Intelligent analysis task standardization generation method based on large model and multi-industry scene demand

Abstract

The application relates to the technical field of big data analysis, and discloses a multi-industry scene demand intelligent analysis task standardization generation method based on a large model, which comprises the following steps: converting a demand text into a semantic tensor and decoupling the semantic tensor into an intention vector and a context vector; extracting a parameter boundary to construct a five-dimensional constraint matrix; splitting task nodes and establishing a directed edge to generate a directed acyclic graph based on a compatibility score; calculating a node deviation score, contracting a constraint boundary according to a sensitivity partial derivative to update the directed acyclic graph; calculating a matching score to bind a cascade pipeline; and outputting a standardized description file. The application converts probabilistic language generation into controlled tensor operation, eliminates semantic drift and topological dead lock during multi-source data mapping, realizes in-situ closed-loop error correction through partial derivative feedback, and greatly improves the logical self-consistency of a big data management process.

Description

Technical Field

[0001] This invention relates to a standardized generation method for intelligent analysis of multi-industry scenario requirements based on a large model, belonging to the field of big data analysis technology. Background Technology

[0002] Currently, in big data analytics systems, parsing unstructured requirement texts from various sources and in diverse formats into structured task sequences is a crucial step in business process management. With the development of generative large language model technology, existing solutions typically utilize the model's contextual understanding capabilities to identify requirement content in natural language form and generate corresponding business process task lists. These solutions have been applied to scenarios such as industrial production line collaborative scheduling, supply chain gap matching, and equipment inspection management. The aforementioned requirement texts often simultaneously contain information such as environmental conditions, technical specifications, equipment capabilities, and safety boundaries. Since large language models gradually generate text based on the probability of each word appearing in the vocabulary, their output has a certain degree of randomness. However, business management processes and automated control engines typically require input results to be clear, stable, and consistent with actual operating conditions.

[0003] In complex production environments, this probabilistic generation method is prone to parameter omissions, semantic shifts, or incorrect task node order. When the requirements text contains environmental constraints or mandatory safety specifications that are not easily identifiable, without explicit constraints on the model generation process, it is difficult to guarantee that the generated task topology conforms to actual operational patterns and safety requirements. To reduce the randomness of the generated results, existing solutions mainly improve the situation by adding prompt word rules or increasing the model parameter size. However, the continuous accumulation of prompt word rules can easily occupy too many context windows and cause conflicts between different rules; while increasing the model parameter size increases computational overhead, it is less effective in the face of... When extreme conditions not covered in the training data are encountered, it is still difficult to guarantee that the parameters will always be within the specified range, and it is also difficult to guarantee that the generated results will converge stably. For example, Chinese invention patent application with publication number CN120386603A discloses a method and device for data analysis based on multi-agent systems. Although this scheme introduces a large model to extract intent and plan a serial-parallel strategy, its data transmission relies excessively on static attribute mapping. In the dynamic evolution scenario of multi-source heterogeneous data intertwining, its underlying constraints present a fixed pattern and cannot be adaptively adjusted with the nonlinear changes of real-time load. Under complex conditions, it is easy to cause parameter transmission distortion and it is difficult to meet the requirements of high-precision real-time control.

[0004] Therefore, how to maintain the flexibility of unstructured text parsing while imposing clear parameter boundaries on the generation process, and how to reduce logical conflicts in the task topology through node compatibility verification and adaptive error correction, remains a problem that existing technologies need to solve. Summary of the Invention

[0005] To address the problems in the background technology, the technical solution of this invention is as follows: A standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, comprising the following steps:

[0006] Step S1: The sequence of requirement texts is converted into a semantic tensor through an embedding layer, and then decoupled into a business intent vector and an environmental context vector using a slicing operator.

[0007] Step S2: Extract the parameter boundaries of the two vectors and construct a five-dimensional constraint matrix containing the task objectives, environmental conditions, technical indicators, capability requirements, and safety boundary constraint vectors.

[0008] Step S3: Decompose the business intent vector into a set of task nodes. When the compatibility score between the output state vector of the preceding node and the capability requirement constraint vector of the following node is greater than or equal to 0.85, establish a directed dependency edge and generate a directed acyclic graph of the task.

[0009] Step S4: Input the task node into the retrieval module to recall the standard case node, calculate the deviation score between the two, and when the deviation score is greater than or equal to 0.30, solve the partial derivatives of each dimension to lock the highest sensitivity dimension, use a penalty factor of 0.15 to calculate the shrinkage amount of the boundary of this dimension and update the task directed acyclic graph.

[0010] Step S5: Calculate the matching score between the tool requirement vector of the task node and the capability representation vector of the candidate tools in the pre-set candidate tool library. When the matching score is greater than or equal to 0.70 and less than 0.90, connect the two candidate tools into a composite pipeline and bind it to the task node. Convert the task directed acyclic graph into a standardized description file output.

[0011] Preferably, the step of constructing the five-dimensional constraint matrix in step S2 includes: step S21, extracting data containing the ambient temperature range, spatial coordinate red line, and network latency upper limit from the environmental context vector to establish an environmental condition constraint vector; step S22, extracting data containing system accuracy indicators, time response window, and throughput threshold from the business intent vector to establish a technical indicator constraint vector; step S23, combining the environmental condition constraint vector, technical indicator constraint vector, task objective constraint vector, capability requirement constraint vector, and security boundary constraint vector into a multi-dimensional tensor to obtain the five-dimensional constraint matrix.

[0012] Preferably, the step of calculating the compatibility score in step S3 includes: step S31, extracting the output state vector of the preceding node and the capability requirement constraint vector of the following node; step S32, calculating the cosine similarity and normalized distance between the output state vector and the capability requirement constraint vector; step S33, weighted summing of the cosine similarity and normalized distance to obtain the compatibility score.

[0013] Preferably, the step of generating the directed acyclic graph of the task in step S3 further includes: step S34, traversing the pairwise node combinations in the task node set, setting the weight of the corresponding directed dependent edge to 0 when the compatibility score is less than 0.85, and setting the weight of the corresponding directed dependent edge to 1 when the compatibility score is greater than or equal to 0.85.

[0014] Preferably, when calculating the deviation score between the two in step S4, the following formula is used for calculation: ,in, Scoring based on deviation; It is set as the first weighting coefficient and its value is set to 0.6; The cosine similarity of the feature vectors between the task node and the standard case node; It is the second weighting coefficient and its value is set to 0.4; This is the average absolute value of the differences between the task node and the standard case node in each component of the five-dimensional constraint matrix.

[0015] Preferably, the step of updating the directed acyclic graph of the task in step S4 includes: step S41, marking the task node in the current state as an abnormal node when the deviation score is greater than or equal to 0.30; step S42, calculating the partial derivatives of the deviation score with respect to each dimension of the five-dimensional constraint matrix, and determining the dimension corresponding to the largest partial derivative as the highest sensitivity dimension; step S43, using a penalty factor of 0.15 to perform gradient adjustment on the boundary of the highest sensitivity dimension, converging the boundary of the highest sensitivity dimension within a specific gradient range reduced by the penalty factor, and calling the intention decomposition unit to generate alternative nodes to replace the abnormal nodes.

[0016] Preferably, the step of calculating the matching score in step S5 includes: step S51, extracting the tool requirement vector from the task node and retrieving the capability representation vector of each candidate tool in the candidate tool library; step S52, calculating the normalized value of the inner product between the tool requirement vector and the capability representation vector, and using the calculated normalized value of the inner product as the matching score.

[0017] Preferably, the method of binding to the task node in step S5 further includes: step S53, when the matching score is greater than or equal to 0.90, the corresponding tool interface information is directly written into the attribute field of the task node to complete the direct binding.

[0018] Preferably, after step S5, a closed-loop adaptive fine-tuning step is also included: Step S6, when a task failure signal is received from the downstream automated management engine, the running log vector of the corresponding failed node is extracted; the partial derivatives of the running log vector with respect to each dimension of the five-dimensional constraint matrix are calculated to locate the target constraint dimension; the shrinkage correction amount is calculated for the constraint boundary of the target constraint dimension and updated to the industry knowledge graph.

[0019] Preferably, the following preprocessing steps are included before step S1: Step S0, data cleaning is performed on the input original demand text sequence, and plaintext identity data and specific entity labels in the original demand text sequence are removed using text masking and high-dimensional hash desensitization mechanism to obtain the demand text sequence.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. In the standardized generation of intelligent parsing tasks for multi-industry scenario requirements in large models, the requirement text is decoupled into business intent vectors and environmental context vectors. A five-dimensional constraint matrix is ​​used to uniformly restrict task objectives, environmental conditions, technical indicators, capability requirements, and security boundaries. This ensures that the various constraints in natural language requirements can be accurately implemented at specific task nodes, thereby reducing task deviations and logical breaks caused by semantic ambiguity, parameter omissions, or unrecognized implicit conditions, and improving the accuracy and consistency of multi-industry requirement parsing results.

[0022] 2. By verifying the compatibility between the output status of the preceding node and the capability requirements of the following node, it is possible to determine whether each node can be properly connected before establishing task dependencies. This makes the generated directed acyclic graph of tasks more consistent with the actual business process, thereby reducing missing preconditions, incorrect node order, and task deadlock, and improving the logical consistency and execution stability of the task topology after it is input into the downstream management system.

[0023] 3. By comparing task nodes with standard case nodes and identifying the constraint dimensions with the greatest impact based on deviation scores, abnormal nodes can be corrected in a targeted manner, avoiding redundant regeneration. This method achieves precise positioning of problems and errors, enabling constraint boundaries to gradually and steadily converge, thereby reducing task nodes that do not meet business conditions or security requirements and ensuring the compliance of the generated results. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the requirements analysis and task topology generation process for this invention.

[0025] Figure 2 This is a functional module diagram of the task standardization generation system of the present invention.

[0026] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0028] A standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model includes the following steps:

[0029] Step S1: The sequence of requirement texts is converted into a semantic tensor through an embedding layer, and then decoupled into a business intent vector and an environmental context vector using a slicing operator.

[0030] Step S2: Extract the parameter boundaries of the two vectors and construct a five-dimensional constraint matrix containing the task objectives, environmental conditions, technical indicators, capability requirements, and safety boundary constraint vectors.

[0031] Step S3: Decompose the business intent vector into a set of task nodes. When the compatibility score between the output state vector of the preceding node and the capability requirement constraint vector of the following node is greater than or equal to 0.85, establish a directed dependency edge and generate a directed acyclic graph of the task.

[0032] Step S4: Input the task node into the retrieval module to recall the standard case node, calculate the deviation score between the two, and when the deviation score is greater than or equal to 0.30, solve the partial derivatives of each dimension to lock the highest sensitivity dimension, use a penalty factor of 0.15 to calculate the shrinkage amount of the boundary of this dimension and update the task directed acyclic graph.

[0033] Step S5: Calculate the matching score between the tool requirement vector of the task node and the capability representation vector of the candidate tools in the pre-set candidate tool library. When the matching score is greater than or equal to 0.70 and less than 0.90, connect the two candidate tools into a composite pipeline and bind it to the task node. Convert the task directed acyclic graph into a standardized description file output.

[0034] Preferably, the step of constructing the five-dimensional constraint matrix in step S2 includes: step S21, extracting data containing the ambient temperature range, spatial coordinate red line, and network latency upper limit from the environmental context vector to establish an environmental condition constraint vector; step S22, extracting data containing system accuracy indicators, time response window, and throughput threshold from the business intent vector to establish a technical indicator constraint vector; step S23, combining the environmental condition constraint vector, technical indicator constraint vector, task objective constraint vector, capability requirement constraint vector, and security boundary constraint vector into a multi-dimensional tensor to obtain the five-dimensional constraint matrix.

[0035] Preferably, the step of calculating the compatibility score in step S3 includes: step S31, extracting the output state vector of the preceding node and the capability requirement constraint vector of the following node; step S32, calculating the cosine similarity and normalized distance between the output state vector and the capability requirement constraint vector; step S33, weighted summing of the cosine similarity and normalized distance to obtain the compatibility score.

[0036] Preferably, the step of generating the directed acyclic graph of the task in step S3 further includes: step S34, traversing the pairwise node combinations in the task node set, setting the weight of the corresponding directed dependent edge to 0 when the compatibility score is less than 0.85, and setting the weight of the corresponding directed dependent edge to 1 when the compatibility score is greater than or equal to 0.85.

[0037] Preferably, when calculating the deviation score between the two in step S4, the following formula is used for calculation: ,in, Scoring based on deviation; It is set as the first weighting coefficient and its value is set to 0.6; The cosine similarity of the feature vectors between the task node and the standard case node; It is the second weighting coefficient and its value is set to 0.4; This is the average absolute value of the differences between the task node and the standard case node in each component of the five-dimensional constraint matrix.

[0038] Preferably, the step of updating the directed acyclic graph of the task in step S4 includes: step S41, marking the task node in the current state as an abnormal node when the deviation score is greater than or equal to 0.30; step S42, calculating the partial derivatives of the deviation score with respect to each dimension of the five-dimensional constraint matrix, and determining the dimension corresponding to the largest partial derivative as the highest sensitivity dimension; step S43, using a penalty factor of 0.15 to perform gradient adjustment on the boundary of the highest sensitivity dimension, converging the boundary of the highest sensitivity dimension within a specific gradient range reduced by the penalty factor, and calling the intention decomposition unit to generate alternative nodes to replace the abnormal nodes.

[0039] Preferably, the step of calculating the matching score in step S5 includes: step S51, extracting the tool requirement vector from the task node and retrieving the capability representation vector of each candidate tool in the candidate tool library; step S52, calculating the normalized value of the inner product between the tool requirement vector and the capability representation vector, and using the calculated normalized value of the inner product as the matching score.

[0040] Preferably, the method of binding to the task node in step S5 further includes: step S53, when the matching score is greater than or equal to 0.90, the corresponding tool interface information is directly written into the attribute field of the task node to complete the direct binding.

[0041] Preferably, after step S5, a closed-loop adaptive fine-tuning step is also included: Step S6, when a task failure signal is received from the downstream automated management engine, the running log vector of the corresponding failed node is extracted; the partial derivatives of the running log vector with respect to each dimension of the five-dimensional constraint matrix are calculated to locate the target constraint dimension; the shrinkage correction amount is calculated for the constraint boundary of the target constraint dimension and updated to the industry knowledge graph.

[0042] Preferably, the following preprocessing steps are included before step S1: Step S0, data cleaning is performed on the input original demand text sequence, and plaintext identity data and specific entity labels in the original demand text sequence are removed using text masking and high-dimensional hash desensitization mechanism to obtain the demand text sequence.

[0043] Example 1: Taking the receipt of multi-source unstructured requirement text regarding the inspection of high-voltage distribution cabinets under smart factory operating conditions as an example, plaintext identity data is replaced with mask characters using text masks according to the position of the identity field. Then, a high-dimensional hash desensitization mechanism is used to replace specific entity tags with fixed-dimensional hash codes, resulting in a requirement text sequence. Subsequently, a preset large model embedding layer is called to map this requirement text sequence to a high-dimensional semantic space, obtaining a 768-dimensional original requirement semantic tensor. And utilize semantic slicing operators to Independent component extraction and orthogonal decoupling are performed to separate and obtain the business intent vector representing the business action. The context vector representing the boundary of the parameters By combining the knowledge graph of the inspection domain to extract the parameter boundaries of the two vectors, from Data is extracted from the environment, including the temperature range, spatial coordinate red line, and network communication latency with an upper limit of 100ms, to establish an environmental condition constraint vector. ;from Data on system accuracy metrics, time response windows, and throughput thresholds are extracted to establish a technical metric constraint vector. The time response window is limited by a sampling period of 7200s; then the task objective constraint vector is... Environmental constraint vector Technical indicator constraint vector Capability requirement constraint vector and safety boundary constraint vector By concatenating the columns, we obtain the five-dimensional constraint matrix. Safety boundary constraint vector Includes Ex-d-II-C-T4 explosion-proof rating gate control value.

[0044] The intent decomposition unit decomposes the business intent vector according to the sequence of business actions and the input-output relationship between adjacent actions. Break it down into nodes that include navigation to the distribution cabinet. Nodes for acquiring infrared thermal imaging data Local temperature data comparison and analysis node and over-temperature alarm and image upload node The set of task nodes, the topology building units are based on the five-dimensional constraint matrix. Extracting the predecessor node The output state vector With subsequent nodes Capability requirement constraint vector Calculate the cosine similarity and normalized distance between the two and perform a weighted sum to obtain the compatibility score: ,in, The first weighting coefficient for the compatibility score is set to 0.6; This is the second weighting coefficient for the compatibility score, with a value set to 0.4; Preorder node The output state vector; For subsequent nodes The capability requirement constraint vector; To normalize the distance, before calculation, both the output state vector and the capability requirement constraint vector are mapped to the same 768-dimensional isomorphic feature space through a shared linear transformation layer; the normalized Euclidean distance uses the maximum Euclidean distance between all vector pairs in the candidate node set as the global normalization benchmark divisor. Calculated by subtracting the normalized Euclidean distance from 1, its value lies within a closed interval between 0 and 1, and the nodes... With nodes If the compatibility score between nodes reaches 0.92 (greater than or equal to 0.85), the system sets the weight of the corresponding directed dependent edge to 1. With nodes Establish directed dependency edges between tasks, and generate a directed acyclic graph of tasks after traversing the set of task nodes. .

[0045] The dependency verification unit will generate a directed acyclic graph of tasks. The task nodes are input one by one into the retrieval module, which then uses a retrieval enhancement generation method to recall standard case nodes. Calculate the deviation score between the task node and the standard case node: ,in, Scoring based on deviation; The first weighting factor for deviation is set to 0.6; Node to be verified With standard case nodes Cosine similarity of eigenvectors between them; The second weighting coefficient for deviation is set to 0.4. Node to be verified With standard case nodes The formula for calculating the mean of the absolute values ​​of the differences among the components of the five-dimensional constraint matrix is ​​as follows: ,in, Node to be verified The dimensional constraint vector components, Standard case node The dimensional constraint vector components, subscript For constrained dimension indexing, the initial node Data excluding the coordinates of explosion-proof restricted areas results in an increased deviation value for the environmental conditions dimension, leading to a higher calculated deviation score. If the value reaches 0.38, which is greater than or equal to 0.30, the system will update the node in the current state. Mark the node as an anomalous node and trigger the state machine rollback mechanism to calculate the deviation score. Regarding the five-dimensional constraint matrix The partial derivatives of each dimension are calculated, and the values ​​of each partial derivative are compared one by one. The environmental constraint dimension corresponding to the partial derivative with the largest value is selected. The dimension with the highest sensitivity was identified, and a penalty factor of 0.15 was applied. After applying gradient adjustment to the boundary of this dimension and calculating the boundary shrinkage, the updated environmental constraint boundary is as follows: Boundary constraints based on the environmental conditions before adjustment and the environmental condition constraint boundary reduced by the penalty factor As two boundaries within a specific gradient range, the adjusted boundary converges within this range; subsequently, the explosion-proof restricted area coordinate segment is written as a hard exclusion term into the intention decomposition unit to generate alternative nodes containing avoidance coordinates. To replace abnormal nodes The recalculated deviation score dropped to 0.08.

[0046] From the verified task node Extract tool requirement vectors from input / output parameters, interface protocols, and time response requirements. And retrieve the candidate tool capability representation vectors corresponding to each candidate tool in the pre-set candidate tool library. The normalized value of the inner product between the two vectors is used as the matching score: ,in, To determine the matching score, when the matching score is greater than or equal to 0.70 and less than 0.90, the pipeline reconstruction operator is activated. This connects the sensor acquisition tool and the data transmission tool in series among the candidate tools, forming a composite pipeline through the output interface of the sensor acquisition tool and the input interface of the data transmission tool. This composite pipeline is then bound to the task node. Finally, the schema mapper will bind the directed acyclic graph of tasks following the toolchain. The projection is converted into a standardized schema description file containing node identifiers, prerequisites, input and output parameters, and tool calls, and then output.

[0047] Example 2: This example selects 1000 sets of multi-source unstructured demand texts, including industrial inspection, supply chain gap diagnosis, and park operation and maintenance, and inputs them into a big data analysis test platform composed of computing power nodes and data storage nodes. To simulate the transmission environment of the industrial site, Gaussian noise with a signal-to-noise ratio of 18dB and interference harmonics used to simulate transmission delay fluctuations are superimposed at the signal input end. The original input features, key intermediate calculation data and final output results of each group are collected during the processing.

[0048] In this embodiment, the compatibility score threshold is set to 0.85. This threshold is used to balance the accuracy of establishing directed dependency edges with the disconnection rate of task nodes. When cross-validating based on 1000 historical industry cases, a threshold below 0.85 will increase the erroneous dependency establishment rate by more than 15%, while a threshold above 0.92 will cause the disconnection rate of normal nodes to exceed 10%. The deviation score judgment threshold is set to 0.30, which is used to trigger state machine backtracking when the constraint deviation of a task node reaches the judgment threshold. (Penalty factor) Set to 0.15; when When the value is less than 0.15, the number of convergence iterations for the constraint boundary increases. When the score is greater than 0.15, the constraint boundary shrinks excessively and compliant nodes are eliminated. A score of 0.15 can make the boundary converge stably within 3 to 5 iterations. The cascade interval of the matching score is set to be greater than or equal to 0.70 and less than 0.90. When the matching score is less than 0.70, the candidate tool cannot cover the tool requirements of the task node. When the matching score is greater than or equal to 0.90, a single tool is directly bound. The two weight coefficients in the compatibility score are set to 0.6 and 0.4, respectively. Their values ​​are determined according to the proportion of the discrete variance contribution rate of cosine similarity and normalized distance in representing semantic consistency.

[0049] Under the same testing environment, the sample of this invention adopts the complete technical solution. After inputting the requirement text with a signal-to-noise ratio of 18dB, the large model embedding layer extracts the original requirement semantic tensor, and the semantic slicing operator decouples it into a business intent vector and an environmental context vector, and then according to the five-dimensional constraint matrix... Calculate the compatibility score between task nodes. The group The initial deviation score is 0.92, which is greater than 0.85. Based on this, the system establishes directed dependency edges, and the retrieval module calculates the initial deviation score. If the value is 0.38, which is greater than 0.30, a penalty factor of 0.15 is applied after the state machine rolls back. The amount of boundary shrinkage is obtained by shrinking the boundary of the environmental constraint dimension. The value is 0.057; after regenerating the replacement node, The score dropped to 0.08, and then the matching score was calculated. The value is 0.78, which is within the range of greater than or equal to 0.70 and less than 0.90. The system connects the two candidate tools into a composite pipeline and binds them to the task node. The logical conflict rate of the directed acyclic graph of the task group is 0.2%, and the success rate of the standardized schema description file is 99.8%.

[0050] The control group 1 does not construct a five-dimensional constraint matrix. Under an input condition of 18dB, this group does not produce a compatibility score. and boundary contraction Deviation score The matching score is 0.48. The value was 0.52. Due to the lack of boundary constraints of the five-dimensional constraint matrix for task nodes, semantic drift occurred in the generated results. The logical conflict rate of the directed acyclic graph of the task was 28.5%, and the success rate of the standardized schema description file was 71.2%.

[0051] Compared to control group two, which retains the five-dimensional constraint matrix and task node compatibility calculation but removes the deviation feedback and boundary shrinkage of the retrieval module, the compatibility score of this group is [missing information - likely a specific score]. The deviation score is 0.88. The value was 0.38, indicating no boundary contraction. Matching score The error rate was 0.76. Since the abnormal nodes were not corrected after the deviation reached the judgment limit, the logical conflict rate of the directed acyclic graph of the task was 14.6%, and the success rate of the standardized schema description file was 85.1%.

[0052] To test the impact of the penalty factor value, control group three will... Set to 0.03, the compatibility score for this group under an input condition of 18dB. The deviation score is 0.91. The boundary contraction is 0.31. The match score is 0.011. The deviation score was 0.77. Due to the small boundary contraction, the deviation score was still higher than 0.30. The logical conflict rate of the directed acyclic graph was 8.3%, and the success rate of distributing the standardized schema description file was 91.4%. (The last sentence appears to be incomplete and possibly refers to a different sample group.) Set to 0.45; under the same 18dB input conditions, the compatibility score is... The deviation score is 0.90. The boundary contraction is 0.22. The match score is 0.171. The coefficient of variation is 0.75. The large boundary shrinkage eliminates some compliant task nodes and generates new dependency breaks, resulting in a task directed acyclic graph logical conflict rate of 5.7% and a standardized schema description file distribution success rate of 93.8%.

[0053] To test the impact of the compatibility score threshold, control group five had the threshold set to 0.65. Under an input condition of 18 dB, the compatibility score of this group was... The deviation score is 0.68. The boundary contraction is 0.12. The match score is 0.045. The compatibility score was 0.78. After the combination of nodes with low compatibility met the edge construction conditions, weakly dependent edges were formed, resulting in a 19.2% logical conflict rate in the directed acyclic graph of the task. The success rate of distributing the standardized schema description file was 80.5%. The control group six set this threshold to 0.95. Under an input condition of 18dB, the compatibility score was... The deviation score is 0.96. The boundary contraction is 0.09. The matching score is 0.042. The threshold was 0.79, which was relatively high, preventing some normal dependencies from being established. The logical conflict rate of the directed acyclic graph of tasks was 11.4%, and the success rate of distributing the standardized schema description file was 88.2%.

[0054] Example 3: The big data analytics platform receives the inspection request text from the smart factory, inputs the anonymized text obtained after data cleaning into the embedding layer of the large model, and maps it into a 768-dimensional original requirement semantic tensor. The semantic slicing operator uses a pre-defined feature projection matrix to... Slice the data and extract the feature components from dimensions 1 to 384 into a business intent vector. The feature components from dimension 385 to 768 are extracted as environmental context vectors. The feature projection matrix is ​​determined through joint pre-training of principal component analysis and autoencoder. During offline training, orthogonal covariance matrix decomposition is performed on corpus samples containing different scenarios to extract mutually orthogonal technical action basis features and environmental constraint basis features, and the original demand semantic tensor is obtained. After input, the semantic slicing operator first multiplies it with a linear transformation matrix, projects it onto a feature space composed of two sets of orthogonal bases, and then performs channel segmentation according to the technical meaning corresponding to each feature dimension to obtain the business intent vector. and environment context vector .

[0055] The multilayer perceptron decoder receives the service intent vector respectively. and environment context vector Its output layer is configured with a classification head and a regression head. The classification head outputs the task target classification code, inspection equipment hardware dependency label, and explosion-proof level gate control code based on the multi-label classification probability. The regression head outputs the environmental temperature limit, coordinate range, communication latency limit, system accuracy index, time response window, and throughput threshold through a fully connected neural network. Combined with the inspection domain knowledge graph, it integrates the business intent vector. The action core and target attributes are converted into scalar classification codes to form a task target constraint vector. From the context vector The environmental temperature range, explosion-proof area coordinates, and 100ms network communication latency upper limit are extracted to form an environmental condition constraint vector. From the business intent vector The system's accuracy metrics, the time response window characterized by a 7200s sampling period, and the throughput threshold are used to form a technical performance constraint vector. The hardware dependency labels of the inspection equipment are used to generate the capability requirement constraint vector. The Ex-d-II-C-T4 explosion-proof gating code is used to generate safety boundary constraint vectors. The five column vectors are concatenated column by column to form a five-dimensional constraint matrix of size 768×5. .

[0056] In the retrieval enhancement generation module verification task, a directed acyclic graph is used. At that time, the task node is calculated by relying on the verification unit. With standard case nodes Deviation score between The calculation result is 0.38, which is greater than 0.30. Therefore, the system will assign the task node... Marking a node as an anomalous and triggering state machine rollback, the system applies a five-dimensional constraint matrix to determine the highest sensitivity dimension. Each of the five constraint vectors is subjected to a step perturbation of 0.01. Record number The difference in deviation score of each constraint dimension before and after perturbation subscript The constraint dimension index is set to 1 to 5. Each variation difference is divided by the step perturbation to obtain the partial derivative of the deviation score with respect to the corresponding constraint dimension. The values ​​of the five partial derivatives are compared, and the column containing the environmental condition constraint vector corresponding to the partial derivative with the largest value is determined as the highest sensitivity dimension.

[0057] The control module retrieves the environmental condition constraint boundaries stored in the constraint register, with a penalty factor of 0.15. The corresponding partial derivatives are scaled, the boundary adjustment ratio is calculated, and the environmental constraint boundary is updated within the gradient shrinkage range defined by the penalty factor. The updated environmental constraint data is then rewritten into the input buffer of the intention decomposition unit, which uses this data to generate alternative nodes that avoid the coordinates of the explosion-proof restricted area. Used to replace abnormal nodes After recalculation, the deviation score is... To reduce the value to 0.08, for discrete engineering features such as explosion-proof rating and spatial coordinate red line, the system first converts the discrete classification code and coordinate interval into a continuous probability density feature vector, then applies a step perturbation in the normalized continuous feature space, and calculates the partial derivative based on the difference in deviation score before and after the perturbation.

[0058] For the task nodes that have passed the verification System retrieval tool requirement vector And the candidate tool capability representation vectors of each candidate tool in the candidate tool library. Calculate the normalized inner product value to obtain the matching score. The matching score of the candidate sensing tool is 0.78, which falls within the range of 0.70 or higher and less than 0.90. The pipeline reconstruction operator selects two candidate tools from the candidate tool library: the output data structure of the first candidate tool and the task node. The input requirements partially overlap, and the input data structure of the second candidate tool matches the output data structure of the first candidate tool. The pipeline reconstruction operator constructs a data type adapter layer between the output interface of the first candidate tool and the input interface of the second candidate tool, connecting the two candidate tools in series into a composite cascaded pipeline.

[0059] When constructing the data type adapter layer, the pipeline refactoring operator parses the output data architecture of the first candidate tool and the input data architecture of the second candidate tool, extracting the differences between them in field names, data types, and data units. If the data types are inconsistent, a data format conversion module is inserted between the two interfaces; if the data units are inconsistent, a pre-defined unit conversion coefficient matrix is ​​called for conversion; if the data throughput rates are inconsistent, a buffer queue is established between the two interfaces. After completing the interface adaptation, the interface description information of the composite cascaded pipeline is written to the task node. The attribute fields of the schema mapper will contain a directed acyclic graph of tasks with composite toolchain information. The projection is converted into a standardized schema description file containing node identifiers, prerequisite dependencies, input and output parameters, and tool calls, and then sent to the downstream control system.

[0060] Example 4: This example combines Figures 1 to 2 This paper explains the standardized generation method for intelligent analysis of multi-industry scenario requirements based on large models, such as... Figure 1 As shown, in the execution flow of the standardized generation method for intelligent parsing of multi-industry scenario requirements in large-scale models, step S1 is first executed, converting the requirement text sequence into a semantic tensor through an embedding layer, and then decoupling it into a business intent vector and an environmental context vector using a slicing operator; then step S2 is executed, extracting the parameter boundaries of the two vectors and constructing a five-dimensional constraint matrix containing task objectives, environmental conditions, technical indicators, capability requirements, and safety boundary constraint vectors; next, step S3 is executed, splitting the business intent vector into a set of task nodes, and ensuring that the compatibility score between the output state vector of the preceding node and the capability requirement constraint vector of the following node is greater than 1. When the deviation score is greater than or equal to 0.85, a directed dependency edge is established, generating a directed acyclic graph of the task. Next, step S4 is executed, where the task node is input into the retrieval module to recall standard case nodes. When the deviation score is greater than or equal to 0.30, the partial derivatives of each dimension are calculated to lock the highest sensitivity dimension. The boundary shrinkage is calculated using a penalty factor of 0.15 and the directed acyclic graph of the task is updated. Finally, step S5 is executed, where the matching score between the tool requirement vector of the task node and the capability representation vector of the candidate tool is calculated. When the matching score is in the range [0.70, 0.90), the two candidate tools are connected in series as a composite pipeline and bound to the task node, and a standardized description file is output.

[0061] like Figure 2As shown, the standardized generation architecture for intelligent parsing of multi-industry scenario requirements in a large model includes an input parsing domain, a topology orchestration domain, an interface binding domain, a conditional constraint domain, an adaptive error correction domain, and an output final state. The input parsing domain includes a preprocessing mechanism and semantic decoupling. The preprocessing mechanism performs text masking and hash desensitization, while semantic decoupling separates the intent vector using a slicing operator. The topology orchestration domain includes compatibility verification and topology graph generation. Compatibility verification calculates cosine similarity, and topology graph generation establishes a directed acyclic graph based on the scores. The interface binding domain includes tool matching and a composite pipeline. Tool matching calculates the matching scores between requirements and capabilities, and the composite pipeline connects tools and binds task nodes. The conditional constraint domain includes constraint construction and parameter extraction. Constraint construction combines to generate a five-dimensional constraint matrix, and the parameter extraction architecture extracts the boundaries of business and environmental parameters. The adaptive error correction domain includes boundary contraction and sensitivity analysis. Boundary contraction updates the topology using a penalty factor, and sensitivity analysis calculates partial derivatives to lock sensitive dimensions. The output final state connects to the interface binding domain and transforms the output into a standardized description file.

[0062] Example 5: Before the big data analytics system is integrated into a new industry application environment, the platform's data interface retrieves 5,000 sets of offline historical demand records from the storage node, generates the original demand semantic tensor through the large model embedding layer, and extracts the benchmark feature vectors of standard case nodes in various fields. With reference five-dimensional constraint matrix Based on the characteristics of historical nodes, the covariance matrix between different feature dimensions is calculated, and the dispersion of changes in each feature dimension is sorted and normalized; the first weight coefficient of deviation is determined by the top 60% of feature dimensions in the sorting. Its value is set to 0.6, and the second weighting coefficient of deviation is determined by the normalized mean of the dispersion of the other feature dimensions. Its value is set to 0.4, and the platform is set to a sliding sampling window with a time span of 30 days. When the feature vector of the newly recalled standard case node is compared with the baseline feature vector... When the mean residual is greater than 0.15, the baseline feature vector is triggered. The online weighted replacement is performed, and the replaced benchmark data is written into the knowledge base to update the benchmark of the standard case node.

[0063] Before the standardized schema description file is distributed to the downstream control system, the system tool interface module verifies the physical port mapping of each candidate tool in the pre-built candidate tool library, extracts the tool response latency, data throughput, and interface protocol fields, and converts these fields into a vector corresponding to the tool requirements of the task nodes. Same-dimensional candidate tool capability representation vector Calculate the normalized value of the inner product between the two vectors to obtain the matching score. When the matching score is greater than or equal to 0.70 and less than 0.90, select two candidate tools and read their input and output parameter types. Configure a data format conversion parser between the output interface of the first candidate tool and the input interface of the second candidate tool. Connect the two candidate tools in series as a composite cascaded pipeline. After completing the pipeline configuration, verify the connectivity of the data stream transmission. When the connectivity delay is less than 10ms and the packet loss rate is 0, bind the composite cascaded pipeline to the task node. Update the directed acyclic graph of tasks. The toolchain of this node outputs a standardized schema description file in JSONSchema format and injects it into the downstream execution engine.

[0064] Example 6: After the downstream automation management engine reports a task failure signal, the log parsing module of the data analysis platform retrieves the runtime log of the corresponding failed node, extracts the runtime response latency, memory usage, and error code, and performs standardized numerical transformation on each data item to obtain the runtime log vector. Run log vector and five-dimensional constraint matrix The discrete partial derivatives between the two are calculated based on a pre-constructed fault causal correlation matrix; the log parsing module first uses this matrix to analyze the running log vectors. Projected onto a five-dimensional constrained space, then calculated using a central finite difference method. Compared to Discrete partial derivative vectors of each column vector subscript Choose from 1 to 5, which correspond to the 5 constraint dimensions respectively.

[0065] The analysis module compares the magnitudes of each discrete partial derivative vector and determines the column corresponding to the maximum magnitude as the target constraint dimension. The constraint boundary for this dimension is retrieved from the industry knowledge graph, and a penalty factor of 0.15 is applied. The shrinkage correction is calculated by multiplying the product of the corresponding partial derivative magnitude and the shrinkage correction. This shrinkage correction is then subtracted from the original constraint boundary to obtain the converged new boundary data. Subsequently, the new boundary data is written into the prior distributed storage area of ​​the industry knowledge graph to update the initial parameters of the five-dimensional constraint matrix.

[0066] When network communication transmission delay exceeds 100ms or environmental parameters are missing from the original requirement text, the data acquisition interface triggers an abnormal fault tolerance protection mechanism. The data processing module stops calling the external knowledge base, retrieves the preset baseline five-dimensional constraint matrix from the storage node, and fills the missing features with the corresponding constraint components; the topology building unit reduces the compatibility score. The weights are calculated by establishing directed dependency edges based on the hard matching results of the output data type of the preceding node and the input data type of the following node. This matching method replaces probability calculation, and the task's directed acyclic graph is reconstructed. The schema mapper converts the task topology containing the safety precautions action into a degraded, standardized schema description file and sends the file to the terminal control unit.

[0067] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A large model-based multi-industry scenario demand intelligent analysis task standardization generation method, characterized by, Includes the following steps: Step S1: The sequence of requirement texts is converted into a semantic tensor through an embedding layer, and then decoupled into a business intent vector and an environmental context vector using a slicing operator. Step S2: Extract the parameter boundaries of the two vectors and construct a five-dimensional constraint matrix containing the task objectives, environmental conditions, technical indicators, capability requirements, and safety boundary constraint vectors. Step S3: Decompose the business intent vector into a set of task nodes. When the compatibility score between the output state vector of the preceding node and the capability requirement constraint vector of the following node is greater than or equal to 0.85, establish a directed dependency edge and generate a directed acyclic graph of the task. Step S4: Input the task node into the retrieval module to recall the standard case node, calculate the deviation score between the two, and when the deviation score is greater than or equal to 0.30, solve the partial derivatives of each dimension to lock the highest sensitivity dimension, use a penalty factor of 0.15 to calculate the shrinkage amount of the boundary of this dimension and update the task directed acyclic graph. Step S5: Calculate the matching score between the tool requirement vector of the task node and the capability representation vector of the candidate tools in the pre-set candidate tool library. When the matching score is greater than or equal to 0.70 and less than 0.90, connect the two candidate tools into a composite pipeline and bind it to the task node. Convert the task directed acyclic graph into a standardized description file output.

2. The method according to claim 1, wherein, The steps in step S2 to construct the five-dimensional constraint matrix include: Step S21, extracting data containing the ambient temperature range, spatial coordinate red line, and network latency upper limit from the environmental context vector to establish an environmental condition constraint vector; Step S22, extracting data containing system accuracy indicators, time response window, and throughput threshold from the business intent vector to establish a technical indicator constraint vector; Step S23, combining the environmental condition constraint vector, technical indicator constraint vector, task objective constraint vector, capability requirement constraint vector, and security boundary constraint vector into a multi-dimensional tensor to obtain the five-dimensional constraint matrix.

3. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 1, is characterized in that... The steps for calculating the compatibility score in step S3 include: step S31, extracting the output state vector of the preceding node and the capability requirement constraint vector of the following node; step S32, calculating the cosine similarity and normalized distance between the output state vector and the capability requirement constraint vector; and step S33, weighted summing of the cosine similarity and normalized distance to obtain the compatibility score.

4. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 1, is characterized in that... The step of generating the directed acyclic graph of the task in step S3 also includes: step S34, traversing the pairwise node combinations in the task node set, setting the weight of the corresponding directed dependent edge to 0 when the compatibility score is less than 0.85, and setting the weight of the corresponding directed dependent edge to 1 when the compatibility score is greater than or equal to 0.

85.

5. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 1, is characterized in that... In step S4, when calculating the deviation score between the two, the following formula is used: ,in, Scoring based on deviation; It is set as the first weighting coefficient and its value is set to 0.6; The cosine similarity of the feature vectors between the task node and the standard case node; It is the second weighting coefficient and its value is set to 0.4; This is the average absolute value of the differences between the task node and the standard case node in each component of the five-dimensional constraint matrix.

6. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 5, is characterized in that... The steps in step S4 to update the directed acyclic graph of the task include: Step S41, when the deviation score is greater than or equal to 0.30, marking the task node in the current state as an abnormal node; Step S42, calculating the partial derivatives of the deviation score with respect to each dimension of the five-dimensional constraint matrix, and determining the dimension corresponding to the largest partial derivative as the highest sensitivity dimension; Step S43, using a penalty factor of 0.15 to perform gradient adjustment on the boundary of the highest sensitivity dimension, converging the boundary of the highest sensitivity dimension within a specific gradient range that shrinks according to the penalty factor, and calling the intention decomposition unit to generate alternative nodes to replace the abnormal nodes.

7. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 1, is characterized in that... The steps for calculating the matching score in step S5 include: step S51, extracting the tool requirement vector from the task node and retrieving the capability representation vector of each candidate tool in the candidate tool library; step S52, calculating the normalized value of the inner product between the tool requirement vector and the capability representation vector, and using the calculated normalized value of the inner product as the matching score.

8. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 7, is characterized in that... The method of binding to the task node in step S5 also includes: step S53, when the matching score is greater than or equal to 0.90, the corresponding tool interface information is directly written into the attribute field of the task node to complete the direct binding.

9. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 1, is characterized in that... Following step S5, a closed-loop adaptive fine-tuning step is also included: Step S6, when a task failure signal is received from the downstream automated management engine, the running log vector of the corresponding failed node is extracted; the partial derivatives of the running log vector with respect to each dimension of the five-dimensional constraint matrix are calculated to locate the target constraint dimension; The shrinkage correction amount is calculated for the constraint boundary of the target constraint dimension and updated to the industry knowledge graph.

10. The standardized generation method for intelligent parsing of multi-industry scenario requirements based on a large model, as described in claim 1, is characterized in that... Before step S1, the following preprocessing steps are included: Step S0, data cleaning is performed on the input original requirement text sequence, and plaintext identity data and specific entity labels in the original requirement text sequence are removed by using text masking and high-dimensional hash desensitization mechanism to obtain the requirement text sequence.

Images