monotonic deformation of the sequences maintain the correlation value, and small phase shifts between sequences flip the sign of the lag correlation. This invariance decouples the morphological similarity in the observable signal space from the decision equivalence in the correlation structure space. Several samples selected by business experts based on morphological similarity often repeatedly fall into the same equivalence class in the correlation structure space. The direct consequence is that no matter how many samples are provided by business experts, the decision boundary of the diagnostic code converted from natural language is never sufficiently constrained in the correlation structure space, leading to uncontrollable decision-making behavior in correlation structures not sampled by business experts. This core bottleneck is something that the existing schemes that constrain code generation based on the representativeness of signal space samples cannot overcome.

[0005] Therefore, how to fully constrain the relevant structural decision boundary of sequence correlation diagnostic logic under the constraint that business experts can only provide a small number of labeled samples in the observable signal space, so as to make the behavior of the data analysis code obtained from natural language conversion controllable in the relevant structural space, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the bottleneck in existing intelligent operation and maintenance technologies where the decision boundary of the relevant structure of sequence correlation diagnosis logic cannot be sufficiently constrained by a limited number of signal space samples provided by business experts, this invention provides a method and apparatus for intelligent conversion of business rules from natural language to code. By migrating sample constraints from the observable signal space to the relevant structure space and synthesizing discriminative samples using dual-manifold constraints that simultaneously satisfy both signal space identifiability and relevant structure space discrimination, this invention achieves sufficient constraint of the relevant structure decision boundary and controllable behavior of the generated code's relevant structure, based on the principle of mismatch between the sampling manifold and the decision manifold caused by the invariance of correlation operations, even when business experts provide only a limited number of labeled samples in the observable signal space.

[0007] The technical solution of this invention is: a business rule intelligent conversion method from natural language to code, comprising the following steps: obtaining a natural language description of operation and maintenance diagnosis rules input by a business expert, selecting a set of analysis data fields and a set of labeled examples; parsing the natural language description of the operation and maintenance diagnosis rules into a semantic constraint representation, and extracting a behavioral constraint representation from the set of labeled examples; using the semantic constraint representation and the behavioral constraint representation as constraints to generate a set of candidate data analysis codes that pass through the set of labeled examples, and projecting the relevance judgment of the candidate data analysis code set to the input sequence onto a relevant structure space to obtain a relevant structure decision boundary; measuring the equivalence class collapse degree of the set of labeled examples in the relevant structure space, and locating the constraint blind zone in the relevant structure decision boundary that is not constrained by the set of labeled examples; selecting analysis data... Discriminant samples are synthesized on the value distribution of the field set. The discriminant samples simultaneously fall into the identifiable neighborhood of the signal space of the labeled sample set and span the relevant structure discrimination direction in the constraint blind zone. The discriminant samples are fed back to the business expert annotation to update the labeled sample set, and the step of generating a candidate data analysis code set is returned. The behavior of the converged target data analysis code in the relevant structure space is measured to be consistent with the relevant structure represented by the semantic constraint. When the relevant structure consistency meets the solidification criterion, the operation and maintenance diagnosis rule natural language description, the labeled sample set and the target data analysis code are bound into a versioned business expert experience model unit and solidified online. When the relevant structure consistency does not meet the solidification criterion, the inconsistency location information is fed back to request the refinement of the operation and maintenance diagnosis rule natural language description.

[0008] This invention also provides a business rule natural language to code intelligent conversion device, comprising: a dual representation parsing module, used to acquire the operation and maintenance diagnosis rule natural language description, selected analysis data field set, and labeled sample set input by business experts, parse the operation and maintenance diagnosis rule natural language description into semantic constraint representation, and extract behavioral constraint representation from the labeled sample set; a candidate code and related structure boundary generation module, used to generate a candidate data analysis code set that passes through the labeled sample set by jointly constraining the semantic constraint representation and the behavioral constraint representation, and project the correlation judgment of the candidate data analysis code set to the input sequence onto the related structure space to obtain the related structure decision boundary; and a sampling manifold collapse diagnosis module, used to measure the equivalence class collapse degree of the labeled sample set in the related structure space, and determine... The constraint blind zone in the relevant structural decision boundary that is not constrained by the labeled sample set; the dual-manifold constraint sample synthesis and backflow module is used to synthesize discrimination samples on the value distribution of the selected analysis data field set. The discrimination samples simultaneously fall into the identifiable neighborhood of the signal space of the labeled sample set and span the relevant structural discrimination direction in the constraint blind zone. The discrimination samples are backflowed to the business expert annotation to update the labeled sample set and trigger the candidate code and relevant structural boundary generation module to regenerate; the relevant structural consistency solidification module is used to measure the consistency between the behavior of the converged target data analysis code in the relevant structural space and the relevant structural representation of the semantic constraint, and solidify the versioned business expert experience model unit or backflow request the refinement of the natural language description of the operation and maintenance diagnosis rules.

[0009] The beneficial effects of this invention are as follows: Firstly, this invention measures the equivalence class collapse degree of the labeled sample set in the relevant structural space, using the coverage of relevant structural equivalence classes as the constraint sufficiency criterion, rather than the number of samples in the signal space. The mechanism lies in the invariance of correlation operations, which decouples the similarity of signal space morphology from the equivalence determination in the relevant structural space. Samples selected by business experts based on morphological similarity collapse into a few equivalence classes in the relevant structural space. Using equivalence class coverage as the criterion directly exposes the truly unconstrained constraint blind spots in the relevant structural decision boundary, fundamentally overcoming the shortcomings of existing schemes that misjudge constraint sufficiency based on the representativeness of signal space samples.

[0010] Secondly, this invention synthesizes the discriminative examples using a dual-manifold constraint that simultaneously satisfies both signal space identifiability and related structural space discrimination. The mechanism is that examples synthesized solely by spanning the discrimination direction in the related structural space often fall into unidentifiable regions of the signal space, making them unrecognizable and unlabelable by business experts. Conversely, examples obtained by random perturbation within the identifiable neighborhood of the signal space fall into the original equivalence class due to the invariance of correlation operations, thus failing to constrain the constraint blind zone. This invention maximizes the related structural displacement in the constraint blind zone direction under morphological distance constraints, enabling the synthesized discriminative examples to be both identifiable and labeled by business experts, while also effectively constraining the related structural decision boundaries. The synergy of these two aspects produces an effect that neither constraint alone can achieve, namely, constraining the related structural decision boundaries with a minimal number of labels within the business expert's capability boundary.

[0011] Third, this invention establishes a closed-loop coupling between the candidate code and related structural boundary generation, the sampling manifold collapse diagnosis, the dual-manifold constraint sample synthesis and reflow, and the related structural consistency solidification. The mechanism is that the result of the related structural consistency solidification triggers the dual-manifold constraint sample synthesis and reflow or requests the refinement of the natural language description of the operation and maintenance diagnosis rules. The annotation results of the discrimination samples update the sampling manifold collapse diagnosis and the related structural decision boundaries, making the annotation rounds logarithmically related to the size of the candidate data analysis code set rather than linearly related. With this closed loop, when rules change, business experts only need to update the natural language description of the operation and maintenance diagnosis rules and the annotation sample set to automatically regenerate and verify the target data analysis code, compressing the new rule deployment cycle from weekly to hourly, and achieving readability, maintainability, and rollbackability of diagnostic experience through versioned business expert experience model units. Attached Figure Description

[0012] Figure 1 This is a flowchart of the intelligent conversion method for business rules from natural language to code according to the present invention.

[0013] Figure 2 This is a schematic diagram of the structure of the business rule natural language to code intelligent conversion device of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0015] This embodiment provides a method for intelligent conversion of business rules from natural language to code. (See also...) Figure 1The method is applied to an intelligent operation and maintenance scenario for industrial equipment. In this scenario, business experts need to define an operation and maintenance diagnostic rule for early warning of thermo-mechanical coupling faults. The diagnostic logic is as follows: when the vibration amplitude sequence of the equipment spindle and the temperature sequence of the bearing show a strong positive correlation within a window with a lag of 0 to 5 sampling points, the correlation strength is not less than 0.8, and this strong positive correlation lasts for no less than 10 sampling points, it is determined as an early warning of a thermo-mechanical coupling fault. The equipment operation data is stored in the form of a data frame, with fields including timestamp, vibration amplitude, and bearing temperature, etc., where the vibration amplitude ranges from 0 to 50 mm, the bearing temperature ranges from 0 to 150 °C, and the sampling period is 1 second. The method includes the following steps.

[0016] Step S1, Dual Representation Parsing. Obtain the natural language description of the operation and maintenance diagnosis rules, the selected set of analysis data fields, and the set of labeled examples input by the business expert. The natural language description of the operation and maintenance diagnosis rules is the diagnostic logic text written in natural language by the business expert; the selected set of analysis data fields consists of the vibration amplitude field and the bearing temperature field selected by the business expert from the data frame; the set of labeled examples consists of several sequence fragments extracted and labeled by the business expert from historical operating data. Each labeled example includes a vibration amplitude sequence, a corresponding bearing temperature sequence, and a judgment label provided by the business expert, where the judgment label is either positive or negative.

[0017] The operation and maintenance diagnosis rules are parsed from natural language descriptions into semantic constraint representations. This step uses a pre-trained language model to perform semantic parsing on the natural language descriptions of the operation and maintenance diagnosis rules, extracting diagnostic conditions, data field references, and logical relationships to obtain the structured semantic constraint representations. These semantic constraint representations explicitly record the expected relevant structure of the diagnostic logic, namely, a strong positive correlation with a correlation strength of not less than 0.8 within a lag window of 0 to 5 sampling points, a correlation strength threshold of 0.8, and a time-series constraint lasting for at least 10 sampling points. This semantic parsing is an existing natural language processing technique in the field, and its specific implementation is not limited in this embodiment.

[0018] Behavioral constraint representations are extracted from the labeled sample set. For each labeled sample, the correspondence between its input sequence fragment and the judgment label given by the business expert is recorded, constituting the behavioral constraint representation. The behavioral constraint representation describes the diagnostic logic in an extended manner, that is, what the diagnostic result should be on these specific sequence fragments.

[0019] Therefore, the natural language description of the operation and maintenance diagnosis rules is parsed into the semantic constraint representation at the connotation level, and the set of labeled examples is extracted into the behavioral constraint representation at the extension level. The two serve as the dual constraint source for subsequent steps.

[0020] Step S2: Candidate code and related structural boundaries are generated. The semantic constraint representation and the behavioral constraint representation are used together to generate a set of candidate data analysis codes that pass the labeled sample set. This step uses both the semantic constraint representation and the behavioral constraint representation as constraints for code generation. Multiple candidate data analysis codes are generated by the code generation model, and the candidate data analysis codes that pass all labeled samples in the labeled sample set are retained to form the candidate data analysis code set. Each candidate data analysis code in the candidate data analysis code set is Python code based on a data frame, capable of reading the selected set of analysis data fields and outputting diagnostic judgments. During the generation process, constraints are applied to the field types, missing value distribution, and time index of the selected set of analysis data fields, ensuring that the candidate data analysis code set performs type alignment, forward filling of missing values, and time index alignment for the vibration amplitude field and bearing temperature field, guaranteeing that the code can be stably executed on the data frame. To obtain the candidate data analysis code set with behavioral diversity, code generation uses diversity sampling with a temperature coefficient, preferably 0.8, and the initial candidate data analysis code set size is preferably 16 samples.

[0021] It should be noted that the analysis codes from multiple candidate data sets in the labeled sample set may yield different diagnostic judgments on inputs not covered by the labeled sample set. This is precisely the external manifestation of the underconstraint of the correlation structure decision boundary. To explicitly characterize this underconstraint, this step further projects the correlation judgments of the candidate data analysis code set on the input sequence onto the correlation structure space to obtain the correlation structure decision boundary.

[0022] Specifically, for any input vibration amplitude sequence and bearing temperature sequence, the normalized correlation strength at each hysteresis value is calculated using the following formula, forming a hysteresis-correlation strength characterization: ,in: For lag The normalized correlation strength at point is a scalar, and its value range is . Dimensionless, calculated by this formula, characterizing the hysteresis between the vibration amplitude sequence and the bearing temperature sequence. The degree of linear correlation at each location; It is a value with lag, and is an integer scalar with a range of values. The unit is sampling points, determined by enumeration of the lag window length L, representing the relative shift of the two sequences on the time axis; L is the lag window length, a positive integer scalar, with a value range of... Sampling points, preferred The sampling point, measured in units of sampling points, is determined by the hysteresis window semantics in the natural language description of the operation and maintenance diagnosis rules. If L is too small, it cannot cover the required hysteresis range for diagnosis; if L is too large, irrelevant hysteresis noise is introduced. It represents the span of the hysteresis dimension of the relevant structure. t is the sampling time index, a positive integer scalar, with a value range of... The unit is the sampling point, determined by the sequence length and the lag value, representing the time position of the summation; N is the length of the input sequence, a positive integer scalar, with a value range of... Sampling points, measured in units of sampling points, are determined by the length of the sequence segment of the labeled sample and represent the number of sequence points involved in the correlation calculation. Let be the value of the vibration amplitude sequence at sampling time t, and let be a scalar with a range of values. The unit is mm, which is read from the selected set of analytical data fields and characterizes the vibration amplitude of the equipment spindle. For the bearing temperature sequence at sampling time The value of is a scalar, and its range is . The unit is ℃, which is read from the selected set of analytical data fields and characterizes the bearing temperature; Let be the mean of the vibration amplitude sequence, and be a scalar with a range of values. The unit is mm. The calculated value serves as a reference for characterizing the vibration amplitude. Let be the mean of the bearing temperature series, and be a scalar with a value range of . The unit is ℃, from The calculated reference temperature characterizes the bearing. For summation operators, For the arithmetic square root operator, This is the absolute value operator.

[0023] Stacking the above normalized correlation strengths over all lag values ​​yields the correlation structure characterization: ,in: For the characterization of related structures, for A column vector of dimensions, with each component taking values ​​within a range of 1. , dimensionless, is calculated and stacked by the above formula over all lag values, and represents the position of candidate data analysis code or annotation sample in the relevant structure space; Indicates vector transpose; The definitions of L and L are the same as in the aforementioned formula. The diagnostic logic corresponding to each candidate data analysis code in the candidate data analysis code set is mapped to a decision region in the relevant structure space, and the boundary of each decision region is the relevant structure decision boundary. The difference between the candidate data analysis codes in the candidate data analysis code set on the relevant structure decision boundary is a quantitative characterization of the underconstraint of the relevant structure decision boundary.

[0024] Step S3, sampling manifold collapse diagnosis. The equivalence class collapse degree of the labeled sample set is measured in the relevant structure space to locate constraint blind spots in the relevant structure decision boundary that are not constrained by the labeled sample set.

[0025] First, each labeled example in the labeled example set is projected onto the related structure space according to the aforementioned related structure representation formula to obtain the related structure representation of the labeled example. Then, the related structure similarity between any two labeled examples is calculated using the following formula: ,in: Let be the structural similarity between the k-th labeled example and the 1st labeled example, and let be a scalar with a value range of . , dimensionless, is calculated by this formula and characterizes the degree of closeness between two labeled samples in the relevant structural space; , For the relevant structural representation of the k-th and 1-th labeled examples, A dimensionless column vector, defined as in the aforementioned structural characterization formula; The L2 norm operator for vectors; is the similarity bandwidth, which is a positive scalar with a value range of . Dimensionless, preferred The scale is determined by the related structural characterization. If the value is too small, the equivalence class division will be too fine. If the value is too large, the equivalence class division will be too coarse, which is the scale for judging the equivalence classes of related structures. This is the natural exponentiation operator; k and 1 are the annotation sample indices, which are positive integer scalars with a value range of 1 to 1. Furthermore, eigenvalue decomposition is performed on the similarity matrix composed of all relevant structural similarities, and the equivalence class collapse degree of the labeled sample set is calculated using the following formula: Where: C is the sampling manifold collapse degree, which is a scalar with a value range of 10 ... The dimensionless value C, calculated using this formula, indicates that the labeled sample set collapses into fewer related structural equivalence classes in the related structural space, and the constraint blind zone is larger. Conversely, C approaching 0 indicates a more dispersed distribution of the labeled sample set in the related structural space, serving as a constraint sufficiency criterion. M is the number of labeled samples in the labeled sample set, a positive integer scalar, with a value range of... The unit is "one", which is determined by the set of labeled samples and represents the number of labeled samples participating in the collapse metric. Let be the i-th eigenvalue of the similarity matrix, and let be a scalar with a value range of . , dimensionless, is obtained by eigenvalue decomposition of the similarity matrix, and characterizes the strength of the i-th main distribution pattern in the related structure space; Here, is the normalized eigenvalue, and is a scalar with a range of values. and Dimensionless, calculated from the right-hand side of this formula, representing the proportion of the i-th related structural equivalence class pattern; i and j are eigenvalue indices, positive integer scalars, with values ​​ranging from 1 to 1. ; The effective equivalence class number is the exponent of the Shannon entropy of the normalized eigenvalue distribution, representing the number of relevant structural equivalence classes actually occupied by the labeled sample set. The operator is the natural logarithm operator. When the equivalence class collapse degree C is high, it indicates that although the number of labeled sample sets is considerable, they only occupy a few related structural equivalence classes in the related structural space, and there are large unconstrained regions in the related structural decision boundary. This step uses the coverage of related structural equivalence classes, i.e., the ratio of the number of effective equivalence classes to the number of equivalence classes required for the related structural decision boundary, as the sufficiency criterion for the constraint of the labeled sample set on the related structural decision boundary. The lower the coverage, the larger the constraint blind zone.

[0026] Taking this embodiment as an example, the business expert intuitively extracted 12 labeled samples from historical operating data. These labeled samples exhibited morphological differences in amplitude and waveform length in the signal space, leading the business expert to believe that they covered various typical operating conditions. However, after the aforementioned collapse metric, it was found that the number of effective equivalence classes for these 12 labeled samples was only about 2.4, corresponding to an equivalence class collapse degree of about 0.8. This indicates that these seemingly diverse labeled samples in the signal space actually only fall near about two related structural equivalence classes in the correlation structure space. The reason for this is that the correlation operation maintains the correlation value unchanged when scaling the overall vibration amplitude sequence. The multiple labeled samples distinguished by the business expert based on amplitude are classified into the same equivalence class in the correlation structure space; while the correlation structure that the diagnostic logic focuses on, where the correlation strength undergoes sign flipping at the hysteresis boundary, happens to fall within the constraint blind zone that the business expert has never sampled. If the sufficiency of constraints is still judged based on the number of signal space samples, it will be mistakenly determined that the set of labeled samples is sufficient, thus solidifying a set of target data analysis code whose behavior is uncontrollable in the constraint blind zone. This step uses the equivalence class collapse degree as the criterion, directly exposing the aforementioned constraint blind zone.

[0027] Furthermore, this step utilizes the sign change of the normalized correlation strength with respect to the lag value in the lag-correlation strength characterization to further locate the phase-sensitive segment. Phase sensitivity is calculated using the following formula: , in: Let be the phase sensitivity, and be a scalar with a value range of . Dimensionless, obtained by calculation using this formula. The larger the value, the more drastic the change in normalized correlation strength with hysteresis near the peak hysteresis, and the more unstable the sign, which is used to locate phase-sensitive segments; For peak lag, is an integer scalar, and its value range is . The unit is the sampling point, which is... The location with the maximum absolute value of the normalized correlation strength is used to determine the lag position where the correlation is strongest; Let be the radius of the phase-sensitive neighborhood, and be a positive integer scalar with a value range of . Sampling points, preferred Sampling points, the unit is the number of sampling points, determined empirically. If the value is too small, the statistics will be unstable. If the value is too large, irrelevant lags will be mixed in, which is the statistical window that characterizes phase sensitivity; This is the point operator for finding the maximum value; , L The definition is the same as the aforementioned formula. The relevant structural space constraint blind zone reflected by the equivalence class collapse degree, together with the phase-sensitive segment located by the phase sensitivity, constitutes the constraint blind zone.

[0028] Step S4, dual-manifold constraint sample synthesis and backflow. Discriminant samples are synthesized on the value distribution of the selected analysis data field set. These discriminant samples simultaneously fall within the identifiable neighborhood of the labeled sample set's signal space and span the relevant structure discrimination direction within the constraint blind zone. To balance the identifiability for business experts with the discriminative power regarding the relevant structure decision boundary, this step uses an anchored sample from the labeled sample set as the center and searches for the sequence segment with the largest relevant structure displacement in the constraint blind zone direction within its identifiable neighborhood of signal space; this segment is then used as the discriminant sample. The optimization objective is as follows: , in: The synthesized discriminant sample is a sequence fragment of length N, i.e., an N-dimensional vector, whose units are consistent with the corresponding fields in the selected set of analytical data fields. It is obtained by this optimization solution and represents a new sample used to constrain the decision boundary of the relevant structure; z is an optimization variable sequence fragment, an N-dimensional vector. An anchor sample is an N-dimensional vector selected from the set of labeled samples, representing the center of the identifiable neighborhood in the signal space; The value distribution support set for the selected set of analytical data fields is determined by the value range of historical running data, and represents the feasible domain of the discrimination sample; , For sequence fragment z, The relevant structural characterization, for A dimensionless column vector, defined as in the aforementioned structural characterization formula; To constrain the direction of the blind spot, A unit column vector of dimension, satisfying , dimensionless, given by the main direction of the constraint blind zone, characterizing the relevant structural discrimination direction that needs to be constrained; This is the vector dot product operator; For z and The signal spatial morphology distance is a scalar, with a value range of [value missing]. After amplitude normalization, it becomes a dimensionless quantity, calculated from dynamic time warp distance or normalized Euclidean distance, and characterizes the morphological differences between two sequence segments in the signal space. The preset morphological distance threshold is a positive scalar with a value range of [value range missing]. Dimensionless, preferred This is determined by the capabilities that business experts can identify. If it is too small, the synthesis space will be limited. If the value is too large, business experts will be unable to identify the discrimination sample, representing the identifiable neighborhood radius in the signal space. The above optimization is solved using the projection gradient method, and its iterative format is as follows: ,in: For the first The sequence segment of the next iteration is an N-dimensional vector, with the same unit. This is obtained by updating the format in this iteration, representing the current solution in the optimization process; Let be the sequence segment of the i-th iteration, and be an N-dimensional vector; To reach the feasible region The projection operator represents the distribution support set of the updated result constraint retrieved values; The gradient operator for sequence segment z; Let be the iteration step size, and be a positive scalar with a range of values ​​of . Dimensionless, preferred Determined by convergence, it characterizes the magnitude of the optimization update. Here, is the constraint penalty coefficient, and is a positive scalar with a range of values ​​of . Dimensionless, preferred , characterizing the degree of enforcement of the signal spatial morphology distance constraint; i is the iteration number, a non-negative integer scalar, with a value range of . , The optimal value is 200, which represents the number of projection gradient iterations; The operator for extracting the non-negative part; , , , The definition is the same as the previous formula.

[0029] The synthesized discrimination example thus falls within the signal space identifiable neighborhood of the labeled example set, where the signal space morphological distance between it and the anchored example does not exceed the preset morphological distance threshold. Business experts can identify and label the anchored example based on their knowledge of it. At the same time, it has the largest relevant structural displacement in the constraint blind zone direction, thereby spanning the relevant structural discrimination direction in the constraint blind zone and forming an effective constraint on the relevant structural decision boundary.

[0030] The discriminant examples are fed back to the business expert annotations to update the annotation example set. This step presents the discriminant examples to the business expert using a binary or single-choice interactive method. The business expert only needs to determine whether the discriminant example is positive or negative, without writing any code. The annotation results from the business expert on the discriminant examples are received and incorporated into the annotation example set. Then, the process returns to step S2, where the updated annotation example set re-constrains the candidate data analysis code set.

[0031] Furthermore, when the diagnostic logic corresponding to the natural language description of the operation and maintenance diagnosis rule includes a combination of correlation conditions, threshold conditions, and duration conditions, the output divergence on a single discrimination example may be caused by the coupling of multiple conditions, making it difficult to attribute to a specific condition. Therefore, this step orthogonally decomposes the output divergence of the candidate data analysis code set on the discrimination example along the correlation structure dimension, scalar threshold dimension, and time-series duration dimension: , in: Let be the total output divergence of the candidate data analysis code set on the discriminant sample z, and let be a scalar with a value range of . , dimensionless, is given by the variance of the output of each candidate data analysis code in the candidate data analysis code set, and represents the total discriminative power of z in distinguishing the candidate data analysis code set; To determine the gradient, let be an N-dimensional vector, dimensionless, derived from... The calculation shows that the overall output divergence is characterized by the direction of the perturbation of the sample features. , , These are the orthogonal projection matrices for the relevant structure dimension, scalar threshold dimension, and temporal duration dimension, respectively. A matrix, dimensionless, satisfies Furthermore, the features are orthogonal to each other and constructed from the feature subspaces corresponding to the three types of diagnostic conditions, representing the decomposition of the total discriminative power into each diagnostic dimension; , , These are the discriminative power components for the relevant structural dimension, scalar threshold dimension, and temporal persistence dimension, respectively. They are scalars, all with values ​​greater than or equal to 0, and are dimensionless. They are calculated from the projection terms of this formula and represent the contribution of each dimension to distinguishing the candidate data analysis code set. for identity matrix; The L2 norm operator for vectors; This is the gradient operator. Based on the discriminative force component corresponding to the target condition, the discriminative examples activating the target condition are synthesized in a directional manner. For example, when it is necessary to constrain correlation conditions, in the aforementioned optimization, the discriminative examples are synthesized in a directional manner with the discriminative force component of the correlation structure dimension as the dominant direction. This allows the labeling results of the discriminative examples to directionally constrain the correlation structure decision boundary corresponding to the correlation condition, avoiding attribution difficulties caused by multi-condition coupling.

[0032] Step S5, Consolidation of Relevant Structures. The consistency between the behavior of the converged target data analysis code in the relevant structure space and the relevant structure represented by the semantic constraints is measured. This step measures the deviation between the expected region of the relevant structure defined by the semantic constraints and the behavior of the target data analysis code in the relevant structure space, and calculates the relevant structure consistency using the following formula: , in: For consistency of related structures, is a scalar with a value range of . Dimensionless, obtained by calculation using this formula. The closer the value is to 1, the more the behavior of the target data analysis code in the relevant structural space conforms to the semantic constraint representation, which is used as a solidification criterion. To verify the fragment set, The number of elements in the set of verification fragments is a positive integer scalar with a value range of 1. The sample base, represented by the number of historical operational data, is determined by dividing the data into units; j is the validation fragment index, a positive integer scalar with a value range of [value missing]. ; The relevant structural representation of the target data analysis code output on the j-th verification segment is... The dimensionless column vector is calculated using the aforementioned structural characterization formulas. The relevant structural expectation region defined by the semantic constraint representation is a subset of the relevant structural space, obtained by parsing the natural language description of the operation and maintenance diagnosis rules, and represents the relevant structure expected by the diagnosis logic; The distance from the relevant structure to the expected region of the relevant structure is a scalar with a value range of greater than or equal to 0. It is dimensionless and is calculated from the Euclidean distance from a point to a set. It represents the degree of deviation of the target data analysis code behavior from the expected relevant structure. For consistency tolerance, is a positive scalar, and its value range is . Dimensionless, preferred Determined by the allowable deviation in diagnosis. If the value is too small, the consistency judgment will be too strict. If the value is too large, the consistency judgment is too lenient, representing the allowable scale for the degree of deviation; The natural exponential operator. Steps S4 and S5 are executed iteratively, with each iteration reducing the candidate data analysis code set using one of the discriminant samples. Given that the discriminant sample has a single-round reduction ratio on the candidate data analysis code set, the number of annotation rounds required for convergence satisfies the following upper bound: ,

[0033] Where: T is the annotation round, a positive integer scalar, in rounds, determined by the upper bound, representing the number of business expert annotation interaction rounds required for convergence; Let be the single-round candidate reduction ratio, and let be a scalar with a value range of . Dimensionless, preferred This represents the proportion of the candidate data analysis code set reduced by each of the discrimination examples; The initial candidate data analysis code set size is denoted by , where is a positive integer scalar with a value range of . The unit is a set of candidate data analysis codes; This is the floor operator. The upper bound indicates that the number of annotation rounds in the labeled sample set is logarithmically related to the size of the candidate data analysis code set, rather than linearly related to the full coverage requirement of the relevant structural decision boundary. Therefore, a very small number of annotation rounds within the business expert capability boundary are sufficient to converge to the target data analysis code, which is unique in the relevant structural space.

[0034] Taking this embodiment as an example, the initial candidate data analysis code set is 16 pieces, and the single-round candidate reduction ratio is 0.5. Therefore, according to the upper bound, the number of annotation rounds required for convergence is no more than 4 rounds. This means that business experts only need to make binary judgments on 4 of the aforementioned discrimination examples to gradually reduce the 16 candidate data analysis codes from the original annotation example set to the target data analysis code with unique behavior in the relevant structural space. In contrast, if the existing scheme uses signal space examples to constrain the relevant structural decision boundary, since the relevant structural equivalence classes cannot be directly identified in the signal space, business experts often need to add dozens or even hundreds of annotation examples without achieving convergence, and repeatedly oscillate between conflicting constraints. Thus, the directional constraint force of the discrimination examples synthesized in this step on the direction of the constraint blind zone, combined with the logarithmic reduction of the candidate data analysis code set, produces an effect that cannot be achieved by simply increasing the number of signal space examples, by converging the relevant structural decision boundary with a very small number of annotations.

[0035] When the consistency of the relevant structures meets the solidification criterion (i.e., the consistency of the relevant structures is not lower than the consistency threshold and the equivalence class collapse degree is lower than the preset collapse threshold, preferably 0.9 and preferably 0.2), the natural language description of the operation and maintenance diagnosis rules, the set of labeled examples, and the target data analysis code are bound into a versioned business expert experience model unit and solidified online. When the consistency of the relevant structures does not meet the solidification criterion, the inconsistency location information is fed back to step S1, requesting the refinement of the natural language description of the operation and maintenance diagnosis rules; wherein the inconsistency location information indicates the specific dimension by which the target data analysis code deviates from the expected region of the relevant structure in the relevant structure space, and the business expert refines the natural language description of the operation and maintenance diagnosis rules accordingly.

[0036] Furthermore, when the natural language description of the operation and maintenance diagnosis rule changes, a triple comparison is performed between the revised version of the business expert experience model unit and the original version. The triple includes the natural language description of the operation and maintenance diagnosis rule, the set of labeled examples, and the target data analysis code. Based on the comparison result, the version of the business expert experience model unit can be rolled back, thus achieving readability, maintainability, and rollback capability of the diagnostic experience. With this closed loop, when rules change, business experts only need to update the natural language description of the operation and maintenance diagnosis rule and the set of labeled examples to automatically regenerate and verify the target data analysis code, compressing the new rule deployment cycle from weekly to hourly.

[0037] This embodiment provides a device for intelligent conversion of business rules from natural language to code. See [link / reference] Figure 2This system is used to execute the intelligent transformation method for business rules in the natural language to code conversion described in Embodiment 1. The system includes a dual-representation parsing module, a candidate code and related structure boundary generation module, a sampling manifold collapse diagnosis module, a dual-manifold constraint sample synthesis and reflow module, and a related structure consistency solidification module.

[0038] The dual-representation parsing module is used to acquire the natural language description of the operation and maintenance diagnosis rules, the selected set of analysis data fields, and the set of labeled samples input by business experts. It parses the natural language description of the operation and maintenance diagnosis rules into semantic constraint representations and extracts behavioral constraint representations from the set of labeled samples. In its specific implementation, the dual-representation parsing module includes a natural language parsing unit and a sample extraction unit: the natural language parsing unit uses a pre-trained language model to perform semantic parsing on the natural language description of the operation and maintenance diagnosis rules, extracting diagnostic conditions, data field references, and logical relationships to obtain the semantic constraint representations. These semantic constraint representations explicitly record the relevant structure, relevant strength thresholds, and temporal constraints expected by the diagnostic logic; the sample extraction unit records the correspondence between the input sequence fragment and the judgment label for each labeled sample in the set of labeled samples to obtain the behavioral constraint representations. The processing procedure of the dual-representation parsing module is the same as step S1 in Embodiment 1, and will not be repeated here.

[0039] The candidate code and related structure boundary generation module is used to generate a set of candidate data analysis codes that pass through the labeled sample set, constrained by the semantic constraint representation and the behavioral constraint representation. It then projects the relevance determination of the candidate data analysis codes to the input sequence onto the related structure space to obtain the related structure decision boundary. In a specific implementation, the candidate code and related structure boundary generation module includes a code generation unit and a related structure projection unit: the code generation unit, constrained by the semantic constraint representation and the behavioral constraint representation, generates multiple candidate data analysis codes from the code generation model and retains the candidate data analysis codes that pass through the labeled sample set. It also imposes constraints on the field type, missing value distribution, and time index of the selected analysis data field set, making the candidate data analysis code set a Python code based on a data frame; the related structure projection unit projects the relevance determination of the candidate data analysis codes to the input sequence onto the related structure space according to the related structure representation formula described in Embodiment 1 to obtain the related structure decision boundary. The processing procedure of the candidate code and related structure boundary generation module is the same as step S2 in Embodiment 1.

[0040] The sampling manifold collapse diagnosis module is used to measure the equivalence class collapse degree of the labeled sample set in the relevant structure space and locate the constraint blind zone in the relevant structure decision boundary that is not constrained by the labeled sample set. In a specific implementation, the sampling manifold collapse diagnosis module includes a collapse measurement unit and a blind zone location unit: the collapse measurement unit calculates the equivalence class collapse degree of the labeled sample set in the relevant structure space according to the relevant structure similarity formula and the equivalence class collapse degree formula described in Embodiment 1, and uses the coverage of the relevant structure equivalence class as the constraint sufficiency criterion; the blind zone location unit combines the equivalence class collapse degree with the phase-sensitive segment located according to the phase-sensitive formula described in Embodiment 1 to jointly determine the constraint blind zone. The processing procedure of the sampling manifold collapse diagnosis module is the same as step S3 in Embodiment 1.

[0041] The dual-manifold constraint sample synthesis and backflow module is used to synthesize discrimination samples on the value distribution of the selected set of analytical data fields. The discrimination samples simultaneously fall into the identifiable neighborhood of the signal space of the labeled sample set and span the relevant structure discrimination direction in the constraint blind zone. The discrimination samples are backflowed to the business expert annotation to update the labeled sample set, and the candidate code and relevant structure boundary generation module is triggered to regenerate. The dual-manifold constraint sample synthesis and reflow module, in its specific implementation, includes a constraint synthesis unit, a discriminant force decomposition unit, and an annotation interaction unit. The constraint synthesis unit, according to the optimization target and projection gradient iteration format described in Embodiment 1, maximizes the relevant structural displacement in the constraint blind zone direction under signal space morphological distance constraints to obtain the discriminant sample. When the diagnostic logic includes a combination of correlation conditions, threshold conditions, and duration conditions, the discriminant force decomposition unit decomposes the output divergence along the relevant structural dimension, scalar threshold dimension, and temporal duration dimension according to the discriminant force orthogonal decomposition formula described in Embodiment 1, and synthesizes the discriminant sample according to the discriminant force components corresponding to the target conditions. The annotation interaction unit presents the discriminant sample to the business expert using a binary selection or single-choice interaction method, receives the annotation results, and incorporates them into the annotation sample set. The processing procedure of the dual-manifold constraint sample synthesis and reflow module is the same as step S4 in Embodiment 1.

[0042] The related structure consistency solidification module is used to measure the consistency between the behavior of the converged target data analysis code in the related structure space and the related structure of the semantic constraint representation. When the related structure consistency meets the solidification criterion, the operation and maintenance diagnosis rule natural language description, the labeled sample set, and the target data analysis code are bound into a versioned business expert experience model unit and solidified online. When the related structure consistency does not meet the solidification criterion, the inconsistency location information is fed back to the dual representation parsing module to request the refinement of the operation and maintenance diagnosis rule natural language description. In specific implementation, the related structure consistency solidification module includes a consistency measurement unit, a solidification unit, and a version management unit: the consistency measurement unit calculates the related structure consistency according to the related structure consistency formula described in Embodiment 1; the solidification unit binds and solidifies the versioned business expert experience model unit when the related structure consistency meets the solidification criterion, otherwise the inconsistency location information is fed back; the version management unit performs a triplet comparison of the versioned business expert experience model unit before and after the change and supports version rollback when the operation and maintenance diagnosis rule natural language description changes. The processing procedure of the related structure consistency solidification module is the same as step S5 in Embodiment 1.

[0043] The above modules form a closed-loop coupling: the measurement result of the related structure consistency solidification module triggers the dual-manifold constraint sample synthesis and backflow module to continue synthesizing the discrimination sample, or backflows to the dual-representation parsing module to request refinement; the annotation result obtained by the dual-manifold constraint sample synthesis and backflow module updates the equivalence class collapse degree of the sampling manifold collapse diagnosis module and the related structure decision boundary of the candidate code and related structure boundary generation module, so that the entire system converges with an annotation round that is logarithmically related to the size of the candidate data analysis code set.

[0044] Each module in the system can be implemented by a processor executing computer programs stored in memory. The processor can be a central processing unit (CPU), other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). The memory can be a read-only memory (ROM), random access memory (RAM), or other read-write storage media. Those skilled in the art will understand that the above module division is only a logical functional division, and in actual implementation, there can be other division methods; for example, multiple modules can be combined or integrated into another system.

[0045] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A business rule intelligent conversion method for natural language to code conversion, characterized in that, Includes the following steps: S1. Obtain the natural language description of the operation and maintenance diagnosis rules input by the business expert, select the set of analysis data fields and the set of labeled examples, parse the natural language description of the operation and maintenance diagnosis rules into semantic constraint representation, and extract the behavioral constraint representation from the set of labeled examples. S2. Using the semantic constraint representation and the behavioral constraint representation as constraints, a candidate data analysis code set is generated through the labeled sample set, and the relevance determination of the input sequence by the candidate data analysis code set is projected onto the relevant structure space to obtain the relevant structure decision boundary; S3. Measure the equivalence class collapse degree of the labeled sample set in the relevant structure space, and locate the constraint blind zone in the relevant structure decision boundary that is not constrained by the labeled sample set; S4. Synthesize discrimination examples on the value distribution of the selected set of analytical data fields. The discrimination examples simultaneously fall into the signal space identifiable neighborhood of the labeled example set and span the relevant structure discrimination direction in the constraint blind zone. The discrimination examples are then fed back to the business expert annotation to update the labeled example set, and the process returns to S2. S5. The behavior of the target data analysis code after convergence in the relevant structural space is consistent with the relevant structural representation of the semantic constraints. When the consistency of the relevant structural representation satisfies the solidification criterion, the natural language description of the operation and maintenance diagnosis rules, the set of labeled examples, and the target data analysis code are bound as a versioned business expert experience model unit and solidified online. When the consistency of the relevant structural representation does not satisfy the solidification criterion, the inconsistency location information is fed back to S1 to request the refinement of the natural language description of the operation and maintenance diagnosis rules.

2. The intelligent conversion method for business rules from natural language to code as described in claim 1, characterized in that, The S3 step of measuring the equivalence class collapse degree of the labeled sample set in the relevant structural space includes: Project each annotation example in the annotation example set onto the relevant structure space to obtain the relevant structural representation of the annotation example in the relevant structure space; Based on the similarity between the related structural representations, the set of labeled examples is divided into several related structural equivalence classes; The coverage of the relevant structural equivalence classes is used as the constraint sufficiency criterion for the set of labeled examples on the decision boundary of the relevant structure. The lower the coverage, the larger the constraint blind zone.

3. The intelligent conversion method for business rules from natural language to code as described in claim 2, characterized in that, The correlation structure is characterized as a hysteresis-correlation strength characterization, which is composed of the normalized correlation strength of the labeled sample at multiple hysteresis values. The S3 further includes: determining the phase-sensitive segment of the diagnostic logic corresponding to the natural language description of the operation and maintenance diagnosis rule based on the sign change of the normalized correlation strength with the lag value in the lag-correlation strength characterization, and incorporating the phase-sensitive segment into the constraint blind zone.

4. The intelligent conversion method for business rules from natural language to code as described in claim 3, characterized in that, The process of synthesizing discriminative samples on the value distribution of the selected set of analytical data fields in step S4 includes: Using the value distribution of the selected set of analytical data fields as the support set, and under the constraint that the signal spatial morphological distance between the discriminant sample and an anchor sample in the set of labeled samples does not exceed a preset morphological distance threshold, the relevant structural displacement between the discriminant sample and the anchor sample in the direction of the constraint blind zone is maximized to obtain the discriminant sample.

5. The intelligent conversion method for business rules from natural language to code as described in claim 4, characterized in that, When the diagnostic logic corresponding to the natural language description of the operation and maintenance diagnostic rule includes a combination of correlation conditions, threshold conditions, and duration conditions, S4 further includes: The output divergence of the candidate data analysis code set on the discrimination sample is orthogonally decomposed along the correlation structure dimension, scalar threshold dimension and time series persistence dimension to obtain the discriminative power components of each dimension; Based on the discriminative power component corresponding to the target condition, the discriminative sample that activates the target condition is synthesized in a targeted manner.

6. The intelligent conversion method for business rules from natural language to code as described in claim 5, characterized in that, The behavior of the target data analysis code after metric convergence in S5 in the relevant structure space is consistent with the relevant structure represented by the semantic constraints, including: The degree of deviation between the expected region of the relevant structure defined by the semantic constraint representation and the behavior of the target data analysis code in the relevant structure space is measured to obtain the consistency of the relevant structure. S4 and S5 are executed iteratively. In each iteration, the candidate data analysis code set is reduced by one of the discrimination examples until the consistency of the related structure satisfies the solidification criterion and the equivalence class collapse degree is lower than the preset collapse threshold, so that the annotation rounds of the annotation example set and the size of the candidate data analysis code set are logarithmically related.

7. The intelligent conversion method for business rules from natural language to code as described in claim 1, characterized in that, The step S2 generates a set of candidate data analysis codes that pass the set of labeled samples, including: Based on the field types, missing value distribution, and time index of the selected set of analytical data fields, field type alignment constraints, missing value handling constraints, and time index alignment constraints are applied to the candidate data analysis code set, so that the candidate data analysis code set is Python code based on a data frame.

8. The intelligent conversion method for business rules from natural language to code as described in claim 1, characterized in that, Also includes: When the natural language description of the operation and maintenance diagnosis rule changes, the modified version of the business expert experience model unit is compared with the original version of the business expert experience model unit using a triplet. The triplet includes the natural language description of the operation and maintenance diagnosis rule, the set of labeled examples, and the target data analysis code. Based on the comparison results, support the version rollback of the versioned business expert experience model unit.

9. The intelligent conversion method for business rules from natural language to code as described in claim 1, characterized in that, In step S4, the discrimination sample is fed back to the business expert for annotation, including: The discrimination examples are presented to the business experts using a binary or single-choice interactive method. Receive the annotation results of the business experts on the discrimination examples, and incorporate the annotation results into the annotation example set.

10. A business rule intelligent conversion device for natural language to code conversion, used to implement the business rule intelligent conversion method for natural language to code conversion as described in any one of claims 1-9, characterized in that, include: The dual representation parsing module is used to obtain the natural language description of the operation and maintenance diagnosis rules input by business experts, the selected set of analysis data fields and the set of labeled examples, parse the natural language description of the operation and maintenance diagnosis rules into semantic constraint representations, and extract behavioral constraint representations from the set of labeled examples. The candidate code and related structure boundary generation module is used to generate a set of candidate data analysis codes that pass through the labeled sample set by jointly constraining the semantic constraint representation and the behavioral constraint representation, and to project the relevance judgment of the input sequence of the candidate data analysis code set onto the related structure space to obtain the related structure decision boundary; The sampling manifold collapse diagnosis module is used to measure the equivalence class collapse degree of the labeled sample set in the relevant structure space and locate the constraint blind zone in the relevant structure decision boundary that is not constrained by the labeled sample set. The dual-manifold constraint sample synthesis and backflow module is used to synthesize discrimination samples on the value distribution of the selected set of analysis data fields. The discrimination samples simultaneously fall into the signal space identifiable neighborhood of the labeled sample set and span the relevant structure discrimination direction in the constraint blind zone. The discrimination samples are backflowed to the business expert annotation to update the labeled sample set and trigger the candidate code and relevant structure boundary generation module to regenerate. The related structure consistency solidification module is used to measure the consistency between the behavior of the converged target data analysis code in the related structure space and the related structure of the semantic constraint representation. When the related structure consistency meets the solidification criterion, the operation and maintenance diagnosis rule natural language description, the labeled sample set and the target data analysis code are bound into a versioned business expert experience model unit and solidified online. When the related structure consistency does not meet the solidification criterion, the inconsistency location information is fed back to the dual representation parsing module to request the refinement of the operation and maintenance diagnosis rule natural language description.

Images