Production data analysis method and system applied to metal product processing

Abstract

This invention discloses a production data analysis method and system applied to metal product processing. The method first extracts independent processing operation fragments and performs multi-dimensional semantic decomposition on the process text to generate a processing operation semantic sequence. Second, it performs multi-scale morphology primitive decomposition on the surface micro-morphology data to establish a surface morphology evolution sequence. Then, it synchronously correlates the processing operation semantic sequence with the surface morphology evolution sequence in a time sequence, generating a bidirectional reciprocating causal relationship chain through positive action logic evaluation and reverse tracing logic evaluation, and combining them into an initial causal association structure. Next, it performs latent disturbance factor detection processing, detecting asymmetric deviation paths and extracting latent processing factors. Finally, based on the latent processing factors, it adjusts the multi-dimensional descriptive semantics in the processing operation semantic sequence to generate a process parameter adjustment instruction sequence, thereby uncovering latent processing factors and generating a process parameter adjustment strategy with causal explanatory power.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a method and system for analyzing production data in metal product processing. Background Technology

[0002] In metal product processing, the process specification defines the complete processing route of the part in the form of process text, covering elements such as operation actions, action areas, and process parameters. Surface micromorphology, as a direct representation of processing quality, is accurately recorded by height field data acquired at each acquisition time point. Currently, the parsing of process texts largely relies on keyword extraction, making it difficult to form a joint semantic representation of operation methods, action areas, and energy application forms. In terms of surface morphology analysis, obtaining morphological primitives at different levels through multi-scale geometric decomposition has become a common method. This allows for the extraction of morphological units and the labeling of their scale and area attributes, and the establishment of morphological state transition sequences through primitive matching between adjacent time points.

[0003] However, the shortcomings of existing technologies are that mainstream methods usually treat process text analysis and morphological evolution analysis as independent tasks, and there is a lack of strict causal logical relationship mechanism between the two. This leads to the fact that the cause of processing anomalies is highly dependent on human experience judgment, and it is difficult to discover unrecorded hidden processing factors from the data level, and it is also difficult to automatically generate process parameter adjustment strategies with causal explanation. Summary of the Invention

[0004] The purpose of this invention is to provide a production data analysis method and system for metal product processing, so as to solve the problems mentioned in the background art.

[0005] This invention provides a production data analysis method for metal product processing, comprising: Independent processing operation segments are extracted from the process text in the metal product process specification to perform multidimensional semantic decomposition, generating multidimensional descriptive semantics corresponding to each independent processing operation segment and connecting them in time sequence to form a processing operation semantic sequence; the multidimensional descriptive semantics include operation mode description semantics, action area description semantics, and energy action form description semantics. Multi-scale morphology primitive decomposition is performed on the surface micromorphology data acquired at each acquisition time node in the metal product processing production line to obtain a set of morphology primitives. Based on the acquisition time node, a morphology state transition sequence of the morphology primitive set is established as a surface morphology evolution sequence. Each independent processing operation segment in the processing operation semantic sequence is temporally synchronized with the morphology state transition in the corresponding time period in the surface morphology evolution sequence. A bidirectional reciprocating causal relationship chain is generated through positive action logic evaluation and reverse tracing logic evaluation. All bidirectional reciprocating causal relationship chains are combined into an initial causal relationship structure. The initial causal relationship structure is subjected to latent perturbation factor detection processing. Asymmetric deviation paths between the forward and reverse causal relationship chains are detected along the bidirectional reciprocating causal relationship chain. Latent processing factors not recorded in the semantic sequence of the processing operation are extracted from the asymmetric deviation paths. Based on the implicit processing factors, the multidimensional description semantics of the corresponding independent processing operation segments in the processing operation semantic sequence are adjusted to generate a process parameter adjustment instruction sequence.

[0006] This invention provides a production data analysis system for metal product processing, comprising: A processor; a storage device on which a computer program is stored; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the above-described production data analysis method applied to metal product processing.

[0007] The present invention provides a readable storage medium on which a program or instruction is stored, and when the program or instruction is executed by a processor, it implements the above-described production data analysis method applied to metal product processing.

[0008] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing two independent processing links—multidimensional semantic parsing of process text and multi-scale evolution analysis of surface micromorphology—and innovatively introducing a bidirectional reciprocating causal relationship chain as the core mechanism for cross-link association, the correlation between processing operations and morphological responses no longer remains at the temporal synchronization level, but delves into the level of causal logical mutual verification. By actively detecting the asymmetric deviation path between the positive effect strength and the negative support strength of the causal relationship chain, the systematic detection and extraction of implicit processing factors that are not recorded in the process specifications but substantially affect processing quality are achieved for the first time. Finally, driven by these implicit processing factors, the semantic descriptions of operation mode, action area, and energy action form in the semantic sequence of processing operations are directionally adjusted to generate a physically interpretable sequence of process parameter adjustment instructions. This elevates traditional post-processing quality analysis to a closed-loop process optimization based on causal inference, enhancing the interpretability of the metal product processing process and the initiative of process optimization. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1This is a flowchart illustrating a production data analysis method for metal product processing, provided as an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of a production data analysis method for metal product processing provided in an embodiment of this application.

[0012] Figure 3 This is a schematic diagram of the basic structure of a production data analysis system for metal product processing provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0014] Please see Figure 1 , Figure 1 This is a flowchart of a production data analysis method for metal product processing provided in an embodiment of this application. The method can be executed by a production data analysis system for metal product processing, or it can be jointly executed by a production data analysis system for metal product processing and a server. The method includes steps 110-150.

[0015] This invention aims to address the problem in existing technologies where process specifications only record explicit operations, while many implicit disturbances during processing are difficult to capture and record, leading to difficulties in tracing processing quality and a lack of physical basis for process optimization. This invention, through deep integration of semantic parsing of process text and multi-scale evolution analysis of surface microstructure, constructs a bidirectional causal relationship between processing operations and morphological responses. It actively detects unrecorded implicit processing drivers and ultimately generates physically interpretable process parameter adjustment instructions, thus providing a causal-driven process optimization method for high-precision metal product processing lines.

[0016] This invention uses an exemplary precision metal parts processing production line as the application scenario throughout. This production line includes a process specification database based on a computer-aided design model, multiple CNC machining centers equipped with online surface topography measurement devices, and a production data analysis system. The executing entity of this invention is the central computing server (i.e., the production data analysis system applied to metal parts processing), which interacts with the database, measurement devices, and the controllers of the machining centers via a network, implementing the various steps of this invention in the form of software programs.

[0017] Step 110: Extract independent processing operation segments from the process text in the metal product process specification to perform multidimensional semantic decomposition, generate multidimensional descriptive semantics corresponding to each independent processing operation segment, and connect them in time sequence to form a processing operation semantic sequence; the multidimensional descriptive semantics includes operation mode description semantics, action area description semantics, and energy action form description semantics.

[0018] The core of this step is to transform the process specifications recorded in natural language or semi-structured text into a time-ordered semantic representation sequence that computers can understand and compute. This process specification is a typical machining procedure for a valve body part on this production line, and its process text details each step from the raw material to the finished product. To achieve this transformation, step 110 contains several progressive processing sub-steps: Step 111: Perform syntactic analysis and semantic segmentation on the process text of the metal product process specification to obtain independent processing operation segments.

[0019] First, the production data analysis system loads a pre-trained process text syntactic dependency parsing model and a processing domain proper noun recognition model. The architecture of the syntactic dependency parsing model is a multi-layer bidirectional long short-term memory network, with its top layer connected to a dependency relation decoder with a dual affine attention mechanism. When the process text "rough milling of the reference surface followed by fine grinding of the step surface" is input into the model, the model first decomposes the text into a sequence of sub-word units using a byte-pair encoding word segmenter, and then extracts the contextual features of each sub-word unit through the bidirectional long short-term memory network layer.

[0020] The dependency decoder calculates a score on these features for the specific dependency relationship between each pair of word units. By identifying the verb-object dependency arc between the verb "rough milling" and the object "datum surface," and the verb-object dependency arc between "fine milling" and "step surface," and by detecting the coordinating conjunction "and" connecting two predicate structures, the model semantically segments the original text into two independent syntactic units.

[0021] Meanwhile, the processing domain-specific terminology recognition model uses a conditional random field as the top layer of sequence labeling. It receives the feature sequence output from a bidirectional long short-term memory network and performs BIO tag prediction on each sub-word unit, labeling "datum surface" and "step surface" as "area of ​​action" entities, and "rough milling" and "fine grinding" as "operation actions" entities. Finally, based on the identified core verbs and the entity groups they govern, the system segments the text into two independent processing operation segments: "rough milling datum surface" and "fine grinding step surface."

[0022] Step 112: Extract the operation action events of each independent processing operation segment and classify them to generate operation mode description semantics; determine the geometric boundary representation of the processing contact area based on the workpiece geometry description to generate action area description semantics; extract the process parameters in the independent processing operation segment and combine them into energy action form description semantics.

[0023] Taking the independent machining operation segment "rough milling of the reference surface" as an example, the operation action extractor maps the verb "rough milling" to a predefined machining action hierarchy. This hierarchy is a multi-branch tree structure, with the root node being "physical machining," which is subdivided into "material removal," "plastic deformation," etc., and "material removal" is further divided into "cutting," "grinding," etc. "Rough milling" is ultimately mapped to a standardized coding sequence of the complete hierarchy of "cutting - milling - end milling - rough machining." This hierarchical coding sequence is the semantic description of the operation mode.

[0024] Next, the geometric boundary resolver accesses the 3D model of the valve body part in the pre-machining stage. Based on the markings of the solid "datum plane" in the model, it extracts all the triangular faces that constitute the datum plane and calculates the connected convex hulls of the vertices of these triangular faces. The boundary of the convex hull is represented by a set of ordered point coordinate sequences, which is the semantic description of the action area of ​​this operation segment. Finally, the parameter combiner extracts all numerical process parameters from the original segment text using regular expressions, including spindle speed, feed rate, axial depth of cut, and radial width of cut. These parameters are grouped: spindle speed represents "rotational mechanical energy," and feed rate and depth of cut together represent "feed mechanical energy." In a multi-dimensional vector space, the combiner assigns the values ​​of the corresponding dimensions of each energy group as normalized values ​​of the parameters, forming a multi-dimensional energy action vector. This vector is the semantic description of the energy action form.

[0025] Step 113: Perform joint rationality judgment and conceptual superposition correction on the semantic description of the operation mode, the semantic description of the action area, and the semantic description of the energy action form. Based on the correction result, perform ternary semantic encapsulation to generate processing operation semantic nodes.

[0026] A rationality discriminator, composed of physical rules, receives the results of filling these three semantic slots. The rules embedded within the discriminator are expressed using first-order predicate logic. For example, one rule might be: "If the operation belongs to the 'surface modification-heat treatment' subclass, then the energy action form must include a 'thermal energy' component, and its value should be greater than the 'mechanical energy' component." The discriminator performs truth evaluation on these logical expressions. If all joint conditions are true, the discriminator passes the test.

[0027] If a conflict is detected, a higher-level concept correction is initiated. This correction process is guided by a knowledge graph that stores the hierarchical, equivalence, and substitution relationships between processing concepts. For example, if the operation method is "laser surface hardening" but the energy application form only contains "protective airflow" without "laser power," the inference engine traverses upwards along the knowledge graph to find the higher-level concept "auxiliary energy form" for "protective airflow" and the lower-level relationship for "laser power." Ultimately, the energy application form is corrected to include both a vector dimension representing "laser thermal energy" and a vector dimension representing "auxiliary energy form - airflow." After correction, the values ​​of the three semantic slots are encapsulated into a unified data object, namely the processing operation semantic node.

[0028] Step 114: Determine the sequential relationship between semantic nodes of each processing operation based on the statement order of the process text and assign a temporal connection interval attribute label to obtain a directed sequence chain.

[0029] The sequential relation parser reads the relative position of each processing operation semantic node in the original process text and assigns it a monotonically increasing node sequence index. Simultaneously, the parser analyzes the connecting text between the segments from which the nodes originate.

[0030] If the original text uses the phrase "immediately afterwards" between two segments, the parser queries the corresponding temporal attribute in a node relationship knowledge base. This knowledge base stores the mapping relationship between text forms such as "immediately afterwards" and the time intervals they represent. Based on this, the parser assigns the temporal connection interval attribute label "immediately following" to the connection.

[0031] If two segments are separated by a period and there is no time adverb in between, the parser matches them against an empirical time sequence rule base based on the standard processing rhythm of the production line and the types of preceding and following processes, confirming them as "regular intervals". Through this analysis, all semantic nodes of processing operations are connected by directed edges with semantic attributes, forming a directed sequence chain.

[0032] Step 115: The semantic descriptions of each processing operation semantic node are converted into vectors by semantic vector mapping and fused into node semantic representation vectors. The processing operation semantic sequence is generated by concatenating the nodes of the directed order chain.

[0033] For each processing operation semantic node, a semantic vector mapping module is invoked. This module contains three independent encoder sub-networks: a graph convolutional network for encoding the semantic description of the operation mode, which receives the hierarchical encoding sequence of the operation, aggregates the hierarchical structure information through multi-layer graph convolutional operations, and outputs a fixed-length operation mode semantic vector; a point cloud feature extraction network for encoding the semantic description of the action region, which receives the point coordinate sequence of the action region, extracts spatial features through a multi-layer perceptron, and aggregates them through a symmetric function, outputting a fixed-length action region semantic vector; and a fully connected network for encoding the semantic description of the energy action form, directly outputting a fixed-length energy action form semantic vector. These three vectors are then concatenated along the principal dimension to obtain a composite node semantic representation vector. Finally, the system extracts the node semantic representation vectors of each node sequentially according to the node sequence index of the directed sequential chain, arranging them into a time-series shaped data structure, namely the processing operation semantic sequence.

[0034] Step 120: Perform multi-scale morphology primitive decomposition on the surface micromorphology data acquired at each acquisition time node in the metal product processing production line to obtain a set of morphology primitives, and establish the morphology state transition sequence of the set of morphology primitives as the surface morphology evolution sequence based on the acquisition time node.

[0035] This step is performed in parallel with the text parsing in step 110. In the production line, whenever a machining operation is completed or a preset online inspection time node is reached, the surface topography measurement device located next to the machine tool scans a specific inspection area of ​​the valve body part to acquire its surface micro-topography data. The purpose of step 120 is to transform this continuous, high-dimensional surface topography data into a structured sequence of topography primitives and quantify its evolution process based on this sequence.

[0036] Step 121: Extract the surface micromorphology height field matrix corresponding to each acquisition time node from the surface micromorphology data acquired at each acquisition time node in the metal product processing production line.

[0037] The production data analysis system receives raw data streams from surface topography measurement equipment. For each data acquisition point, it parses the file header to determine the data dimensions, and then fills the height values ​​of the measurement points into a two-dimensional matrix according to their row and column coordinates. The row and column indices of the matrix correspond to the sampling point numbers of the Cartesian coordinate system on the workpiece inspection surface, respectively. The value of each element in the matrix corresponds to the vertical height of the sampling point relative to the reference plane.

[0038] Step 122: Perform multi-level topographic undulation layering analysis on the surface micro-topographic height field matrix to generate topographic undulation morphological primitives at different level scales. Extract independent contour closed boundaries for topographic undulation morphological primitives at each level scale to obtain topographic morphological units with closed perimeters. Label each topographic morphological unit with a level label and area descriptor.

[0039] First, a multi-level topographic undulation analysis module performs spatial frequency domain decomposition on the height field matrix. This module uses a set of Gaussian bandpass filters, each defined by a high-pass cutoff frequency and a low-pass cutoff frequency. The cutoff frequencies of different filter sets correspond to different spatial wavelength scales, from the smallest scale labeled S1 to the largest scale labeled Sk. The original height field matrix is ​​passed through this set of filters sequentially, resulting in k filter response matrices at different levels of scale. The connected high-response regions in each filter response matrix constitute the topographic undulation morphological primitives at that scale.

[0040] Next, for each scale's morphological undulation primitive, a contour closure boundary extraction algorithm is initiated. This algorithm first binarizes the response matrix, then employs a boundary tracking algorithm based on gradient magnitude and direction, traversing along the high gradient direction of the response region boundary until returning to the starting point, forming a closed contour. Each region defined by a closed contour is defined as a morphological unit. The number of pixels in this morphological unit is recorded as its area value, and a hierarchical label is assigned to it to indicate its scale level.

[0041] Step 123: Combine morphological units with spatial adjacency at the same level scale into morphological unit groups according to the degree of contact between adjacent boundaries, and assign a morphological group identifier and a spatial distribution density descriptor within the group to each morphological unit group.

[0042] Spatial clustering is performed within each scale level. The system calculates the shortest Euclidean distance between point sets on the closed contours of any two morphological units. If this distance is lower than a preset contact threshold related to that scale level, the two morphological units are considered spatially adjacent. Morphological units that have direct or indirect adjacency relationships are clustered into a morphological unit group. Each newly formed morphological unit group is assigned a globally unique morphological group identifier, which is a combination of the scale level number and an auto-incrementing group number within that scale. Simultaneously, the convex hull area of ​​the morphological unit group is calculated, and the sum of the areas of all morphological units within the group is divided by the convex hull area. The quotient obtained is the intra-group spatial distribution density descriptor of the group.

[0043] Step 124: Perform cross-level nesting relationship analysis on the morphological unit groups at all levels under each acquisition time node, establish the inclusion relationship structure of the fine-level morphological unit groups located inside the coarse-level morphological unit groups, and generate the multi-scale morphological primitive set for that acquisition time node.

[0044] Starting with each morphological unit group at the smallest scale level, the geometric centroid coordinates of its convex hull are calculated. Then, the system queries upwards to larger scale levels to find which morphological unit groups have convex hulls containing these centroid coordinates. If most key points (such as contour vertices) of a finer-scale morphological unit group fall within the convex hull of a coarser-scale morphological unit group, an inclusion edge is established from the coarser-scale group to the finer-scale group. This inclusion edge forms a cross-level directed acyclic graph. This graph, along with all morphological unit groups and their attributes at all levels, constitutes a complete multi-scale representation of the surface morphology at the time of acquisition—that is, a multi-scale morphological primitive set.

[0045] Step 125: Perform morphological unit group matching and tracing on the multi-scale morphological primitive sets of adjacent acquisition time nodes. For each morphological unit group located at the previous acquisition time node, find the corresponding morphological unit group in the subsequent acquisition time node whose morphological group identifier meets the matching condition and satisfies the requirement of spatial proximity, and establish temporally adjacent morphological unit group matching pairs.

[0046] The system uses each morphological unit group in the multi-scale morphological primitive set of the previous time point as the query object. The matching and tracing process first searches for morphological unit groups at the same scale level in the next time point based on the scale level number in the morphological group identifier. The matching condition is achieved by calculating the comprehensive similarity between two morphological unit groups, which is a weighted sum of the difference in group area descriptors and the difference in spatial distribution density within the group. Only when the comprehensive similarity is less than a preset matching threshold, and the offset of the geometric centroids of the two groups in spatial coordinates is less than a preset displacement threshold, is a morphological unit group matching pair considered successful.

[0047] Step 126: For each temporally adjacent morphological unit group matching pair, calculate the morphological group identifier change description, area descriptor change description, and intra-group spatial distribution density descriptor change description of the morphological unit group, and aggregate them into a morphological state transition description vector.

[0048] For each matching pair, the changes in its three attributes are calculated. The morphological group identifier change description is a Boolean change vector, where each dimension corresponds to a feature of a predefined morphological group category, generated by comparing the group features before and after. The area descriptor change description is represented by the ratio obtained by dividing the area descriptor of the later time point by the area descriptor of the previous time point. The intra-group spatial distribution density descriptor change description is represented by calculating the difference between the density descriptors before and after.

[0049] Finally, the Boolean value change vector, ratio value, and difference value are concatenated into a one-dimensional vector, which is the morphological state transition description vector corresponding to the matching pair.

[0050] Step 127: Arrange the topographic state transition description vectors of all time-adjacent matching pairs according to the time progression direction of the acquisition time node to obtain the topographic state transition vector sequence.

[0051] The system extracts the topographic state transition description vectors calculated for all matching pairs between each adjacent time node in step 126 in the order of the acquisition time nodes from first to last. After sorting them according to their adjacency relationship in spatial distribution, the system flattens and concatenates them into a time series, which is the topographic state transition vector sequence.

[0052] Step 128: Perform morphological transition pattern merging on the morphological state transition vector sequence, merge consecutive matching pairs with similar morphological state transition description vectors into the same morphological transition stage, assign a stage start timestamp and stage duration length attribute to each same morphological transition stage, and encapsulate the stage start timestamp and stage duration length attribute of different same morphological transition stages, as well as all morphological state transition description vectors within that stage, into a surface morphological evolution sequence.

[0053] The system performs segmented aggregation on the topographic state transition vector sequence, specifically employing a bottom-up segmentation method. Initially, each vector point in the sequence is an independent segment. Then, adjacent segments whose Euclidean distance falls within a preset low variance interval are iteratively merged. The iteration terminates when no more segments can be merged. Each segment obtained is defined as a topographic transition stage, with the start of the stage corresponding to the timestamp of the first vector in the segment, and the duration determined by multiplying the number of vector points in the segment by the sampling interval. These topographic transition stages with temporal attributes, along with their internal vector sequences, are then combined and output to form the surface topographic evolution sequence.

[0054] Step 130: Each independent processing operation segment in the processing operation semantic sequence is temporally synchronized with the morphology state transition of the corresponding time period in the surface morphology evolution sequence. A bidirectional reciprocating causal relationship chain is generated through positive action logic evaluation and reverse tracing logic evaluation. All bidirectional reciprocating causal relationship chains are combined into an initial causal association structure.

[0055] After obtaining the two time series, one semantic and one morphological, the core task of step 130 is to align the two time series in the time dimension and establish a strong causal relationship.

[0056] Step 131: Extract the stage start timestamp and stage duration attribute corresponding to each isomorphic transfer stage in the surface morphology evolution sequence; using the stage start timestamp and stage duration attribute as time interval scales, delineate the occupied interval of each isomorphic transfer stage on the time axis, and map the processing operation semantic nodes within the time axis coverage area to the occupied interval of the corresponding isomorphic transfer stage according to their temporal position, so as to perform time alignment mapping between processing operation semantic nodes and isomorphic transfer stages.

[0057] The production data analysis system plots the corresponding continuous time intervals for each homomorphic transfer stage on a unified timeline. Simultaneously, for each node in the processing operation semantic sequence, based on its order index in its directed sequence chain and assigned temporal connection attributes such as "immediately following" or "regular interval," combined with the standard processing cycle time of the production line, the system dynamically calculates the absolute timestamp interval of that node on the unified timeline. Then, an interval matching algorithm is executed to find homomorphic transfer stage intervals that intersect with the timestamp interval of each processing operation semantic node. If an intersection exists, a time alignment mapping relationship is established between the two.

[0058] Step 132: For each time-aligned and successfully mapped processing operation semantic node and the same shape transition stage pair, extract the operation mode description semantic of the processing operation semantic node in the pair as the antecedent descriptor, and extract the shape state transition description vector in the same shape transition stage as the consequence descriptor.

[0059] For each pair established in step 131, the system retrieves the value of the operation mode description semantic slot generated in step 112 from its processing operation semantic node as a potential cause. Simultaneously, it extracts the set of all encapsulated morphology state transition description vectors from the object in the same morphology transition stage as a potential result.

[0060] Step 133: Combining the antecedent descriptor and the consequence descriptor, initiate the positive action logic evaluation and the reverse tracing logic evaluation, and determine the initial causal relationship structure based on the positive causal relationship establishment identifier and the reverse tracing relationship establishment identifier.

[0061] This step is one of the core technologies of this invention, and aims to cross-verify the causal relationship from both positive and negative perspectives.

[0062] Step 1331: Initiate positive action logic evaluation, determine the ideal morphological state transition change of the processing action type pointed to by the processing method physical influence rule base, obtain a positive presumed morphological state transition change description, and logically compare the positive presumed morphological state transition change description with the consequence descriptor to generate a positive causal relationship establishment identifier.

[0063] The physical impact rule base for processing methods is a tensor-based expert knowledge base. The system converts the hierarchical encoding of operation methods in the antecedent descriptors into a query vector, which is then retrieved from the rule base. For each atomic-level processing operation and its combination, the rule base stores one or more "positive presumed morphological state transition tensors" representing the desired morphological change result. The tensor that is matched in the search is taken as a positive presumed morphological state transition change description. This tensor is then matched with the dot product of each morphological state transition description vector from the consequence descriptor set to obtain a set of similarity scores. If the highest similarity score falls into a preset high confidence interval, a positive causal relationship establishment identifier for a positive state is generated.

[0064] Step 1332: Initiate reverse tracing logic evaluation, and based on the morphological backtracking rule base, back-deduce the set of candidate processing action types that can produce the morphological state transition from the consequence descriptor, determine whether the antecedent descriptor falls into the set of candidate processing action types, and generate a reverse tracing relationship establishment identifier.

[0065] The shape backtracking rule base is a knowledge base indexed by shape change patterns. The system fuses and averages the shape state transition description vectors in the consequence descriptor set to obtain an average consequence description vector. Using this vector, the system retrieves K rules with similarity exceeding a threshold by calculating the cosine similarity with each rule index vector in the shape backtracking rule base. The set of processing action types associated with these retrieved rules constitutes the candidate processing action type set. Subsequently, it checks whether the operation mode hierarchy code corresponding to the antecedent descriptor belongs to this set. If it does, a reverse homing relationship establishment identifier for the affirmative state is generated.

[0066] Step 1333: When both the positive causal relationship establishment identifier and the reverse tracing relationship establishment identifier are in a positive state, establish a positive causal edge from the processing operation semantic node to the same shape transfer stage and a reverse causal edge from the same shape transfer stage to the processing operation semantic node for the pair, and pair the positive causal edge and the reverse causal edge into a bidirectional reciprocating causal relationship chain.

[0067] It is understandable that this step establishes a strong causal relationship between two data nodes only if both the forward and reverse logical evaluations pass. The system creates a causal edge data structure, where the forward causal edge stores the highest similarity score calculated in step 1331 as the positive effect strength attribute, and the reverse causal edge stores the average retrieval similarity calculated in step 1332 as the reverse support strength attribute. These two edges in opposite directions are paired into a whole, that is, a bidirectional reciprocating causal relationship chain.

[0068] Step 1334: Traverse all processing operation semantic nodes that have been successfully time-aligned and mapped, and pair them with the same shape transition stage, generating the corresponding bidirectional reciprocal causal relationship chain one by one, and generating the complete set of bidirectional reciprocal causal relationship chains.

[0069] The system uses a loop structure to traverse all pairs successfully mapped in step 131, and executes the processing logic from steps 132 to 1333 for each pair, generating a corresponding bidirectional reciprocal causal relationship chain. All links are added to a set without repetition, forming the complete set of bidirectional reciprocal causal relationship chains.

[0070] Step 1335: Collect multiple bidirectional reciprocal causal relationship chains with the same processing operation semantic node from the complete set of bidirectional reciprocal causal relationship chains. Within the collection unit, evaluate the heterogeneity of responses to the same processing operation semantic node at different morphological transition stages to generate a processing semantic divergence description. Then, collect multiple bidirectional reciprocal causal relationship chains with the same morphological transition stage from the complete set of bidirectional reciprocal causal relationship chains. Within the collection unit, evaluate the cumulative contribution of different processing operation semantic nodes to the same morphological transition stage to generate a morphological-driven aggregation description. Finally, associate and combine the complete set of bidirectional reciprocal causal relationship chains with the processing semantic divergence description and the morphological-driven aggregation description to obtain an initial causal association structure.

[0071] The system hashes and groups the entire chain set using semantic nodes of processing operations as keys. For each group, it analyzes the vector distribution of consequence descriptors for multiple identical transition stages they point to. By calculating the covariance matrix of all consequence descriptor vectors within the group and taking the trace of this matrix as a description of the processing semantic divergence, the uncertainty of the result caused by the same operation is quantified.

[0072] Next, the descriptors are grouped based on the same morphology transition stage. For each group, the combined influence of multiple antecedent descriptors on the consequent descriptors is evaluated. A contribution estimation algorithm is employed, which calculates the variance explained rate of each principal component by performing principal component analysis or nonnegative matrix factorization on the set of antecedent descriptors, serving as a description of morphology-driven aggregation degree.

[0073] Finally, the complete set of chains, the divergence description, and the aggregation description are combined into a graph structure to form the initial causal relationship structure.

[0074] Step 140: Perform latent perturbation factor detection processing on the initial causal relationship structure, detect asymmetric deviation paths of the forward causal relationship chain and the reverse causal relationship chain along the bidirectional reciprocating causal relationship chain, and extract latent processing factors not recorded in the semantic sequence of the processing operation from the asymmetric deviation paths.

[0075] This step is a key advancement in the embodiments of the present invention, aiming to "detect" traces of implicit physical factors that are not explicitly recorded from the "asymmetry" of causal relationships.

[0076] Step 141: Analyze all bidirectional causal relationship chains in the initial causal relationship structure, and extract the positive action strength attribute on the positive causal edge and the negative support strength attribute on the negative causal edge for each bidirectional causal relationship chain.

[0077] The system traverses all edge objects in the initial causal relationship structure graph model and reads the two strength attribute values ​​stored inside each bidirectional causal relationship chain, which are from steps 1331 and 1332 respectively.

[0078] Step 142: Perform a symmetry deviation measurement on the positive action strength attribute and the negative support strength attribute of the same bidirectional reciprocating causal relationship chain. When the difference between the positive action strength attribute and the negative support strength attribute falls into the preset asymmetric interval, mark the bidirectional reciprocating causal relationship chain as a suspected asymmetric causal chain. Collect all suspected asymmetric causal chains to obtain a temporary set of asymmetric causal chains.

[0079] For each chain, the system calculates the absolute value of the difference between the positive action strength attribute value and the negative support strength attribute value. The system maintains a dynamic asymmetric interval, with a lower bound of a basic sensitivity threshold and an upper bound set to the theoretical maximum difference. When the calculated absolute value falls within this interval, it indicates an anomaly in the causal symmetry of the chain, i.e., there is a gap between "positive inference of the ideal" and "reverse tracing of the reality" that cannot be explained by model error. The chain is marked as a suspected asymmetric causal chain and placed in a temporary storage set.

[0080] Step 143: For each suspected asymmetric causal chain in the temporary asymmetric causal chain set, obtain the multi-dimensional descriptive semantics in the corresponding processing operation semantic node, and obtain the morphological state transition description vector and stage duration attribute in the corresponding morphological transition stage.

[0081] The system uses reverse indexing to quickly locate the semantic nodes of the processing operations and the same-morphological transition stages associated with each suspected asymmetric causal chain from the mapping relationship established in step 130, and extracts all their attribute data as the context for analysis.

[0082] Step 144: Based on the causal completeness template of the processing mechanism, perform causal gap analysis on the multidimensional description semantics and the morphological state transition description vector to determine whether there are missing processing motivations that do not appear in the current processing operation semantic node. If there are missing processing motivations, generate a causal gap record.

[0083] The processing mechanism causal completeness template is a conditional network with an "AND-OR graph" structure. The leaf nodes represent various possible morphological change features (decoded from the morphological state transition description vector), and the root node represents the set of processing drivers explaining their causes. The intermediate nodes of the template graph define the "AND" or "OR" logical relationships between these features and drivers.

[0084] When a factual piece of evidence (i.e., a combination of features in the morphological state transition description vector) activates some leaf nodes in the template graph, logical reasoning propagates upwards along the graph to infer the set of necessary motivations that led to this result. This set of necessary motivations is then compared with the explicit multidimensional descriptive semantics contained in the current processing operation semantic node using a set difference operation. If the difference set is not empty, it indicates the existence of missing processing motivations, and the system generates a causal gap record indicating the type of missing motivation.

[0085] Step 145: Extract the processing motivation type pointed to by the causal gap record, and create a latent processing motivation description body based on the processing motivation type and the corresponding action area description semantics. The latent processing motivation description body includes a motivation category label, an action space range description, and an action time interval description.

[0086] The system assigns each missing motivation in the difference set from step 144 as a motivation category label. Its scope of action is directly inherited from the "scope of action description semantics" of the current processing operation semantic node. Its time interval of action is determined based on the stage start timestamp and stage duration attribute of the current same-morphology transition stage. Combining these three elements constructs a new data structure: the implicit processing motivation description body.

[0087] Step 146: Merge and deduplicate the implicit processing motivation descriptions extracted from suspected asymmetric causal chains associated with different processing operation semantic nodes within the same morphological transfer stage to obtain a set of implicit processing factors for that morphological transfer stage; for each set of implicit processing factors, extract the dominant motivation description and associated motivation description from the motivation category label, and transform the description of the action space range into a continuous region coordinate band on the surface of the workpiece, and transform the description of the action time interval into a continuous interval position based on the processing process time sequence to generate implicit processing factors.

[0088] Multiple latent processing drivers belonging to the same time window are likely to originate from the same root physical disturbance. The system clusters these drivers, with the cluster centers representing the latent processing factors at that stage. For each latent processing factor, its most significant driver category label is extracted as the dominant factor, and the rest are related descriptions. Spatially, by calculating the joint coverage area of ​​all the descriptions of the spatial extent of action, a continuous region coordinate zone on the work-in-process surface is obtained. Temporally, the time window in which this factor appears is mapped to the process time sequence position calculated from the start of processing, forming the final latent processing factor object.

[0089] Step 150: Based on the implicit processing factors, adjust the multidimensional description semantics of the corresponding independent processing operation segments in the processing operation semantic sequence to generate a process parameter adjustment instruction sequence.

[0090] This step is the decision-making and execution phase, which aims to transform the detected latent factors into specific process adjustment actions, forming a closed-loop optimization.

[0091] Step 151: Extract the motivation category label, the spatial range description, and the time interval description from the implicit processing factors and form a localization triplet for the implicit processing factors.

[0092] The system deconstructs the implicit processing factor object generated in step 146, extracts these three core attributes, and combines them into a query triplet for location and matching.

[0093] Step 152: Using the time interval description of the action time in the latent processing factor positioning triple as the time constraint condition, traverse each processing operation semantic node in the processing operation semantic sequence, and include the processing operation semantic nodes whose time sequence connection interval attribute symbol falls into the action time interval description into the candidate adjustment node set.

[0094] The system extracts the start and end positions of the time interval descriptions and forms a filter window on a unified timeline. Then, it iterates through all processing operation semantic nodes, checking if their timestamp intervals intersect with this window. All nodes with intersections are grouped into a temporary set of candidate adjustment nodes.

[0095] Step 153: Using the description of the action space range in the latent processing factor positioning triplet as a spatial constraint condition, extract the action region description semantics corresponding to the independent processing operation fragment of each candidate adjustment node in the candidate adjustment node set, and determine the candidate adjustment node whose processing contact surface range contour represented by the action region description semantics overlaps with the action space range description as the processing operation semantic node to be adjusted.

[0096] The system further performs spatial filtering on the candidate adjustment node set. For each candidate node, the geometric boundary representation of its operational region stored in its operational region description semantics is extracted, and spatial polygon intersection or overlap region calculation is performed with the continuous region coordinate band described in the operational spatial range of the triplet. If the calculated overlap area is greater than zero, the candidate node is marked as a semantic node for the processing operation to be adjusted. This means that the operation region of this node coincides with the latent disturbance region, making it the most directly affected by the disturbance and the most suitable object for compensation adjustment.

[0097] Step 154: Invoke the cause type semantic adjustment strategy library, match the cause category label in the cause type semantic adjustment strategy library with the corresponding semantic adjustment strategy according to the latent processing factor location triple, perform semantic adjustment based on the matched semantic adjustment strategy, obtain the adjusted operation mode description semantic, the adjusted energy action form description semantic, and the adjusted action area description semantic, and perform conflict detection, generate the process parameter adjustment instruction sequence according to the conflict detection result and the semantic inverse mapping rule set.

[0098] The motivation type semantic adjustment strategy library is a set of strategies, each bound to a motivation category label. Each strategy consists of three function objects: an operation mode adjustment function, a domain adjustment function, and an energy action form adjustment function. The system uses the motivation category label in the triple to retrieve the corresponding semantic adjustment strategy from the strategy library through exact matching. The matched strategy is then applied to each semantic node of the processing operation to be adjusted.

[0099] Step 1541: Extract the operation mode adjustment function from the matched semantic adjustment strategy, input the operation mode description semantic of the semantic node to be adjusted into the operation mode adjustment function, and generate the adjusted operation mode description semantic through the motivation-driven semantic appending and replacement operation.

[0100] The operation mode adjustment function internally encapsulates a finite state machine. Based on the type of implicit driver input, it executes preset editing logic on the input hierarchical encoding sequence, replacing or appending a certain level of the encoding sequence. For example, when a "vibration" driver is detected, an encoding representing "damping compensation" might be appended to the end of its operation mode encoding sequence.

[0101] Step 1542: Extract the energy action form adjustment function from the matched semantic adjustment strategy, input the energy action form description semantic of the semantic node to be adjusted into the energy action form adjustment function, and generate the adjusted energy action form description semantic through the motivation-driven energy level correction operation.

[0102] This function operates on a multidimensional vector representing the energy interaction form. At its core are a mask vector and an amplification vector. Based on different latent motives, the function generates a mask vector with the same dimensions as the energy vector, identifying which dimensions of energy need to be modified. Then, by weighted summing the original energy vector and the amplification vector along the selected dimensions of the mask, the adjusted energy interaction form description semantics are obtained.

[0103] Step 1543: Extract the scope adjustment function from the matched semantic adjustment strategy, input the scope description semantic of the semantic node to be adjusted into the scope adjustment function, and generate the adjusted scope description semantic through the motivation-driven boundary offset operation.

[0104] This function receives the original region boundary representation and calculates an offset direction and amount based on the motivation type. Specifically, it moves the calculated offset by moving each vertex outward or inward along the normal vector direction of the region boundary and recloses the contour to form the adjusted semantic description of the region of action.

[0105] Step 1544: Combine the adjusted operation mode description semantics, the adjusted action area description semantics, and the adjusted energy action form description semantics into a ternary semantic combination to obtain an adjusted semantic unit to be verified. Then, input the adjusted semantic unit to be verified into the semantic conflict detection rule set for processing and physical common sense conflict verification to generate a conflict detection result.

[0106] The semantic conflict detection rule set contains a series of hard verification rules, such as the conflict between the "coolant on" operation mode and the "high temperature preheating" energy form, or that no parameter value can exceed the machine tool's performance limit. The rule set iterates through and solves new ternary semantic combinations, and once a rule is triggered, it generates a detection result containing a conflict identifier.

[0107] Step 1545: If the conflict detection result contains a conflict identifier, then the conceptual superposition substitution rule set is invoked to recursively correct the conflict semantic field in the semantic unit to be checked and adjusted until the conflict is eliminated, thus obtaining the final adjusted operation mode description semantic, the final adjusted action area description semantic, and the final adjusted energy action form description semantic; if the conflict detection result does not contain a conflict identifier, then the adjusted operation mode description semantic, the adjusted action area description semantic, and the adjusted energy action form description semantic are directly used as the final adjusted operation mode description semantic, the final adjusted action area description semantic, and the final adjusted energy action form description semantic.

[0108] When a conflict is detected, the system identifies the specific semantic field that caused the conflict and then iteratively calls the concept superordinate replacement rule set. This rule set works similarly to the knowledge graph reasoning in step 113. It attempts to find superordinate, parallel, or alternative concepts for the conflicting field, replaces them, and then sends them back to the conflict detection in step 1544. This "detection-replacement" loop will continue to execute until the conflict is completely eliminated or the maximum number of iterations is reached.

[0109] Step 1546: Invoke the semantic reverse mapping rule set to convert the final adjusted operation mode description semantics into a processing operation instruction description, the final adjusted action area description semantics into a processing contact area boundary coordinate description, and the final adjusted energy action form description semantics into a processing energy level index description. Then, assemble the processing operation instruction description, the processing contact area boundary coordinate description, and the processing energy level index description into a process parameter adjustment instruction sequence according to the sequential position of the processing operation semantic nodes to be adjusted in the processing operation semantic sequence.

[0110] The semantic reverse mapping rule set is a set of rule templates that translate standardized semantics back into executable instructions for specific equipment. It encodes and decodes the hierarchical structure of operation methods into a combination of G-codes and M-codes in CNC code; directly uses the list of boundary point coordinates as the boundary coordinates of the machining contact area; and restores the values ​​of each dimension of the multidimensional energy vector to process parameter settings with physical units through inverse normalization. Finally, these instructions are assembled into a sequence according to the order of the corresponding machining operation semantic nodes and sent to the control module of the machining center.

[0111] In a preferred embodiment, to further improve the robustness of the adjustment instructions, the method further includes a verification step based on offline simulation after step 150: Step 161: Input the adjusted processing operation semantic sequence into the surface morphology inference network. Using the node semantic representation vector of each processing operation semantic node in the adjusted processing operation semantic sequence as the inference driver, the morphology state is recursively deduced along the directed sequence chain to obtain the predicted morphology evolution sequence of the entire processing process.

[0112] The surface morphology inference network is a sequence-to-sequence Transformer model. The encoder receives a sequence of adjusted node semantic representation vectors and fuses them with the context of a processing operation affecting the morphology evolution before and after the operation through a multi-head self-attention mechanism. Its decoder then progressively outputs a sequence of predicted morphology state transition vectors that are completely consistent with the format of step 127.

[0113] Step 162: Extract the morphological state transition description vector corresponding to the same morphological transition stage in the initial causal association structure that has a bidirectional reciprocal causal relationship chain with the semantic node of the adjusted processing operation. Use the morphological group identifier change pattern, area descriptor change pattern and intra-group spatial distribution density descriptor change pattern represented by the morphological state transition description vector as the expected morphological evolution reference benchmark.

[0114] The system extracts the historical morphological evolution pattern from the raw data before the processing operation to be adjusted, forming a desired evolutionary baseline.

[0115] Step 163: Perform evolution trend deviation analysis on the predicted morphology evolution sequence of the entire processing process and the expected morphology evolution reference benchmark. Compare the consistency of the change direction of the morphology group identifier, the change direction of the area descriptor, and the change direction of the spatial distribution density descriptor within the same morphology transfer stage along the time axis segment by segment. Mark the predicted morphology evolution trajectory segment with a deviation exceeding the preset trend tolerance threshold as an evolution anomaly segment, and trace the processing operation semantic node corresponding to the evolution anomaly segment to generate semantic-level deviation location information. Using the semantic-level deviation location information as an index, retrieve the implicit processing driving factor descriptor that matches the description of the time interval and the spatial range of the abnormal segment from the implicit processing factor set. Mark the implicit processing driving factor descriptor as a disturbance factor to be suppressed, and match the reverse compensation type semantic adjustment strategy in the driving factor type semantic adjustment strategy library according to the disturbance factor to be suppressed. Perform secondary semantic correction on the adjusted processing operation semantic node corresponding to the abnormal segment, and output the process parameter compensation instruction sequence.

[0116] If, within a certain time interval, the direction of the morphological transfer vector of the predicted sequence deviates from the direction of the expected baseline, and the cosine similarity is lower than the tolerance threshold, then that interval is determined to be abnormal. Subsequently, the set of latent processing factors is retrieved in reverse using the spatiotemporal information of this abnormal interval to locate the cause, and a stronger reverse compensation strategy from the strategy library is invoked for secondary adjustment to generate compensation instructions.

[0117] In yet another preferred embodiment, the method further includes a dynamic adjustment closed loop based on online real-time data after step 150.

[0118] Step 171: Collect the set of online morphology primitives that characterize the real-time evolution of surface micromorphology during the execution of adjusted process parameters on the production line, and establish an online morphology state transition sequence based on the collection time nodes.

[0119] During the parameter adjustment process, the online measurement probe located in the machining center continuously collects data and generates an online morphology state transition sequence in real time using the same processing flow as steps 121 to 128.

[0120] Step 172: Extract the online homomorphic transition stage in the initial causal association structure that has a bidirectional reciprocating causal relationship chain with the semantic node of the currently executing processing operation, and perform synchronization detection between the real-time change trend of the morphological group identifier represented by the morphological state transition description vector of the online homomorphic transition stage and the positive inferred morphological state transition change description pointed to by the positive causal edge in the bidirectional reciprocating causal relationship chain.

[0121] The system compares the real-time generated topographic state transition vector with the ideal positive inferred topographic state transition tensor stored in the positive edges of the causal chain.

[0122] Step 173: When the synchronization detection result indicates that the change direction of the morphological group identifier deviates from the positive inferred morphological state transition change description, it is determined that the current processing operation semantic node is in an active state of latent processing factor perturbation. The latent processing motivation description body is extracted from the asymmetric deviation path of the bidirectional reciprocating causal relationship chain corresponding to the processing operation semantic node. The latent processing motivation description body is then used to reproduce the perturbation propagation path with the instantaneous morphological state transition description vector in the online morphological state transition sequence to generate a real-time perturbation propagation chain of latent processing factors.

[0123] Once the actual effect of the current processing step is detected to deviate from the theoretical prediction in real time, the system immediately marks the state as active disturbance and reconstructs a specific disturbance propagation chain from the current real-time data along a similar logic to steps 140 and 146.

[0124] Step 174: Based on the description of the spatial range of each disturbance node in the real-time disturbance propagation chain of the latent processing factors, delineate the disturbed morphological unit group in the online morphological primitive set, and use the time interval of the real-time disturbance propagation chain of the latent processing factors as the boundary. Extract the evolution curve of the spatial distribution density descriptor within the group of disturbed morphological units, input the evolution curve of the spatial distribution density descriptor into the disturbance pattern recognition classifier, and output the disturbance attenuation type label. Based on the disturbance attenuation type label, retrieve the online process parameter fine-tuning strategy adapted to the current processing stage from the preset disturbance suppression strategy library, and generate the online process parameter real-time correction instruction sequence.

[0125] A pre-trained perturbation pattern recognition classifier, such as a temporal convolutional network, receives the evolution curve of the density descriptor of the perturbed region as input. It extracts temporal features such as decay, divergence, or oscillation through operations like convolution and pooling. The output, after passing through fully connected layers and a Softmax function, categorizes these features into labels such as "fast decay type," "continuous perturbation type," or "slow divergence type." A corresponding perturbation suppression strategy library, based on these labels, retrieves appropriate fine-tuning strategies, such as modifying the feed rate or injecting active damping signals, to generate real-time correction instructions.

[0126] Based on the above, the various models, networks, or algorithm modules involved in the embodiments of the present invention will be further described below as a further disclosure of the foregoing technical solutions, so that those skilled in the art can implement the technical solutions of the present invention without any doubt.

[0127] The pre-trained process text syntactic dependency parsing model used in step 111 adopts a classic dependency parser structure based on a bidirectional long short-term memory network and a dual affine attention mechanism.

[0128] Specifically, the word embedding layer of the model has an output dimension of 100, followed by two stacked bidirectional long short-term memory network layers, each with a hidden state dimension of 200, used to extract context-sensitive features for each sub-word unit. The key module of the model is the dual affine attention dependency decoder, which predicts the existence of a dependency relationship and the type of dependency arc by calculating the dual affine transformation score between the feature vectors of any two sub-word units. Regarding training data and processes, the model is pre-trained on a corpus containing hundreds of thousands of sentences labeled with general domain information and tens of thousands of sentences labeled with specific process domain information. The input is segmented and part-of-speech tagged process text sentences, and the output is a directed dependency tree with dependency arc labels.

[0129] During the pre-training phase, the AdamW optimizer was used, with an initial learning rate set to a low value (e.g., 0.001) and a batch size of 16. The cross-entropy loss based on dependency arc prediction was used to optimize the entire network parameters. Early stopping was determined based on the unlabeled dependency sentence accuracy on the validation set during training epochs. In terms of model application, the byte-pair encoded word segmentation results of the process text were directly input into the model during the inference phase of step 111. The dependency tree structure output by the model was then used for subsequent segmentation logic.

[0130] Furthermore, the technical implementation of the three encoder sub-networks within the semantic vector mapping module used in step 115 also needs to be clarified. First, the graph convolutional network used to encode the semantic description of the operation mode consists of three stacked graph convolutional layers. The message passing formula for each graph convolutional layer can be summarized as linear transformation and non-linear activation of the aggregation result of the node's own features and the features of its neighboring nodes. Second, the point cloud feature extraction network used to encode the semantic description of the action region adopts a PointNet structure, containing two multilayer perceptron layers with shared weights. The output dimension of each layer increases sequentially to 1024 dimensions, followed by a max-pooling symmetric function layer to aggregate global features. Third, the fully connected network used to encode the semantic description of the energy action form is a three-layer multilayer perceptron, with the intermediate layer dimension set to 128 dimensions, and the output being a fixed-length vector. These three sub-networks are not pre-trained but are jointly optimized in the end-to-end simulation training described below as part of the overall method pipeline. Therefore, their training process and data are detailed in the subsequent training description of the overall method.

[0131] The surface topography inference network introduced in step 161 is specifically structured as a standard Transformer encoder-decoder sequence-to-sequence model. Both the encoder and decoder consist of stacked encoder and decoder layers of their respective numbers. Each encoder layer contains a multi-head self-attention subnetwork and a position-wise feedforward neural network subnetwork, with layer normalization and residual connections applied to their respective outputs. The number of attention heads in the multi-head self-attention subnetwork can be configured to a fixed value (e.g., 8).

[0132] The decoder layer structure is similar, but a multi-head cross-attention subnetwork is inserted between the multi-head self-attention subnetwork and the feedforward neural network subnetwork to focus on the features of the encoder output. This surface morphology inference network requires specialized training. Its training data comes from historical data pairs accumulated through "trial cutting-measurement" on the same or similar production lines, ensuring that the data volume meets statistical requirements. During training, the input is an adjusted semantic sequence of processing operations, and the supervision signal is the corresponding sequence of real morphology state transition vectors.

[0133] The model uses the Adam optimizer for parameter updates, with an initial learning rate of 0.0001 and a batch size of 4. Training epochs are used to achieve early stopping by monitoring the mean squared error between the predicted and true vectors. In application, the adjusted semantic sequence of processing operations is directly fed into the encoder, and the decoder gradually generates the future predicted morphological state transition vector sequence through an autoregressive approach.

[0134] Furthermore, the perturbation pattern recognition classifier used in step 174 employs a temporal convolutional network architecture. This network consists of several residual blocks connected sequentially. Each residual block contains two dilated causal convolutional layers, a weight normalization layer, a ReLU activation layer, and a random deactivation layer to efficiently extract multi-scale patterns from the time series. Its output is passed through a global average pooling layer and then input into a fully connected layer, where the probability of each category is output through a Softmax activation function. In terms of training data and process, the classifier is trained using a sample set of intra-group spatial distribution density descriptor evolution curves containing different perturbation types (such as rapid decay, continuous oscillation, etc.). The sample set must contain a sufficient number of curves for each category. The optimizer used for training is SGD, and the loss function is the classification cross-entropy loss. Training epochs are stopped early based on the macro-average accuracy on the test set. In terms of model application, during runtime, the density evolution curves within a continuous time period are directly input into the network, and the class labels output by the network are used for subsequent policy matching.

[0135] Please refer to the following: Figure 1 and Figure 2 The method provided in this embodiment of the invention revolves around two core data streams: one is the semantic stream of processing operations, and the other is the physical evolution stream of surface morphology. These two data streams are initially parsed independently, then deeply converge through an intermediate causal correlation module, ultimately forming a process adjustment instruction output. Figure 2 The small and medium-sized boxes (e.g., operation segments, semantic analysis, etc.) indicate the key processing actions at each stage, and the arrows indicate the direction of data flow. The production line scenario in this embodiment is a CNC machining production line for precision metal parts. The production data analysis system is connected to the process specification database, online surface morphology measurement equipment, and machining center controller via a data interface to fully execute all steps of this embodiment.

[0136] At the beginning of the process, the method initiates two independent data parsing branches in parallel, corresponding to the left and right input paths in the attached diagram, labeled "extracting process text and decomposing it into multi-dimensional descriptive semantics" and "decomposing surface micro-morphology data into morphology primitives and establishing an evolutionary sequence," respectively. The left branch executes step 110, the purpose of which is to transform the natural language text in the process specification into a machine-computable time-ordered semantic sequence. Within step 110, syntactic analysis and semantic segmentation are first completed in step 111, reliably decomposing the coherent process text into independent processing operation fragments.

[0137] Next, in step 112, multidimensional semantic slots are filled for each segment, generating operation mode description semantics, action area description semantics, and energy action form description semantics, which respectively define a processing operation from the three dimensions of action type, spatial range, and energy input.

[0138] Subsequently, through joint rationality judgment and conceptual superposition correction in step 113, the multidimensional semantics are ensured to be physically and logically compatible and consistent in abstraction level, and the correction results are encapsulated into processing operation semantic nodes. In step 114, based on the parsing of the original text's statement order and temporal connection adverbs, directed edges labeled with "immediately following" or "regular interval" are established between nodes, forming a directed sequence chain. Finally, in step 115, the descriptive semantics are converted into vectors by calling graph convolutional networks, point cloud feature extraction networks, and fully connected networks respectively, and then concatenated and fused to generate node semantic representation vectors. These vectors are then chained together according to the directed sequence chain to obtain the processing operation semantic sequence. Through this branch, the unstructured process text is transformed into a one-dimensional feature sequence with temporal attributes and rich semantic information.

[0139] The right branch executes step 120, the purpose of which is to transform high-dimensional surface micro-topography measurement data into a structured, multi-scale set of topography primitives, and on this basis, capture their dynamic evolution as the processing progresses.

[0140] Within step 120, the surface micro-topography height field matrix is ​​extracted from the measurement data at each acquisition time point through step 121, and the physical topography is converted into a regular numerical matrix.

[0141] Step 122 performs multi-level topographic undulation analysis on the matrix, using a spatial frequency domain filter bank to decompose the topography into topographic undulation morphological primitives at different scales, and defines independent topographic morphological units by extracting contour closure boundaries, assigning level labels and area descriptors. Step 123 Within the same scale, topographic morphological units are clustered into topographic morphological unit groups by judging spatial adjacency, and assigned morphological group identifiers and intra-group spatial distribution density descriptors, thereby organizing isolated units into groups with spatial structural significance. Step 124 analyzes the geometric inclusion relationship between coarse and fine-scale topographic morphological unit groups across levels, establishes a cross-level nested structure, generates a multi-scale topographic primitive set, and achieves a complete representation of the surface topography at a single time point. Step 125 Between adjacent acquisition time points, topographic morphological unit group matching pairs are established through morphological group identifier matching and spatial proximity requirements, realizing the traceability association of topographic primitives on the time axis. Step 126: For each matching pair, calculate the changes in its morphological group identifier, area descriptor, and intra-group spatial distribution density descriptor, and aggregate them into a morphological state transition description vector. Step 127: Arrange the morphological state transition description vectors of all matching pairs in the time direction to obtain a morphological state transition vector sequence. Step 128: Merge the morphological transition patterns of this sequence, merging continuous vector segments with similar change patterns into morphological transition stages, assigning stage start timestamps and stage duration attributes, and finally encapsulating them into a surface morphological evolution sequence. Thus, the right branch compresses and organizes a large number of discrete morphological snapshots into a time-varying sequence consisting of several morphological transition stages, clearly reflecting the dynamic changes in morphology.

[0142] After the left and right branches generate the semantic sequence of processing operations and the evolution sequence of surface morphology respectively, the two data streams enter the core intersection module marked in the attached figure as "synchronous association of processing operations and morphology evolution generation causal relationship chain", corresponding to step 130.

[0143] Step 130 aims to establish a strong causal relationship between the two time series. Step 131 first extracts the time interval information for each homomorphic transition stage, using it as a benchmark. Based on the position and temporal connection interval attributes of the processing operation semantic nodes on the directed sequence chain, the semantic nodes are mapped to the corresponding time intervals, completing the time alignment mapping. Step 132, for each successfully mapped pair, extracts the operation mode description semantics of the processing operation semantic nodes as antecedent descriptors and extracts the homomorphic state transition description vectors of the homomorphic transition stages as consequence descriptors. Step 133 then determines the establishment of a strong causal relationship through bidirectional causal logic evaluation. The positive action logic evaluation, based on the physical influence rule base of the processing mode, derives the ideal consequence from the antecedent and compares it with the actual consequence; the reverse tracing logic evaluation, based on the homomorphic backtracking rule base, infers the possible set of antecedents from the consequence and verifies the membership of the current antecedent. Only pairs with both bidirectional evaluations being positive are established between them as positive and negative causal edges, forming a bidirectional reciprocating causal relationship chain. Finally, all pairs are traversed to generate the complete set of links, and the divergence and aggregation descriptions of the complete set are calculated. The complete set of links is then combined with these macroscopic descriptive information to form the initial causal relationship structure.

[0144] After obtaining the initial causal relationship structure, the process enters the analysis module marked "Detecting latent disturbance factors and extracting unrecorded processing factors" in the attached figure, corresponding to step 140.

[0145] The core basis of step 140 is that the existence of implicit processing factors leads to anomalies in the symmetry of the causal chain. Step 141 analyzes the positive action strength attribute and negative support strength attribute of all bidirectional reciprocal causal relationship chains. Step 142 measures the symmetry deviation of the difference between these two strength attributes, marking causal chains whose differences fall into a preset asymmetric interval as suspected asymmetric causal chains, forming a temporary set of asymmetric causal chains. Step 143 extracts the multidimensional descriptive semantics, morphological state transition description vector, and their temporal attributes corresponding to each suspected asymmetric causal chain, forming the analysis context. Step 144 performs causal gap analysis based on the causal completeness template of the processing mechanism, deducing the set of drivers necessary to explain the current morphological change through the "AND-OR graph" logic of the template, and performing a set difference operation between this set and the explicit drivers in the current processing operation semantic node, thereby locating the missing processing drivers and generating causal gap records. Step 145 creates an implicit processing driver description body based on the missing driver type in the gap record, assigning it a driver category label, a description of the scope of action space, and a description of the time interval of action. Step 146 merges and removes duplicate descriptions of multiple implicit processing factors within the same morphological transfer stage, extracts the dominant and related driving factor descriptions, and transforms the spatial and temporal descriptions into continuous regional coordinate zones and continuous intervals in the process time sequence to generate the final implicit processing factor object.

[0146] Finally, the process enters the decision output module marked "adjust semantic generation process parameter adjustment instruction sequence" in the attached diagram, corresponding to step 150.

[0147] Step 150 aims to make compensatory adjustments to the original processing operations based on the detected latent processing factors. Step 151 extracts the motivation category label, spatial range description, and temporal interval description from the latent processing factors to form a latent processing factor localization triple. Step 152 uses the time interval in the triple as a constraint to filter out temporally related processing operation semantic nodes, forming a candidate adjustment node set. Step 153 uses the spatial range in the triple as a constraint to perform a secondary screening of the candidate nodes, identifying nodes whose area of ​​action overlaps with the area of ​​influence of latent factors as the processing operation semantic nodes to be adjusted. Step 154 ​​calls the motivation type semantic adjustment strategy library and matches the corresponding semantic adjustment strategy according to the motivation category label in the triple. Through the operation mode adjustment function, energy action form adjustment function, and area of ​​action adjustment function in the strategy, motivation-driven semantic correction is performed on the three-dimensional semantics of the node to be adjusted. The corrected semantic unit needs to undergo conflict detection and recursive correction in step 1545 to ensure that it is physically reasonable and free from internal contradictions. Finally, the semantic reverse mapping rule set is called to reverse the adjusted three-dimensional description semantics into processing operation instruction description, processing contact area boundary coordinate description and processing energy level index description, and then assemble them in sequence into a process parameter adjustment instruction sequence and send it to the processing center.

[0148] As a closed-loop optimization branch, the method can optionally introduce a verification closed loop based on offline simulation or a dynamic adjustment closed loop based on online real-time data after step 150. The offline closed loop predicts the adjusted morphology evolution through a surface morphology inference network and performs trend deviation analysis on it with the expected evolution benchmark. For the discovered abnormal evolution sections, it retrieves the implicit processing factors again and matches them with a reverse compensation strategy to generate compensation instructions. The online closed loop collects morphology data in real time and constructs an online state transition sequence during the parameter adjustment process. By synchronizing the real-time evolution trend with the forward inference description, it determines the active state of the disturbance when a deviation is found, reproduces the disturbance propagation chain in real time, and uses a disturbance pattern recognition classifier to identify the disturbance attenuation type. Then, it retrieves an online fine-tuning strategy from the disturbance suppression strategy library and generates real-time correction instructions. These two closed loops ensure the robustness and real-time adaptability of the process adjustment instructions from the two dimensions of offline verification and online compensation, respectively.

[0149] This invention constructs two independent processing links: multi-dimensional semantic parsing of process text and multi-scale evolution analysis of surface micromorphology. It innovatively introduces a bidirectional reciprocating causal relationship chain as the core mechanism for cross-link association, enabling the relationship between processing operations and morphological responses to extend beyond temporal synchronization to a level of causal logical mutual verification. By actively detecting asymmetric deviation paths between the positive and negative support strengths of the causal relationship chain, it achieves, for the first time, the systematic detection and extraction of implicit processing factors that are not recorded in the process specifications but substantially affect processing quality. Finally, driven by these implicit processing factors, it performs targeted adjustments to the semantic descriptions of operation methods, action areas, and energy action forms in the processing operation semantic sequence, generating a physically interpretable sequence of process parameter adjustment instructions. This elevates traditional post-processing quality analysis to a closed-loop process optimization based on causal inference, enhancing the interpretability of metal product processing and the proactiveness of process optimization.

[0150] Please see Figure 3 The figure is a schematic diagram of the basic structure of a production data analysis system 200 for metal product processing provided in an embodiment of this application. The production data analysis system 200 for metal product processing includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the production data analysis methods for metal product processing described above.

[0151] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0152] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. A processor in a production data analysis system for metal product processing reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the production data analysis system for metal product processing to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0153] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A production data analysis method applied to metal product processing, characterized by, include: Independent processing operation segments are extracted from the process text in the metal product process specification to perform multidimensional semantic decomposition, generating multidimensional descriptive semantics corresponding to each independent processing operation segment and connecting them in time sequence to form a processing operation semantic sequence; the multidimensional descriptive semantics include operation mode description semantics, action area description semantics, and energy action form description semantics. Multi-scale morphology primitive decomposition is performed on the surface micromorphology data acquired at each acquisition time node in the metal product processing production line to obtain a set of morphology primitives. Based on the acquisition time node, a morphology state transition sequence of the morphology primitive set is established as a surface morphology evolution sequence. Each independent processing operation segment in the processing operation semantic sequence is temporally synchronized with the morphology state transition in the corresponding time period in the surface morphology evolution sequence. A bidirectional reciprocating causal relationship chain is generated through positive action logic evaluation and reverse tracing logic evaluation. All bidirectional reciprocating causal relationship chains are combined into an initial causal relationship structure. The initial causal relationship structure is subjected to latent perturbation factor detection processing. Asymmetric deviation paths between the forward and reverse causal relationship chains are detected along the bidirectional reciprocating causal relationship chain. Latent processing factors not recorded in the semantic sequence of the processing operation are extracted from the asymmetric deviation paths. Based on the implicit processing factors, the multidimensional description semantics of the corresponding independent processing operation segments in the processing operation semantic sequence are adjusted to generate a process parameter adjustment instruction sequence.

2. The method of claim 1, wherein, The process text in the metal product process specification is extracted into independent processing operation segments for multidimensional semantic decomposition, generating multidimensional descriptive semantics corresponding to each independent processing operation segment and connecting them in time sequence to form a processing operation semantic sequence. The multidimensional descriptive semantics includes operational mode descriptive semantics, action region descriptive semantics, and energy action form descriptive semantics, including: Syntactic analysis and semantic segmentation are performed on the process text of metal product process specifications to obtain independent processing operation segments; Extract and categorize the operation events of each independent machining operation segment to generate operation mode description semantics; determine the geometric boundary representation of the machining contact area based on the workpiece geometry description to generate action area description semantics; extract the process parameters in the independent machining operation segments and combine them into energy action form description semantics. The semantic descriptions of the operation mode, the semantic description of the action area, and the semantic description of the energy action form are jointly judged for rationality and revised by conceptual superposition. Based on the revision results, a ternary semantic encapsulation is performed to generate a processing operation semantic node. Based on the order of the process text, the sequential relationship between the semantic nodes of each processing operation is determined and a temporal connection interval attribute label is assigned to obtain a directed sequence chain; The semantic descriptions of each processing operation semantic node are converted into vectors by semantic vector mapping and fused into node semantic representation vectors. The processing operation semantic sequence is generated by concatenating the nodes according to the directed order chain.

3. The method of claim 1, wherein, The process involves multi-scale morphology primitive decomposition of surface micro-morphology data acquired at various time points in the metal product processing production line to obtain a set of morphology primitives. A morphology state transition sequence of this set of primitives is then established as a surface morphology evolution sequence based on the acquisition time points, including: From the surface micromorphology data acquired at various time points in the metal product processing production line, the surface micromorphology height field matrix corresponding to each acquisition time point is extracted. The surface micro-topography height field matrix is ​​subjected to multi-level topography undulation layer analysis to generate topography undulation morphology primitives at different level scales. Independent contour closure boundary is extracted for each topography undulation morphology primitive at each level scale to obtain topography morphology units with closed perimeters. Each topography morphology unit is labeled with a level label and area descriptor. Morphological units with spatial adjacency at the same level are grouped into morphological unit groups according to the degree of contact between adjacent boundaries, and each morphological unit group is assigned a morphological group identifier and a spatial distribution density descriptor within the group. For each acquisition time node, cross-level nesting relationship analysis is performed on the morphological unit groups at all levels. The inclusion relationship structure of fine-level morphological unit groups located inside coarse-level morphological unit groups is established, and a multi-scale morphological primitive set for that acquisition time node is generated. The multi-scale morphological primitive sets of adjacent acquisition time nodes are matched and traced by morphological unit group matching. For each morphological unit group located at the previous acquisition time node, the corresponding morphological unit group in the subsequent acquisition time node whose morphological group identifier meets the matching condition and satisfies the requirement of spatial proximity is found, and time-adjacent morphological unit group matching pairs are established. For each temporally adjacent morphological unit group matching pair, the morphological group identifier change description, area descriptor change description, and intra-group spatial distribution density descriptor change description of the morphological unit group are calculated and aggregated into a morphological state transition description vector. Arrange the topographic state transition description vectors of all time-adjacent matching pairs according to the time progression direction of the acquisition time nodes to obtain the topographic state transition vector sequence; The morphological state transition vector sequence is merged into a morphological transition pattern. Continuous matching pairs with similar morphological state transition description vectors are merged into the same morphological transition stage. Each same morphological transition stage is assigned a stage start timestamp and a stage duration attribute. The stage start timestamp and stage duration attributes of different morphological transition stages, as well as the description vectors of all morphological state transitions within that stage, are encapsulated into a surface morphological evolution sequence.

4. The method according to claim 1, characterized in that, The process involves temporally synchronizing and associating each independent processing operation segment in the semantic sequence of the processing operations with the corresponding time period of the morphology state transition in the surface morphology evolution sequence. A bidirectional reciprocal causal relationship chain is generated through positive action logic evaluation and reverse tracing logic evaluation. All bidirectional reciprocal causal relationship chains are combined into an initial causal association structure, including: Extract the stage start timestamp and stage duration attributes corresponding to each homomorphic transition stage in the surface morphology evolution sequence; Using the stage start timestamp and stage duration attribute as time interval scales, the occupied interval of each homomorphic transfer stage is defined on the time axis, and the processing operation semantic nodes within the time axis coverage are mapped to the occupied interval of the corresponding homomorphic transfer stage according to their temporal position, so as to perform time alignment mapping between processing operation semantic nodes and homomorphic transfer stages. For each time-aligned and successfully mapped processing operation semantic node and the same shape transition stage pair, the operation mode description semantic of the processing operation semantic node in the pair is extracted as the antecedent descriptor, and the shape state transition description vector in the same shape transition stage is extracted as the consequence descriptor. By combining the antecedent descriptor and the consequence descriptor, the positive action logic evaluation and the reverse tracing logic evaluation are initiated, and the initial causal relationship structure is determined based on the positive causal relationship establishment identifier and the reverse tracing relationship establishment identifier.

5. The method according to claim 4, characterized in that, The process of combining the antecedent descriptor and the consequence descriptor to initiate forward action logic evaluation and reverse causation logic evaluation, and determining the initial causal relationship structure based on the forward causal relationship establishment identifier and the reverse causation relationship establishment identifier, includes: Initiate a positive action logic evaluation, and determine the ideal morphological state transition and change of the processing action type pointed to by the antecedent descriptor on the metal surface based on the physical influence rule base of the processing method. Obtain a positive presumed morphological state transition and change description, and logically compare the positive presumed morphological state transition and change description with the consequence descriptor to generate a positive causal relationship establishment identifier. Initiate reverse tracing logic evaluation, and based on the morphological backtracking rule base, back-infer the set of candidate processing action types that can produce the morphological state transition from the consequence descriptor, determine whether the antecedent descriptor falls into the set of candidate processing action types, and generate a reverse tracing relationship establishment identifier; When both the positive causal relationship establishment flag and the reverse tracing relationship establishment flag are in a positive state, a positive causal edge is established for the pair, pointing from the processing operation semantic node to the same shape transfer stage, and a reverse causal edge is established from the same shape transfer stage to the processing operation semantic node. The positive causal edge and the reverse causal edge are then paired into a bidirectional reciprocating causal relationship chain. Iterate through all processing operation semantic nodes and similar shape transition stage pairs that have successfully time-aligned mappings, generate corresponding bidirectional reciprocal causal relationship chains one by one, and generate a complete set of bidirectional reciprocal causal relationship chains. The multi-dimensional bidirectional reciprocating causal relationship chains with the same processing operation semantic node in the complete set of the bidirectional reciprocating causal relationship chains are aggregated. Within the aggregation unit, the response heterogeneity of different homomorphic transition stages to the same processing operation semantic node is evaluated, and a description of processing semantic divergence is generated. The multi-dimensional bidirectional reciprocating causal relationship chains with the same morphological transition stage in the complete set of the bidirectional reciprocating causal relationship chains are aggregated. Within the aggregation unit, the contribution superposition of different processing operation semantic nodes to the same morphological transition stage is evaluated to generate a morphological-driven aggregation degree description. The complete set of bidirectional reciprocating causal relationship chains is associated and combined with the description of processing semantic divergence and the description of morphology-driven aggregation to obtain the initial causal relationship structure.

6. The method according to claim 1, characterized in that, The process of performing latent perturbation factor detection on the initial causal relationship structure, detecting asymmetric deviation paths between the forward and reverse causal chains along the bidirectional reciprocating causal relationship chain, and extracting latent processing factors not recorded in the semantic sequence of the processing operation from the asymmetric deviation paths includes: Analyze all bidirectional cyclic causal relationship chains in the initial causal relationship structure, and extract the positive action strength attribute on the positive causal edge and the negative support strength attribute on the negative causal edge for each bidirectional cyclic causal relationship chain; The positive action strength attribute and the negative support strength attribute of the same bidirectional reciprocating causal relationship chain are measured for symmetry deviation. When the difference between the positive action strength attribute and the negative support strength attribute falls into the preset asymmetric interval, the bidirectional reciprocating causal relationship chain is marked as a suspected asymmetric causal chain. All suspected asymmetric causal chains are collected to obtain a temporary set of asymmetric causal chains. For each suspected asymmetric causal chain in the asymmetric causal chain temporary storage set, obtain the multi-dimensional description semantics in the corresponding processing operation semantic node, and obtain the morphological state transition description vector and stage duration attribute in the corresponding morphological transition stage. Based on the causal completeness template of the processing mechanism, causal gap analysis is performed on the multidimensional description semantics and the morphological state transition description vector to determine whether there are missing processing motivations that do not appear in the current processing operation semantic node. If there are missing processing motivations, a causal gap record is generated. Extract the processing motivation type pointed to by the causal gap record, and create a latent processing motivation description body based on the processing motivation type and the corresponding action area description semantics. The latent processing motivation description body includes motivation category label, action space range description and action time interval description. The implicit processing motivation descriptions extracted from suspected asymmetric causal chains associated with different processing operation semantic nodes within the same homomorphic transfer stage are merged and deduplicated to obtain the set of implicit processing factors for that homomorphic transfer stage. For each set of implicit processing factors, the dominant and associated driver descriptions in the driver category labels are extracted, and the description of the action space range is transformed into a continuous region coordinate band on the surface of the workpiece, and the description of the action time interval is transformed into a continuous interval position based on the processing time sequence, thus generating implicit processing factors.

7. The method according to any one of claims 1-6, characterized in that, The step of adjusting the multidimensional descriptive semantics of the corresponding independent processing operation segments in the processing operation semantic sequence based on the implicit processing factors, and generating a process parameter adjustment instruction sequence, includes: Extract the motivation category label, the description of the spatial range of action, and the description of the time interval of action from the implicit processing factors and form a localization triplet for the implicit processing factors; Using the time interval of action in the latent processing factor location triple as the time constraint, traverse each processing operation semantic node in the processing operation semantic sequence, and classify the processing operation semantic nodes whose temporal connection interval attribute symbol falls into the time interval of action description into the candidate adjustment node set. Using the description of the action space range in the latent processing factor positioning triplet as a spatial constraint, extract the action region description semantics corresponding to the independent processing operation fragment of each candidate adjustment node in the candidate adjustment node set, and determine the candidate adjustment node whose processing contact surface range contour represented by the action region description semantics overlaps with the action space range description as the processing operation semantic node to be adjusted. The motivation type semantic adjustment strategy library is invoked. Based on the motivation category label in the latent processing factor location triple, the corresponding semantic adjustment strategy is matched in the motivation type semantic adjustment strategy library. The motivation type semantic adjustment strategy library stores the mapping relationship between motivation category label and operation mode adjustment function, action area adjustment function and energy action form adjustment function. Based on the matched semantic adjustment strategy, semantic adjustment is performed to obtain the adjusted operation mode description semantic, the adjusted energy action form description semantic, and the adjusted action area description semantic. Conflict detection is then performed, and a process parameter adjustment instruction sequence is generated based on the conflict detection results and the semantic inverse mapping rule set.

8. The method according to claim 7, characterized in that, The semantic adjustment based on the matched semantic adjustment strategy is performed to obtain the adjusted operation mode description semantics, the adjusted energy action form description semantics, and the adjusted action region description semantics. Conflict detection is then performed, and a process parameter adjustment instruction sequence is generated based on the conflict detection results and the semantic inverse mapping rule set, including: Extract the operation mode adjustment function from the matched semantic adjustment strategy, input the operation mode description semantic of the semantic node to be adjusted into the operation mode adjustment function, and generate the adjusted operation mode description semantic through the motivation-driven semantic appending and replacement operation. Extract the energy action form adjustment function from the matched semantic adjustment strategy, input the energy action form description semantic of the semantic node to be adjusted into the energy action form adjustment function, and generate the adjusted energy action form description semantic through the motivation-driven energy level correction operation. Extract the scope adjustment function from the matched semantic adjustment strategy, input the scope description semantic of the semantic node to be adjusted into the scope adjustment function, and generate the adjusted scope description semantic through the boundary offset operation driven by the motivation. The adjusted operation mode description semantics, the adjusted action area description semantics, and the adjusted energy action form description semantics are combined into a ternary semantic combination to obtain an adjusted semantic unit to be verified. The adjusted semantic unit to be verified is then input into a semantic conflict detection rule set for processing and physical common sense conflict verification to generate a conflict detection result. If the conflict detection result contains a conflict identifier, then the concept-superior substitution rule set is invoked to recursively correct the conflict semantic field in the semantic unit to be checked and adjusted until the conflict is eliminated, resulting in the final adjusted operation mode description semantic, the final adjusted action area description semantic, and the final adjusted energy action form description semantic; if the conflict detection result does not contain a conflict identifier, then the adjusted operation mode description semantic, the adjusted action area description semantic, and the adjusted energy action form description semantic are directly used as the final adjusted operation mode description semantic, the final adjusted action area description semantic, and the final adjusted energy action form description semantic. The semantic reverse mapping rule set is invoked to convert the semantic description of the final adjusted operation mode into a processing operation instruction description, the semantic description of the final adjusted action area into a processing contact area boundary coordinate description, and the semantic description of the final adjusted energy action form into a processing energy level index description. The processing operation instruction description, the processing contact area boundary coordinate description, and the processing energy level index description are then assembled into a process parameter adjustment instruction sequence according to the sequential position of the processing operation semantic nodes to be adjusted in the processing operation semantic sequence.

9. A production data analysis system applied to metal product processing, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the production data analysis method for metal product processing as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the production data analysis method for metal product processing as described in any one of claims 1-8.

Images