Seamless steel pipe continuous rolling size abnormality attribution method and device based on large language model

Abstract

The application provides a seamless steel pipe continuous rolling size abnormality attribution method and device based on a large language model, and relates to the technical field of quality analysis and intelligent diagnosis. The method comprises the following steps: structurally processing the knowledge of the continuous rolling process to obtain a knowledge unit library; collecting multi-source process data and size quality detection data, constructing a unified time reference sequence for time sequence alignment and length coordinate mapping to obtain a unified data reference; analyzing size quality indexes, determining a size quality abnormality distribution interval, and generating abnormal feature descriptions; constructing an abnormality propagation graph, reversely searching to obtain an abnormality propagation chain, and determining a root cause candidate variable set; calculating root cause scores for the root cause candidate variables and sorting to obtain a structured root cause analysis result; obtaining a constraint evidence set according to the structured root cause analysis result and the knowledge unit library, and inputting the constraint evidence set into a large language model to obtain a mechanism consistency verification conclusion. The application improves the accuracy, interpretability and reviewability of abnormality tracing and root cause analysis.

Description

Technical Field

[0001] This invention relates to the field of quality analysis and intelligent diagnosis technology in the continuous rolling production process of steel pipes, and in particular to a method and device for attributing dimensional anomalies in the continuous rolling of seamless steel pipes based on a large language model. Background Technology

[0002] As seamless steel pipe products develop towards precision and high-end applications, the dimensional quality requirements for wall thickness, outer diameter, and ovality are continuously increasing. The continuous rolling mill process involves multiple stands and strong variable coupling, leading to dimensional quality anomalies often being fragmented along the length of the pipe and potentially propagating across multiple stands, increasing the difficulty of anomaly tracing and root cause localization. Existing analytical methods typically suffer from challenges such as difficulty aligning multi-source process data and dimensional inspection data under the same benchmark, difficulty accurately mapping anomaly evidence intervals, difficulty in structurally characterizing cross-stand propagation relationships, and difficulty in quantitatively ranking root cause candidates. Furthermore, the fragmented nature of process specifications, equipment mechanisms, and historical cases, lacking a unified retrieval and constraint mechanism, makes it difficult to ensure that anomaly interpretation and handling recommendations are based on traceable evidence and that conclusions are auditable, thus affecting the timeliness and consistency of anomaly handling. Summary of the Invention

[0003] To address the technical problems in existing technologies, such as the difficulty in aligning multi-source process data and dimensional quality inspection data under the same benchmark, the difficulty in accurately locating dimensional quality anomalies, the difficulty in structurally representing cross-stand propagation relationships, the difficulty in quantitatively ranking root cause candidates, and the lack of traceability and auditability of anomaly interpretation and handling suggestions, this invention provides a method and apparatus for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model. The technical solution is as follows:

[0004] On the one hand, a method for attributing dimensional anomalies in continuous rolling mills of seamless steel pipes based on a large language model is provided. This method is implemented by a device for attributing dimensional anomalies in continuous rolling mills of seamless steel pipes based on a large language model. The method includes: S1. Obtain knowledge of process specifications, equipment structure and operation mechanism, and historical dimensional quality anomaly cases in the continuous rolling process of seamless steel pipes, and perform structured processing to obtain a knowledge unit library.

[0005] S2. For multiple stands in the continuous rolling process of seamless steel pipe, collect multi-source process data and dimensional quality inspection data during operation and preprocess them. Construct a unified time reference sequence to align the preprocessed data in time. Map the time-aligned data to length coordinates to obtain a unified data reference.

[0006] S3. Analyze the dimensional quality indicators in the dimensional quality inspection data in the unified data benchmark, determine the dimensional quality anomaly distribution range, and generate anomaly feature descriptions.

[0007] S4. Construct an anomaly propagation graph based on unified data benchmarks and anomaly feature descriptions. Perform a reverse search on the anomaly propagation graph for the distribution intervals of size and quality anomalies to obtain anomaly propagation chains and determine the set of root cause candidate variables corresponding to the anomaly propagation chains.

[0008] S5. Calculate the abnormal significance score, temporal consistency score, and propagation contribution score for any root cause candidate variable in the root cause candidate variable set, and then calculate and rank the root cause scores to obtain the structured root cause analysis results.

[0009] S6. Based on the results of the structured root cause analysis and the knowledge unit base, obtain the set of constraint evidence. Input the results of the structured root cause analysis and the set of constraint evidence into the large language model to obtain the mechanism consistency verification conclusion.

[0010] On the other hand, a device for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model is provided. This device is applied to the method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model. The device includes: The knowledge unit building unit is used to acquire knowledge of process specifications, equipment structure and operation mechanism, and historical dimensional quality anomaly cases in the continuous rolling process of seamless steel pipes, and to perform structured processing to obtain the knowledge unit library.

[0011] The data preprocessing and alignment unit is used to collect multi-source process data and dimensional quality inspection data from multiple stands in the continuous rolling process of seamless steel pipes, preprocess the data, construct a unified time reference sequence, perform time-series alignment on the preprocessed data, and perform length coordinate mapping on the time-series aligned data to obtain a unified data reference.

[0012] The anomaly identification unit is used to analyze the dimensional quality indicators in the dimensional quality inspection data in the unified data benchmark, determine the dimensional quality anomaly distribution range, and generate anomaly feature descriptions.

[0013] The anomaly propagation graph construction and tracing unit is used to construct an anomaly propagation graph based on a unified data benchmark and anomaly feature description, perform a reverse search on the anomaly propagation graph for the distribution interval of size and quality anomalies to obtain the anomaly propagation chain, and determine the root cause candidate variable set corresponding to the anomaly propagation chain.

[0014] The root cause quantitative assessment and ranking unit is used to calculate the abnormal significance score, temporal consistency score, and propagation contribution score for any root cause candidate variable in the root cause candidate variable set, and then calculate and rank the root cause scores to obtain the structured root cause analysis results.

[0015] The semantic verification and decision support unit is used to obtain a set of constraint evidence based on the results of structured root cause analysis and the knowledge unit base. The results of structured root cause analysis and the set of constraint evidence are then input into the large language model to obtain the mechanism consistency verification conclusion.

[0016] On the other hand, a device for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model is provided. The device includes: a processor; a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, any one of the above-described methods for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model is implemented.

[0017] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described methods for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model.

[0018] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: In this invention, a unified data benchmark is formed by aligning the time sequence between racks and mapping the steel pipe length coordinates, so that the dimensional quality anomaly segments and process variable evidence are aligned in the same coordinate system, thereby improving the accuracy of anomaly interval location and evidence extraction. By constructing an anomaly propagation graph and performing cross-rack reverse tracing, an interpretable anomaly propagation chain and a set of root cause candidate variables are obtained, enhancing the systematicness and reproducibility of the tracing process; By using a fusion scoring mechanism that combines anomaly significance, temporal consistency, and propagation contribution, quantitative assessment and ranking of root cause candidate variables are achieved, thereby improving the stability and objectivity of root cause localization. By retrieving relevant knowledge units from the knowledge unit base as constraint evidence and requiring the output to include citation identification information, the verification and recommendations of the large language model are made traceable and auditable. By limiting the fields and content scope of the handling suggestions through controlled output templates and prohibiting the output of specific settings that exceed the constraints of process specifications, it is beneficial to the engineering implementation, safety and compliance of abnormal handling suggestions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a flowchart of a seamless steel pipe continuous rolling dimensional anomaly attribution method based on a large language model provided by an embodiment of the present invention; Figure 2 This is a flowchart of the overall process of a method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model, provided in an embodiment of the present invention. Figure 3 This is a flowchart of a method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model, provided by an embodiment of the present invention. It shows the steps of knowledge unit base construction, multi-source data preprocessing and alignment, dimensional quality anomaly identification, anomaly propagation graph construction and reverse tracing, root cause quantitative assessment and ranking, and consistency verification of retrieval enhancement mechanism and output of controlled treatment suggestions. Figure 4 This is a block diagram of a seamless steel pipe continuous rolling dimensional anomaly attribution device based on a large language model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a seamless steel pipe continuous rolling dimensional anomaly attribution device based on a large language model, provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0022] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0023] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0024] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0025] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0026] This invention provides a method for attributing dimensional anomalies in continuous rolling mills of seamless steel pipes based on a large language model. This method can be implemented using a device for attributing dimensional anomalies in continuous rolling mills of seamless steel pipes based on a large language model. This device can be an edge computing node of the continuous rolling production line, a factory server, or a cloud computing platform, capable of interacting with the production process data acquisition system, quality inspection system, and historical knowledge base. Figure 1 The flowchart shown is for the attribution method of dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model. The processing flow of this method may include the following steps:

[0027] S1. Obtain knowledge of process specifications, equipment structure and operation mechanism, and historical dimensional quality anomaly cases in the continuous rolling process of seamless steel pipes, and perform structured processing to obtain a knowledge unit library.

[0028] Optionally, step S1 above may include the following steps S11-S15: S11. Acquire knowledge of the process specifications, equipment structure and operating mechanism, and historical cases of dimensional and quality anomalies in the continuous rolling process of seamless steel pipes.

[0029] In one feasible implementation, the continuous rolling process of this invention is multi-stand continuous rolling, for example, six-stand continuous rolling, with the stands sequentially labeled F1 to F6. The continuous rolling process specification text includes at least: parameter definitions, allowable ranges, logical constraints, linkage relationships, applicable specifications, and handling principles for the corresponding continuous rolling stands. Equipment structure and operating mechanism data includes at least: stands (rolls, bases, etc.), main drive system, measurement system and their component structural relationships, observable state quantities, and mechanistic influence paths. Historical dimensional quality anomaly case records include at least: anomaly type, anomaly morphology description, the steel pipe length range where the anomaly occurred, associated stands, associated variables, and handling records. The above three types of source data retain a "source field" (e.g., specification version, equipment number, case number) upon entry into the database to support subsequent review and verification.

[0030] S12. Break down the process specification knowledge into process rule units, and associate each process rule unit with a rack identifier and parameter identifier.

[0031] In one feasible implementation, the continuous rolling process specification text is broken down into process rule units at the "stand-parameter-constraint" granularity, and each process rule unit is associated with at least the following fields: stand identifier, parameter identifier, constraint type, constraint content, applicable conditions, and original text evidence fragment. Table 1 provides some example data for process rule units. The constraint content is stored in a structured format as a "parameter-comparison operator-threshold / interval-condition" quadruple for automatic comparison during subsequent consistency checks.

[0032] Table 1 Examples of Process Rule Units S13. Decompose the knowledge of equipment structure and operation mechanism into equipment mechanism units, and associate each equipment mechanism unit with equipment component identifiers and observable state quantity identifiers.

[0033] In one feasible implementation, the equipment structure and operating mechanism data are broken down into equipment mechanism units at a granularity of "component-state quantity-mechanism-impact," and each equipment mechanism unit is associated with at least the following fields: equipment component identifier, observable state quantity identifier, mechanism description, type of dimensional quality index that may be affected, and suggested verification items. Table 2 provides some example data for equipment mechanism units.

[0034] Table 2 Examples of Equipment Mechanism Units S14. Decompose the historical dimensional quality anomaly case knowledge into anomaly pattern units, and associate anomaly type identifier and anomaly morphology description identifier with each anomaly pattern unit.

[0035] In one feasible implementation, historical dimensional quality anomaly case records are divided into anomaly pattern units at a granularity of "anomaly type—anomaly morphology—evidence interval characteristics—associated racks—handling experience," and each anomaly pattern unit is associated with at least the following fields: anomaly type identifier, anomaly morphology description identifier, evidence interval characteristic summary, common associated racks, and handling experience summary. See Table 3.

[0036] Table 3 Examples of Abnormal Mode Units S15. Establish retrieval index fields for process rule units, equipment mechanism units, and abnormal mode units to form a knowledge unit library that can be retrieved and called by large language models, so as to provide traceable evidence sources for mechanism consistency verification and controlled treatment suggestions in subsequent steps.

[0037] In one feasible implementation, process rule units, equipment mechanism units, and abnormal mode units are uniformly incorporated into a knowledge unit library, and a retrieval index field is established for each knowledge unit, so that the large language model can be recalled based on fields such as "rack identifier, parameter identifier, abnormal type identifier, and keywords".

[0038] Furthermore, to improve semantic recall capabilities, vector representations can be constructed for the content of knowledge units and query conditions respectively, and cosine similarity can be used for similarity scoring. The top-L knowledge units are then selected as the recall results based on the scoring results. (1) In the formula, For knowledge units Similarity score with query criteria For any knowledge unit, This is a query vector representation generated from the search criteria. For knowledge units The vector representation of , For L2 norm operations, This is for the transpose operation.

[0039] Furthermore, the selection of the recall set based on similarity scores can be expressed as: (2) In the formula, For knowledge unit base, To retrieve the retrieved set of knowledge units, The preset recall quantity. To sort by similarity score from highest to lowest and select the top... The operator for each element. Table 4 provides an example (excerpt) of the recall results for a single retrieval, illustrating that the recall results carry identification information and can serve as reference evidence for subsequent consistency verification.

[0040] Table 4. Examples of Knowledge Unit Retrieval Recall (Excerpt) S2. For multiple stands in the continuous rolling process of seamless steel pipe, collect multi-source process data and dimensional quality inspection data during operation and preprocess them. Construct a unified time reference sequence to align the preprocessed data in time. Map the time-aligned data to length coordinates to obtain a unified data reference.

[0041] In one feasible implementation, multi-source process data and dimensional quality inspection data generated during the operation of seamless steel pipes in six stands (F1 to F6) of the continuous rolling process are collected. The multi-source data is preprocessed, and then a temporal alignment relationship between stands and a steel pipe length coordinate mapping are established based on the steel pipe's passage and tracking relationship in each stand to obtain a unified data benchmark to support subsequent identification of abnormal segments and tracing of propagation between stands. The unified data benchmark includes at least a multi-source process variable sequence and a dimensional quality index sequence aligned under the same steel pipe length coordinate system.

[0042] Optionally, step S2 above may include the following steps S21-S25: S21. For multiple stands in the continuous rolling process of seamless steel pipe, collect multi-source process data and dimensional quality inspection data during operation. Perform outlier removal, noise reduction and standardization on the multi-source process data and dimensional quality inspection data according to the stand identification to obtain the first intermediate data set.

[0043] In one feasible implementation, the process data for the six racks are processed separately according to rack identification. (The rest of the text appears to be a fragmented and incomplete sentence, possibly due to OCR errors.) The first Taking a process variable as an example, outlier removal is performed on its original sampled sequence:

[0044] (3) In the formula, This is the rack identifier, with values ​​1, 2, 3, 4, 5, and 6, corresponding to F1 through F6 respectively. For process variable indexing, For the first Each original sampling time, For rack process variables At any moment The original sampled values, and They are respectively Mean and standard deviation within the statistical window.

[0045] After outlier removal, the process variables are normalized to obtain a normalized sequence of process variables: (4) In the formula, These are the normalized process variable values. and After anomaly removal The minimum and maximum values.

[0046] Furthermore, denoising can be achieved using methods such as moving average, low-pass filtering, or median filtering. To avoid unnecessary limitations on the scope of the claims, this embodiment does not limit the specific denoising algorithm, but only requires that the denoised signal meets the stationarity requirements for subsequent alignment and correlation analysis.

[0047] At the same time, the same anomaly removal and standardization process is performed on the dimensional quality inspection data (such as wall thickness, outer diameter, and ellipticity) to obtain the first intermediate data set.

[0048] S22. Construct a unified time reference sequence in the time dimension, and perform resampling and interpolation alignment on the first intermediate data set according to the unified time reference sequence to obtain the second intermediate data set.

[0049] In one feasible implementation, since the sampling frequencies of different measurement points on the six racks may differ, this embodiment constructs a unified time reference sequence and performs resampling and interpolation alignment on all process variables: (5) In the formula, For the first in the unified time base sequence At that moment, To establish the starting time of a unified time base sequence, For the same time step.

[0050] For any process variable ,exist Linear interpolation is used to obtain the alignment value: (6) In the formula, These are the aligned sequence values ​​after resampling and interpolation. and The envelope in the original sampled sequence The adjacent moments.

[0051] Similarly, resampling and interpolation are performed on the dimensional quality inspection data sequence to bring it to a unified time reference with the process variable sequence. The data is then compared to obtain the second intermediate data set.

[0052] S23. Based on the seamless steel pipe tracking relationship, determine the passage range of the seamless steel pipe in each frame, and calculate the length coordinate increment of the seamless steel pipe between adjacent moments in the unified time reference sequence based on the length measurement signal synchronously collected with the last frame.

[0053] In one feasible implementation, the passage of the steel pipe through the frame is determined based on the "bite-out" event or equivalent tracking signal of the steel pipe in the six-stand assembly. The effective time period. To avoid introducing additional structural components that limit the passage, this embodiment determines the passage interval based on the event time recorded by the rack inlet / outlet detection signal or the control system:

[0054] (7) In the formula, For steel pipes at all times Is it in the rack? The amount of information passing through the interval, and They are respectively steel pipe biting frame With the ejection frame At that moment.

[0055] Furthermore, in this embodiment, the length measurement signal acquired synchronously with the main drive is used to determine the length coordinate increment of the steel pipe. The length measurement signal is extracted and the length coordinate increment is calculated within the passage interval of the steel pipe through the last frame. The length measurement signal is realized by cumulative pulse counting using an encoder. (8) In the formula, To unify the length coordinate increments between adjacent time points of a time base, For a moment The final frame encoder accumulates pulse counts. This is the pulse-length conversion factor.

[0056] S24. Accumulate the increments of the seamless steel pipe length coordinates to obtain the seamless steel pipe length coordinates corresponding to the unified time reference sequence. Introduce the frame elongation coefficient to recalibrate the seamless steel pipe length coordinates to obtain the recalibrated seamless steel pipe length coordinates.

[0057] In one feasible implementation, the length coordinate increments are accumulated, and the length coordinate increments outside the specified intervals are set to zero or not included in the accumulation, thus obtaining the steel pipe length coordinates that correspond one-to-one with a unified time reference. (9) In the formula, For a moment The corresponding coordinates of the steel pipe length.

[0058] Considering the elongation effect along the stand direction during six-stand continuous rolling, this embodiment introduces a stand elongation coefficient to recalibrate the length coordinate, so that data from different stands can be uniformly projected onto the same length coordinate reference: (10) In the formula, To arrange the length coordinates according to the frame The length coordinates after recalibration due to the elongation effect For rack The elongation factor can be obtained from specification parameters, process settings, and historical calibration.

[0059] S25. Based on the recalibrated seamless steel pipe length coordinates, a unified length coordinate grid is constructed. Linear interpolation is used to map the second intermediate data set to the unified length coordinate grid, so that the data of different frames are projected onto the same steel pipe length coordinate reference, and a unified data reference is obtained.

[0060] In one feasible implementation, to facilitate subsequent anomaly evidence interval localization and inter-rack propagation analysis, this embodiment constructs a uniform length coordinate grid and projects the aligned process variables and detection indicators onto this grid: (11) In the formula, To unify the length of the coordinate grid in the first Each length position To ensure a uniform length step size.

[0061] For any Linear interpolation is used to map process variables from the time domain / irregular length points to a uniform length grid: (12) In the formula, For rack process variables In uniform length coordinates Mapping value on, , They are respectively with , The corresponding recalibrated length coordinates.

[0062] Similarly, dimensional quality inspection data is mapped to a grid of uniform length. The above yields a sequence of dimensional quality indicators under a unified length coordinate system. This establishes a unified data benchmark for the frame process variables—dimensional quality indicators—that can be compared within the same length coordinate system. See Table 5 (Note: In the table...) , , This is an example of a process variable with a uniform length mapping across different racks; the wall thickness / outer diameter column maps to the same... The detection index values. In actual implementation, this can be expanded to include more process variables and more quality indicators.

[0063] Table 5. Examples of data under a uniform length coordinate system S3. Analyze the dimensional quality indicators in the dimensional quality inspection data in the unified data benchmark, determine the dimensional quality anomaly distribution range, and generate anomaly feature descriptions.

[0064] In one feasible implementation, based on dimensional quality inspection data in a unified data benchmark, at least one dimensional quality index such as wall thickness, outer diameter, and ellipticity is analyzed to identify dimensional quality anomalies and generate anomaly feature descriptions. The anomaly feature descriptions include at least the anomaly type, anomaly degree, and the distribution range of the anomaly on the steel pipe length coordinate, in order to support subsequent tracing of anomaly propagation between racks and screening of root cause candidate variables.

[0065] Optionally, step S3 above may include the following steps S31-S34: S31. Select any dimension quality index from the dimension quality inspection data in the unified data benchmark, obtain the sequence of the selected dimension quality index on the unified length coordinate grid, obtain the length index set corresponding to the reference interval of the selected dimension quality index, and calculate the mean and standard deviation of the reference baseline or reference distribution based on the sequence and the length index set.

[0066] In one feasible implementation, a reference baseline is established for the sequence of quality indicators of any size formed on a uniform length coordinate system. Taking the first... Taking a single dimension quality indicator as an example, it is based on a uniform length coordinate system. The sequence on is denoted as:

[0067] (13) Select the set of length indices corresponding to the reference interval, denoted as And calculate the mean and standard deviation of the reference baseline: (14) In the formula, Dimensional quality indicators The set of length indices corresponding to the reference interval. For set The number of elements, The mean of the reference baseline.

[0068] (15) In the formula, This is the standard deviation of the reference baseline.

[0069] Note: The reference interval can be determined by historical stable samples, qualified segments of the same specification, or stable segments determined online; this embodiment does not limit the specific method of obtaining the reference interval, but only requires that the reference baseline be used for subsequent deviation calculation.

[0070] S32. Based on the mean and standard deviation of the reference baseline, determine the start and end positions of candidate abnormal segments using sliding window statistical tests or change point detection, and obtain the size quality abnormality distribution range based on the start and end positions and the preset minimum duration threshold.

[0071] In one feasible implementation, this embodiment uses standardized deviation as the basis for anomaly determination, calculated as follows: (16) In the formula, Dimensional quality indicators In length position The standardized deviation.

[0072] Furthermore, set anomaly detection threshold. And construct an anomaly indicator: (17) In the formula, Dimensional quality indicators The anomaly detection threshold For length position Whether the indicator meets the abnormal conditions. This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise.

[0073] Furthermore, on Merge consecutive intervals with a value of 1 to obtain the start and end indices of candidate anomaly segments. and To suppress false alarms caused by isolated noise points, a minimum duration threshold is introduced. And only retain candidate segments that meet the duration requirement:

[0074] (18) In the formula, and These are the start and end indices of the candidate abnormal segments. This is the minimum duration threshold. Based on this, the abnormal distribution interval is determined as follows: .

[0075] Furthermore, another implementation may use change point detection instead of sliding window statistical testing to determine the start and end positions of segments, but this embodiment does not limit the specific change point detection algorithm.

[0076] S33. Calculate the deviation magnitude and duration of the abnormal distribution interval relative to the reference baseline to obtain the degree of abnormality and the length of the abnormal segment; extract the deviation direction characteristics of the abnormal distribution interval to determine the abnormality type; calculate the number of zero crossovers and the local rate of change within the abnormal distribution interval to determine the abnormality morphology.

[0077] In one feasible implementation, the degree of anomaly comprehensively reflects the deviation intensity of the anomalous segment. For identified anomalous segments... Calculate the anomaly index:

[0078] (19) In the formula, Dimensional quality indicators The degree of abnormality corresponding to the abnormal segment. The length of the abnormal segment is the number of points, satisfying the condition. .

[0079] Furthermore, to explicitly represent the impact of duration on severity, the length of the anomalous segment can be calculated: (20) In the formula, The length of the abnormal segment.

[0080] S34. Generate an anomaly feature description based on the size quality index identifier, anomaly degree, anomaly segment length, anomaly type, anomaly morphology, and anomaly distribution range.

[0081] In one feasible implementation, to determine the anomaly type, this embodiment extracts the deviation direction features of the anomaly fragment. First, the sign mean of the standardized deviation within the anomaly fragment is calculated:

[0082] (twenty one) In the formula, This is a measure of the overall deviation of the abnormal segment from the direction. When this happens, the exception type can be determined as "upper bias"; when In such cases, the anomaly type can be classified as "downward bias".

[0083] Furthermore, to distinguish between step, drift, and fluctuation patterns, this embodiment calculates the number of zero-crossings within abnormal segments: (twenty two) In the formula, The zero-crossing count of the standardized deviation within the outlier segment, when When the value exceeds a preset threshold, the abnormal pattern can be identified as "fluctuation type".

[0084] when When the value is not greater than a preset threshold, in order to distinguish between "step-type" and "drift-type" abnormal segments, this embodiment calculates the local rate of change of the abnormal segment (characterized by the mean of adjacent differences): (twenty three) In the formula, For measuring the local rate of change within anomaly segments, when When the value exceeds a preset threshold, the abnormal pattern can be classified as "step type"; when If the value is not greater than the preset threshold, the abnormal pattern can be identified as "drifting".

[0085] Based on this, an anomaly feature description is generated, which includes at least: anomaly indicator identifiers. Anomaly type (upward skew, downward skew), anomaly morphology (step, drift, fluctuation), anomaly degree Abnormal fragment length and abnormal distribution intervals See Table 6.

[0086] Table 6. Examples of Anomaly Identification Results Based on Wall Thickness Index S4. Construct an anomaly propagation graph based on unified data benchmarks and anomaly feature descriptions. Perform a reverse search on the anomaly propagation graph for the distribution intervals of size and quality anomalies to obtain anomaly propagation chains and determine the set of root cause candidate variables corresponding to the anomaly propagation chains.

[0087] In one feasible implementation, based on the abnormal feature description output in step S3 and the unified data benchmark formed in step S2, an anomaly propagation diagram between stands is constructed within the range of six continuous rolling stands (F1 to F6); and reverse tracing is performed on the dimensional quality anomaly segments on the anomaly propagation diagram to obtain an anomaly propagation chain, and the set of root cause candidate variables corresponding to the anomaly propagation chain is determined.

[0088] Optionally, step S4 above may include the following steps S41-S43: S41. The process parameters and equipment status of each stand in the continuous rolling mill are taken as process nodes, the dimensional quality indicators are taken as quality nodes, and a directed connection relationship from the previous stand to the next stand is established according to the sequential relationship of multiple stands to form the topology of the anomaly propagation graph.

[0089] In one feasible implementation, when constructing the anomaly propagation graph, the process variables of the six stands (such as rolling force, reduction, speed, torque, motor current, hydraulic reduction pressure, vibration-related state variables, etc.) are used as process nodes, and dimensional quality indicators such as wall thickness, outer diameter, and ellipticity are used as quality nodes. Furthermore, based on the stand sequence, only directed connections from the preceding stand to the following stand are established, forming the topology of the anomaly propagation graph. (twenty four) In the formula, This is an abnormal propagation diagram. A set of nodes, including at least process nodes and quality nodes. Let be a set of directed edges, with the edge direction satisfying the condition: previous rack → subsequent rack.

[0090] Furthermore, the node set can be further written as the union of the process node set and the quality node set: (25) In the formula, For the set of process nodes, This is the set of quality nodes.

[0091] Note: This embodiment does not limit the specific types of variables for process nodes, but requires that all process nodes originate from observable process data and equipment status data within the range of six continuous rolling mill stands.

[0092] S42. Within the data range corresponding to the abnormal distribution interval, perform time delay compensation alignment on the upstream and downstream nodes, and calculate their Pearson correlation coefficient. Use the absolute value of the correlation coefficient as the edge weight of the corresponding directed edge. When the edge weight is not less than the preset edge weight threshold, the directed edge is retained.

[0093] In one feasible implementation, the abnormal distribution interval given in step S3... The corresponding data range is truncated for edge weight calculation to avoid interference from non-abnormal intervals in assessing propagation intensity. The set of length indices corresponding to abnormal intervals is defined as follows:

[0094] (26) In the formula, Abnormal indicators are marked as The set of indexes for abnormal intervals at that time.

[0095] For any candidate directed edge (in This refers to the process node or intermediate node of the preceding rack. For nodes or quality nodes in subsequent racks, perform length-domain delay compensation alignment within the abnormal interval. Set the set of discrete compensation steps:

[0096] (27) In the formula, This is the set of discrete hysteresis steps used for alignment compensation. This represents the maximum number of compensation steps.

[0097] Lag steps In the index set Calculate the Pearson correlation coefficient over the length domain: (28) In the formula, For nodes With nodes With lag steps of Correlation coefficient at time For nodes In length position The sequence of values ​​on, For nodes In length position The sequence of values ​​on, for The mean on the index set, for In the index set The mean of the above.

[0098] Furthermore, the maximum absolute value of the correlation coefficient under different lag steps is taken as the edge weight of the candidate edge: (29) In the formula, For directed edges The right to the side.

[0099] Then set the edge weight threshold. The directed edge is retained when its weight is not less than the threshold: (30) S43. Starting from the quality node, search backwards along the anomaly propagation graph, select the propagation path with the largest cumulative edge weight as the anomaly propagation chain, and determine the variables corresponding to the process nodes other than the quality node on the anomaly propagation chain as the root cause candidate variable set.

[0100] In one feasible implementation, the quality node where the anomaly occurs is used as the starting point in the anomaly propagation graph. The search is performed in the reverse direction of the directed edge to construct several candidate propagation paths from the quality node to the preceding rack process node; and the path with the largest cumulative edge weight is selected as the anomaly propagation chain, using the cumulative edge weight of the path as the evaluation criterion.

[0101] For any candidate propagation path (Constituted by several directed edges), its cumulative edge weight is defined as: (31) In the formula, For the propagation path The cumulative edge weight score, For path A directed edge on, Let the weight of the directed edge be denoted as .

[0102] Based on this, the abnormal propagation chain is determined to be the propagation path with the highest cumulative edge weight score: (32) In the formula, For the finally determined abnormal propagation chain, This is the set of candidate propagation paths that can be reached from the anomalous quality node in reverse.

[0103] After obtaining the abnormal propagation chain Then, the variables corresponding to the process nodes (excluding quality nodes) in the anomaly propagation chain are determined as the root cause candidate variable set: (33) In the formula, For the set of root cause candidate variables, For nodes Located in the abnormal propagation chain superior, For nodes It does not belong to the set of quality nodes, and thus corresponds to process variables or equipment state variables.

[0104] Taking the wall thickness anomaly segment as an example: the anomaly distribution interval is [18.5, 27.0] m, and the quality node is selected as the starting point corresponding to the "wall thickness index"; the process nodes and quality nodes of the six racks (F1~F6) are constructed according to the rack order relationship, and the edge weights of candidate directed edges are calculated within the anomaly interval. The edge weight threshold is set as follows: (Example values), as shown in Tables 7 and 8.

[0105] Table 7. Examples of edge weight calculation and screening results for candidate edges within abnormal intervals. Table 8. Examples of anomaly propagation chains and root cause candidate variable sets obtained through reverse tracing. S5. Calculate the abnormal significance score, temporal consistency score, and propagation contribution score for any root cause candidate variable in the root cause candidate variable set, and then calculate and rank the root cause scores to obtain the structured root cause analysis results.

[0106] In one feasible implementation, the root cause candidate variable set is quantitatively evaluated, and the root cause score is calculated and ranked based on the anomaly significance, temporal consistency and propagation contribution to form a structured root cause analysis result, which supports the subsequent step S6 in the generation of mechanism consistency verification and controlled treatment recommendations based on the knowledge unit base; wherein, the structured root cause analysis result includes at least the root cause candidate variable identifier, the rack identifier, the root cause score and the evidence interval.

[0107] Optionally, step S5 above may include the following steps S51-S54: S51. For any root cause candidate variable in the root cause candidate variable set, obtain the sequence of the root cause candidate variable on a coordinate grid of uniform length, and calculate the reference mean and reference standard deviation of the root cause candidate variable in the reference interval. Calculate the standardized deviation of the root cause candidate variable based on the sequence, reference mean, and reference standard deviation, and calculate the abnormal significance score of the root cause candidate variable based on the standardized deviation.

[0108] In one feasible implementation, for any candidate root cause variable Take its sequence on a coordinate system of uniform length. Then, the reference mean and reference standard deviation are calculated over the reference interval to obtain the standardized deviation series of the variable: (34) In the formula, Candidate variables for root causes In length position The value of , and These are the root cause candidate variables. The mean and standard deviation over the reference interval. Candidate variables for root causes The standardized deviation.

[0109] Furthermore, anomaly significance scores are calculated within the abnormal distribution interval: (35) In the formula, Candidate variables for root causes Abnormal significance score, Set of indexes for abnormal distribution intervals The number of elements.

[0110] S52. Obtain the sequence of dimensional quality indicators on a uniform length coordinate grid, the mean and standard deviation of the reference baseline, calculate the standardized deviation of the dimensional quality indicators, perform lag-compensated correlation calculation on the standardized deviation of the root cause candidate variables based on the standardized deviation of the dimensional quality indicators, and obtain the temporal consistency score of the root cause candidate variables based on the calculation results.

[0111] In one feasible implementation, anomaly indicators are used. Standardized deviation sequence on a uniform length coordinate system For reference, for candidate variables Standardized deviation series Perform lag-compensated correlation calculations and use the maximum absolute correlation as the time-series consistency score: (36) In the formula, Candidate variables for root causes and abnormal indicators With lag steps of Correlation coefficient at time for exist The mean of the above, for exist The mean of the above.

[0112] Furthermore, the timing consistency score is defined as: (37) In the formula, Candidate variables for root causes Temporal consistency score, This is a preset set of lag compensation steps.

[0113] S53. Perform elimination or perturbation on the root cause candidate variables, and calculate the change in the explanatory power or anomaly score of the size quality anomaly to obtain the propagation contribution score of the root cause candidate variables.

[0114] In one feasible implementation, the propagation contribution is used to characterize the marginal contribution of candidate variables to the explanatory power of anomalies.

[0115] A local explanatory model for quality indicators is established within the abnormal interval, and the changes in abnormal scores are observed after perturbing the candidate variables.

[0116] First, define the abnormal indicators. Baseline abnormality score: (38) In the formula, Abnormal indicators The baseline anomaly score within the anomaly range.

[0117] Then, in the abnormal interval Built-in locally linear model explaining the outlier indicators using the set of root cause candidate variables: (39) In the formula, For the intercept term, Candidate variables Abnormal indicators The local explanatory coefficients, This is the residual term.

[0118] The above model is written in matrix form, and the coefficients are obtained by least squares estimation: (40) In the formula, The design matrix is ​​composed of the sequences of candidate variables within the anomaly interval. For the reason The response vector is composed of The estimated parameter vector contains With each .

[0119] After obtaining the local explanatory model, for any candidate variable Perform a perturbation and construct the perturbed sequence of candidate variables: (41) In the formula, For candidate variables The perturbated sequence.

[0120] Substituting the perturbed variables into the local explanatory model, we obtain the predicted quality sequence under the perturbation conditions. Based on this, the anomaly score after the perturbation is calculated: (42) In the formula, For candidate variables Anomaly scores obtained after perturbation Predicted sequence from perturbation The deviation obtained by S3 isomorphic standardization.

[0121] Therefore, the propagation contribution score is defined as the difference between the baseline anomaly score and the disturbance anomaly score: (43) In the formula, Candidate variables for root causes The score for contribution to the spread.

[0122] If the anomaly score decreases significantly after the candidate variable is perturbed, it indicates that the candidate variable has a stronger explanatory contribution to the formation and propagation of anomalies.

[0123] S54. The anomaly significance score, temporal consistency score, and propagation contribution score are fused to obtain the root cause score. The root cause scores are then ranked, and the structured root cause analysis results are output. The structured root cause analysis results include: root cause candidate variable identifiers, rack identifiers, root cause scores, and structured root cause analysis results of evidence intervals.

[0124] In one feasible implementation, to avoid the influence of units, this embodiment normalizes the three types of scores separately. Taking the anomaly significance score as an example, range normalization is used:

[0125] (44) In the formula, The score for the normalized significance of the anomalies. and Candidate sets The minimum and maximum values ​​of the abnormal significance score, To prevent positive constants with a denominator of zero.

[0126] Similarly, we can obtain and Then, a weighted fusion method was used to obtain the root cause score:

[0127] (45) In the formula, Candidate variables for root causes Root cause score , , The preset weighting coefficients satisfy the following conditions: .

[0128] Finally, for all according to Sort the data from largest to smallest and output the structured root cause analysis results. The structured root cause analysis results should include at least the following: root cause candidate variable identifier, rack identifier, root cause score, and evidence interval.

[0129] In this embodiment, the abnormal indicator identified in step S3 is selected as... Abnormal fragments and their abnormal distribution ranges As evidence intervals, and based on the root cause candidate variable set obtained in step S4. As input, within the evidence interval, calculate the abnormal significance score for each candidate variable. Time series consistency score Score for contribution to dissemination The root cause score is obtained by fusion in step S54. The results are listed in Table 9, which presents an example output format of the structured root cause analysis results.

[0130] Table 9. Examples of quantitative scoring and ranking of root cause candidate variables. S6. Based on the results of the structured root cause analysis and the knowledge unit base, obtain the set of constraint evidence. Input the results of the structured root cause analysis and the set of constraint evidence into the large language model to obtain the mechanism consistency verification conclusion.

[0131] In one feasible implementation, the structured root cause analysis results and relevant knowledge units retrieved from the knowledge unit base are fed into the large language model as input, and the model outputs the mechanism consistency verification conclusion and the constrained anomaly handling suggestions, and outputs the identification information of the referenced knowledge units for review and traceability; wherein, the anomaly handling suggestions are limited to at least one of equipment inspection direction, process adjustment ideas and risk warnings, and the output includes the identification information of the referenced knowledge units.

[0132] Optionally, step S6 above may include the following steps S61-S62: S61. Based on the results of the structured root cause analysis, recall knowledge units from the knowledge unit library that match the rack identifier, parameter identifier, and anomaly type identifier, and use the identifier information and content fragments of the recalled knowledge units as constraint evidence.

[0133] In one feasible implementation, a retrieval query is constructed based on the results of structured root cause analysis, and knowledge units that match rack identifiers, variable identifiers, anomaly indicator identifiers, and anomaly patterns are recalled from the knowledge unit base to form a set of constraint evidence.

[0134] Construct a query semantic vector and use it to calculate the relevance score of candidate knowledge units. The query semantic vector is defined as follows:

[0135] (46) In the formula, To query semantic vectors, For vector normalization operators, , , For preset weighting coefficients, The semantic vector is generated from the root cause candidate variable identifier and the rack identifier. A semantic vector generated from anomaly indicator identifiers and anomaly type information. It is a semantic vector generated from abnormal morphological information.

[0136] For any candidate knowledge unit Calculate the relevance score: (47) In the formula, Candidate knowledge units Relevance score, , To preset the fusion coefficient, This is the cosine similarity calculation function. For knowledge units semantic vectors, This is a bonus item based on metadata matching, used to characterize the matching status of rack identifiers, parameter identifiers, or anomaly identifiers.

[0137] Top-based selection based on relevance scores L Each knowledge unit serves as a set of constraint evidence: (48) S62. Input the structured root cause analysis results and constraint evidence into the large language model to generate a mechanism consistency verification conclusion; when the verification fails, output the conflict evidence and the data items to be verified.

[0138] In one feasible implementation, the results of structured root cause analysis are combined with a set of constraint evidence. A common input large language model is used to perform consistency verification of the execution mechanism; the verification includes at least: consistency of process specifications, rationality of equipment mechanism and experience matching of abnormal modes.

[0139] Each recalled constraint evidence is abstracted into a set of testable constraint predicates. And define the constraint violation indication: (49) In the formula, For the first The violation of the indicator of a constraint predicate. These are the constraint predicates extracted from the evidence unit. This is an indicator function.

[0140] Furthermore, the consistency conflict metric is defined as: (50) In the formula, For consistency conflict measurement, To constrain the number of predicates.

[0141] The validation fails when the conflict metric exceeds the threshold. (51) In the formula, This is a preset conflict threshold.

[0142] When the verification fails, this embodiment requires the output of: conflict evidence (at least including conflict evidence unit identifiers) and a set of data items to be verified. The set of data items to be verified can be defined as:

[0143] (52) In the formula, The set of data items to be verified. Top-ranked candidate variables for root causes M Subset.

[0144] S63. When the verification passes, an anomaly handling suggestion is generated based on the constraint evidence. The anomaly handling suggestion is generated using a controlled output template. The controlled output template must limit the output fields to include at least two of the following: equipment inspection object, process adjustment direction, and risk warning. It is also prohibited to output specific set values ​​that exceed the constraints of the process specification.

[0145] In one feasible implementation, when step S62 determines that the verification has passed, this embodiment is based on the constraint evidence set. Generate abnormal handling suggestions, which are limited to at least one of the following: equipment inspection direction, process adjustment approach, and risk warning. It is prohibited to output specific settings that exceed the constraints of process specifications.

[0146] Specifically, define the set of controlled output fields: (53) In the formula, For a controlled set of output fields, For the device inspection object field, For the process adjustment direction field, This is a risk warning field. This is used to reference the knowledge unit identifier field.

[0147] To suppress out-of-bounds output with specific set values, this embodiment performs numerical out-of-bounds detection on the output text of the large language model and defines a numerical trigger indicator: (54) In the formula, The indicator is triggered when the value exceeds the limit. The set of terms for the output text. For the set of numerical terms that are determined to be specific set values / specific adjustment amounts, when When the output does not meet the requirements of the controlled template, a rewrite or masking is triggered; when Output the final disposal suggestion in the field. The table lists the identifiers of the knowledge units referenced. Table 10 provides an example of the output of sample-level controlled treatment recommendations.

[0148] Table 10 Example of Sample-Level Controlled Disposal Recommendations Output like Figure 2 , Figure 3 As shown, this invention takes multi-source process data and dimensional quality inspection data as input, and outputs structured root cause analysis results, mechanism consistency verification conclusions, and constrained anomaly handling suggestions. Specifically, it first acquires and structures process specification knowledge, equipment structure and operating mechanism knowledge, and historical dimensional quality anomaly case knowledge to form a searchable and callable knowledge unit library; then, it preprocesses the multi-source process data and dimensional quality inspection data of the continuous rolling production process, and establishes the time sequence alignment relationship between stands and the steel pipe length coordinate mapping based on material tracking relationships to form a unified data benchmark; based on the unified data benchmark, it identifies dimensional quality anomaly segments and generates anomaly feature descriptions; it constructs an anomaly propagation graph and performs cross-stand reverse tracing to obtain the anomaly propagation chain and the corresponding root cause candidate variable set; it quantitatively evaluates and sorts the root cause candidate variables to form structured root cause analysis results; finally, it inputs the structured root cause analysis results and the retrieved relevant knowledge units as constraint evidence into a large language model, outputting mechanism consistency verification conclusions and constrained anomaly handling suggestions, and the output includes the identification information of the referenced knowledge units.

[0149] In this embodiment of the invention, a unified data benchmark is formed by aligning the time sequence between racks and mapping the steel pipe length coordinates, so that the dimensional quality anomaly segments and process variable evidence are aligned in the same coordinate system, thereby improving the accuracy of anomaly interval location and evidence extraction. By constructing an anomaly propagation graph and performing cross-rack reverse tracing, an interpretable anomaly propagation chain and a set of root cause candidate variables are obtained, enhancing the systematicness and reproducibility of the tracing process; By using a fusion scoring mechanism that combines anomaly significance, temporal consistency, and propagation contribution, quantitative assessment and ranking of root cause candidate variables are achieved, thereby improving the stability and objectivity of root cause localization. By retrieving relevant knowledge units from the knowledge unit base as constraint evidence and requiring the output to include citation identification information, the verification and recommendations of the large language model are made traceable and auditable. By limiting the fields and content scope of the handling suggestions through controlled output templates and prohibiting the output of specific settings that exceed the constraints of process specifications, it is beneficial to the engineering implementation, safety and compliance of abnormal handling suggestions.

[0150] Figure 4 This is a block diagram illustrating an attribution device for dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model, according to an exemplary embodiment. The device is used in a method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model. (Refer to...) Figure 4 The device includes a knowledge unit base construction unit 310, a data preprocessing and alignment unit 320, an anomaly identification unit 330, an anomaly propagation graph construction and tracing unit 340, a root cause quantitative assessment and ranking unit 350, and a semantic verification and decision support unit 360. Among them:

[0151] Knowledge unit 310 is used to acquire knowledge of process specifications, equipment structure and operation mechanism, and historical dimensional quality anomaly cases in the continuous rolling process of seamless steel pipes, and to perform structured processing to obtain the knowledge unit library.

[0152] The data preprocessing and alignment unit 320 is used to collect multi-source process data and dimensional quality inspection data from multiple stands in the continuous rolling process of seamless steel pipes, preprocess the data, construct a unified time reference sequence, perform time-series alignment on the preprocessed data, and perform length coordinate mapping on the time-series aligned data to obtain a unified data reference.

[0153] The anomaly identification unit 330 is used to analyze the dimensional quality indicators in the dimensional quality inspection data in the unified data benchmark, determine the dimensional quality anomaly distribution range, and generate anomaly feature descriptions.

[0154] The anomaly propagation graph construction and tracing unit 340 is used to construct an anomaly propagation graph based on a unified data benchmark and anomaly feature description, perform a reverse search on the size and quality anomaly distribution interval on the anomaly propagation graph to obtain the anomaly propagation chain, and determine the root cause candidate variable set corresponding to the anomaly propagation chain.

[0155] The root cause quantitative assessment and ranking unit 350 is used to calculate the abnormal significance score, temporal consistency score and propagation contribution score for any root cause candidate variable in the root cause candidate variable set, and then calculate the root cause score and rank them to obtain the structured root cause analysis results.

[0156] The semantic verification and decision support unit 360 is used to obtain a set of constraint evidence based on the results of structured root cause analysis and the knowledge unit base, and input the results of structured root cause analysis and the set of constraint evidence into the large language model to obtain the mechanism consistency verification conclusion.

[0157] Figure 5 This is a schematic diagram of a seamless steel pipe continuous rolling dimensional anomaly attribution device based on a large language model, provided in an embodiment of the present invention. Figure 5 As shown, the seamless steel pipe continuous rolling dimensional anomaly attribution device based on a large language model can include the above-mentioned... Figure 4 The illustrated device is a seamless steel pipe continuous rolling dimensional anomaly attribution apparatus based on a large language model. Optionally, the seamless steel pipe continuous rolling dimensional anomaly attribution apparatus 410 based on a large language model may include a first processor 2001.

[0158] Optionally, the seamless steel pipe continuous rolling dimensional anomaly attribution device 410 based on a large language model may also include a memory 2002 and a transceiver 2003.

[0159] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0160] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model, characterized in that, The method includes: S1. Acquire knowledge of process specifications, equipment structure and operation mechanism, and historical dimensional quality anomaly cases in the continuous rolling process of seamless steel pipes, and perform structured processing to obtain a knowledge unit library; S2. For multiple stands in the continuous rolling process of seamless steel pipe, collect multi-source process data and dimensional quality inspection data during operation and preprocess them. Construct a unified time reference sequence to align the preprocessed data in time. Map the time-aligned data to length coordinates to obtain a unified data reference. S3. Analyze the dimensional quality indicators in the dimensional quality inspection data in the unified data benchmark, determine the dimensional quality anomaly distribution range, and generate anomaly feature descriptions; S4. Construct an anomaly propagation graph based on unified data benchmarks and anomaly feature descriptions. Perform a reverse search on the anomaly propagation graph for the distribution intervals of size and quality anomalies to obtain anomaly propagation chains and determine the root cause candidate variable set corresponding to the anomaly propagation chains. S5. Calculate the abnormal significance score, temporal consistency score, and propagation contribution score for any root cause candidate variable in the root cause candidate variable set, and then calculate and rank the root cause scores to obtain the structured root cause analysis results. S6. Based on the results of the structured root cause analysis and the knowledge unit base, obtain the set of constraint evidence. Input the results of the structured root cause analysis and the set of constraint evidence into the large language model to obtain the mechanism consistency verification conclusion. The S5 includes: S51. For any root cause candidate variable in the root cause candidate variable set, obtain the sequence of the root cause candidate variable on a coordinate grid of uniform length, and calculate the reference mean and reference standard deviation of the root cause candidate variable in the reference interval. Calculate the standardized deviation of the root cause candidate variable based on the sequence, reference mean and reference standard deviation, and calculate the abnormal significance score of the root cause candidate variable based on the standardized deviation. S52. Obtain the sequence of dimensional quality indicators on a uniform length coordinate grid, the mean and standard deviation of the reference baseline, calculate the standardized deviation of the dimensional quality indicators, perform lag-compensated correlation calculation on the standardized deviation of the root cause candidate variables based on the standardized deviation of the dimensional quality indicators, and obtain the temporal consistency score of the root cause candidate variables based on the calculation results. S53. Perform elimination or perturbation on the root cause candidate variables, and calculate the change in the explanatory power or anomaly score of the size quality anomaly to obtain the propagation contribution score of the root cause candidate variables. S54. The anomaly significance score, temporal consistency score, and propagation contribution score are fused to obtain the root cause score. The root cause scores are then ranked, and the structured root cause analysis results are output. The structured root cause analysis results include: root cause candidate variable identifiers, rack identifiers, root cause scores, and structured root cause analysis results of evidence intervals.

2. The method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model according to claim 1, characterized in that, S1 includes: S11. Acquire knowledge of the process specifications, equipment structure and operating mechanism, and historical cases of dimensional quality anomalies in the continuous rolling process of seamless steel pipes; S12. Break down the process specification knowledge into process rule units, and associate each process rule unit with a rack identifier and parameter identifier; S13. Decompose the knowledge of equipment structure and operation mechanism into equipment mechanism units, and associate each equipment mechanism unit with equipment component identifiers and observable state quantity identifiers; S14. Decompose the historical dimensional quality anomaly case knowledge into anomaly pattern units, and associate anomaly type identifier and anomaly morphology description identifier with each anomaly pattern unit. S15. Establish retrieval index fields for process rule units, equipment mechanism units, and abnormal mode units to form a knowledge unit library.

3. The method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model according to claim 1, characterized in that, S2 includes: S21. For multiple stands in the continuous rolling process of seamless steel pipe, collect multi-source process data and dimensional quality inspection data during operation. Perform outlier removal, noise reduction and standardization on the multi-source process data and dimensional quality inspection data according to the stand identification to obtain the first intermediate data set. S22. Construct a unified time base sequence, and perform resampling and interpolation alignment on the first intermediate data set according to the unified time base sequence to obtain the second intermediate data set; S23. Based on the seamless steel pipe tracking relationship, determine the passage interval of the seamless steel pipe in each frame, extract the length measurement signal based on the passage interval, and calculate the seamless steel pipe length coordinate increment between adjacent moments in the unified time reference sequence. S24. Accumulate the increments of the seamless steel pipe length coordinates to obtain the seamless steel pipe length coordinates corresponding to the unified time reference sequence. Introduce the frame elongation coefficient to recalibrate the seamless steel pipe length coordinates to obtain the recalibrated seamless steel pipe length coordinates. S25. Based on the recalibrated seamless steel pipe length coordinates, a unified length coordinate grid is constructed. Linear interpolation is used to map the second intermediate data set to the unified length coordinate grid to obtain a unified data benchmark. The unified data benchmark includes multi-source process data and dimensional quality inspection data aligned under the same seamless steel pipe length coordinate system.

4. The method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model according to claim 1, characterized in that, The S3 includes: S31. Select any dimension quality index from the dimension quality inspection data in the unified data benchmark, obtain the sequence of the selected dimension quality index on the unified length coordinate grid, obtain the length index set corresponding to the reference interval of the selected dimension quality index, and calculate the mean and standard deviation of the reference baseline based on the sequence and the length index set. S32. Based on the mean and standard deviation of the reference baseline, determine the start and end indices of the candidate abnormal segments using sliding window statistical tests or change point detection, and obtain the size quality abnormality distribution range based on the start and end indices and the preset minimum duration threshold. S33. Calculate the deviation magnitude and duration of the abnormal distribution interval relative to the reference baseline to obtain the degree of abnormality and the length of the abnormal segment; extract the deviation direction characteristics of the abnormal distribution interval to determine the type of abnormality; calculate the number of zero crossovers and the local rate of change within the abnormal distribution interval to determine the morphology of the abnormality. S34. Generate an anomaly feature description based on the size quality index identifier, anomaly degree, anomaly segment length, anomaly type, anomaly morphology, and anomaly distribution range.

5. The method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model according to claim 1, characterized in that, The S4 includes: S41. Take the multi-source process data in the unified data benchmark as process nodes, take the dimensional quality inspection data as quality nodes, and establish a directed connection relationship from the preceding rack to the following rack according to the sequential relationship of multiple racks to form the topology of the anomaly propagation graph. S42. Define the length index set corresponding to the abnormal distribution interval, perform time delay compensation alignment on the upstream and downstream nodes, calculate the Pearson correlation coefficient between the upstream and downstream nodes on the length index set, use the Pearson correlation coefficient as the edge weight of the candidate edge, and retain the candidate edge with the edge weight not less than the preset edge weight threshold as the directed edge of the abnormal propagation graph. S43. Starting from the quality node where the anomaly occurred, search in the reverse direction of the directed edges on the anomaly propagation graph to obtain multiple candidate propagation paths from the quality node to the preceding rack process node. Select the candidate propagation path with the largest cumulative edge weight as the anomaly propagation chain, and determine the root cause candidate variable set based on the variables corresponding to the process nodes other than the quality node on the anomaly propagation chain.

6. The method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model according to claim 1, characterized in that, The S6 includes: S61. Construct a query semantic vector based on the results of the structured root cause analysis, calculate the relevance score between the query semantic vector and any candidate knowledge unit in the knowledge unit base, and select multiple knowledge units as a set of constraint evidence based on the relevance score. S62. Input the structured root cause analysis results and the set of constraint evidence into the large language model for mechanism consistency verification. When the mechanism consistency verification result is that the verification fails, output the conflict evidence and the set of data items to be verified. When the mechanism consistency verification result is that the verification passes, generate anomaly handling suggestions based on the set of constraint evidence.

7. A device for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model, the device being used to implement the method for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model as described in any one of claims 1-6, characterized in that, The device includes: The knowledge unit building unit is used to acquire knowledge of process specifications, equipment structure and operation mechanism, and historical dimensional quality anomaly cases in the continuous rolling process of seamless steel pipes, and to perform structured processing to obtain the knowledge unit library. The data preprocessing and alignment unit is used to collect multi-source process data and dimensional quality inspection data from multiple stands in the continuous rolling process of seamless steel pipes, preprocess the data, construct a unified time reference sequence, perform time-series alignment on the preprocessed data, and perform length coordinate mapping on the time-series aligned data to obtain a unified data reference. The anomaly identification unit is used to analyze the dimensional quality indicators in the dimensional quality inspection data in the unified data benchmark, determine the dimensional quality anomaly distribution range, and generate anomaly feature descriptions. The anomaly propagation graph construction and tracing unit is used to construct an anomaly propagation graph based on a unified data benchmark and anomaly feature description, perform a reverse search on the anomaly propagation graph for the distribution interval of size and quality anomalies to obtain the anomaly propagation chain, and determine the root cause candidate variable set corresponding to the anomaly propagation chain. The root cause quantitative assessment and ranking unit is used to calculate the abnormal significance score, temporal consistency score, and propagation contribution score for any root cause candidate variable in the root cause candidate variable set, and then calculate and rank the root cause scores to obtain the structured root cause analysis results. The semantic verification and decision support unit is used to obtain a set of constraint evidence based on the results of structured root cause analysis and the knowledge unit base. The results of structured root cause analysis and the set of constraint evidence are then input into the large language model to obtain the mechanism consistency verification conclusion.

8. A device for attributing dimensional anomalies in continuous rolling of seamless steel pipes based on a large language model, characterized in that, The seamless steel pipe continuous rolling dimensional anomaly attribution device based on a large language model includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 6.

Images