A method for diagnosing error sources in a precision micro connector mold manufacturing process

Abstract

This invention relates to a method for diagnosing error sources in the manufacturing process of precision micro-connector molds. Addressing the challenge of integrating multi-source heterogeneous detection data and attributing errors during the manufacturing process, it proposes a method based on semantic fingerprint extraction and incremental fusion of knowledge graphs driven by multi-source raw data. Through industrial Ethernet, dedicated communication protocols, and optical character recognition, data from CNC machining, coordinate measuring machine (CMM), electrical discharge machining (EDM), and maintenance fault logs are collected in real time, and data feature compression, geometric fingerprinting, material response analysis, and semantic phrase standardization are performed. The core solution automatically constructs and evolves a knowledge graph through low-dimensional semantic fingerprint spatial mapping and anchor point expansion mechanisms, enabling real-time location of manufacturing error sources and the formation of causal chain attribution reasoning. This method achieves dynamic identification and ontology fusion of novel error patterns, improving the efficiency of anomaly diagnosis and the self-evolutionary capability of knowledge in the mold manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis and knowledge management technology in mold manufacturing, and in particular to a method for diagnosing error sources in the manufacturing process of precision micro connector molds. Background Technology

[0002] Precision micro-connector mold manufacturing, due to the extremely small size and complex structure of components, is subject to multiple interferences and has very little tolerance for errors during the processing. Error source diagnosis and self-healing place extremely high demands on manufacturing capabilities. With the increasing trend of mass customization of highly integrated electronic components and rapid replacement of various product types, mold manufacturing workshops... It is gradually evolving towards intelligence, digitalization, and networking. Existing mainstream technologies generally adopt two paths: one is ontology-driven knowledge graphs that rely mainly on manual annotation and domain expert knowledge, and the other is unstructured knowledge extraction methods based on end-to-end neural networks (including large language models).

[0003] Currently, most typical intelligent diagnostic solutions in the industry rely on traditional knowledge graph construction technology. This involves process experts manually identifying and defining error sources, process parameters, fault phenomena, and remedial measures, and then mapping these rules into a searchable graph-structured database using knowledge engineering methods. This ontology-based approach ensures manageability and high cohesion of knowledge, making it suitable for mass production scenarios with stable rules and slow process upgrades. However, for precision micro-mold applications, the error characteristics are highly heterogeneous across equipment, processes, physical components, and semantics. Dynamic changes frequently occur, such as parameter variations in the process flow, material batch disturbances, and unpredictable equipment conditions. The manual ontology is difficult to update in a timely manner, leading to outdated knowledge and limiting the generalization ability of the intelligent diagnostic system.

[0004] Some studies have attempted to use large language models or end-to-end deep semantic networks to automatically extract and construct graphs from unstructured text and sensor data. While this method has certain advantages in semantic richness and knowledge scale, its core limitations include: the model training requires massive amounts of labeled samples, significant degradation after domain boundary migration, and the generation results lack controllable semantic granularity and process physical constraints, making it difficult to support root cause tracing and process constraint explanation, leading to the knowledge reasoning process becoming easily "black box." Furthermore, related systems are difficult to implement in high-reliability industrial manufacturing environments, and edge computing terminals cannot handle the model parameters and inference latency, nor do they meet the demands of fast-paced workshop conditions.

[0005] In practical applications, the knowledge of mold error sources is extremely dispersed, including time-series data from CNC program deviation logs, spatial model files from coordinate measuring machine reports, pulse signal sequences from EDM equipment, and handwritten natural language text in maintenance records. How to unify these multimodal, multidimensional, and highly context-dependent raw knowledge sources into a reasonable and reusable graph is a recognized challenge in the industry. Existing technologies struggle to achieve automatic fusion and global alignment between various heterogeneous knowledge sources, leading to knowledge silos and fragmented reasoning. More significantly, in scenarios involving new processes, new materials, new equipment, or frequent abnormal and unknown faults, traditional ontology and semantic representations cannot adapt and evolve online, resulting in extremely poor knowledge reusability and severely restricting the self-learning and generalization efficiency of intelligent diagnostic systems. Secondly, current systems are mostly constrained by the limitations of ontology structure transfer and equipment manufacturer interface standards, resulting in strong diagnostic targeting but weak transferability, making it difficult to meet the rapid knowledge response requirements of flexible manufacturing with its diverse product types, small batches, and high change rates.

[0006] Therefore, this field A novel method for organizing knowledge sources of mold errors, capable of handling multi-source heterogeneity, semantic variability, and loose structure, not only fully mines and integrates different types of physical signals and on-site textual experience, but also possesses adaptive knowledge structure evolution and efficient causal reasoning capabilities. This effectively breaks through the manual bottleneck of traditional knowledge ontology construction and improves the technological constraints of reasoning and interpretation. Transparency enables knowledge graphs to be highly reusable, low-intrusive, and sustainably self-evolving, providing strong technical support for intelligent diagnosis and digital knowledge management in the manufacturing of precision micro-connector molds. Summary of the Invention

[0007] This application provides a method for diagnosing error sources in the manufacturing process of precision micro connector molds, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0008] This application provides a method for diagnosing error sources in the manufacturing process of precision micro connector molds, specifically including: S1: Collect multi-source heterogeneous data during the manufacturing process of precision micro connector molds, including coordinate offset sequences, surface profile deviation cloud maps, abnormal discharge parameter records, and phenomenon description texts.

[0009] S2: Perform time-frequency joint entropy compression processing, curvature extreme point sampling and main direction projection processing, pulse energy and breakdown delay scatter distribution fitting processing, and entity recognition processing based on process dictionary on the multi-source heterogeneous data to generate error semantic fingerprints.

[0010] S3: Construct an initial anchor point set based on the pre-calibrated typical error case set, and uniformly map the error semantic fingerprint to a low-dimensional semantic fingerprint space, perform nearest neighbor matching operation, and generate fingerprint anchor point mapping relationship.

[0011] S4: Determine whether the matching degree value in the fingerprint anchor mapping relationship is lower than the preset evolution threshold. If the determination result is lower than the preset evolution threshold, trigger the local ontology evolution mechanism to generate an anchor extension cluster.

[0012] S5: Based on the fact that the new fingerprint in the anchor point extension cluster is included in the same semantic granularity level as the original anchor point, the process constraint relationship between the new fingerprint and the neighboring anchor points is automatically deduced, and condition rules are generated.

[0013] S6: Convert the real-time collected mold detection data into a real-time error semantic fingerprint, and perform a nearest neighbor search operation on the real-time error semantic fingerprint in the low-dimensional semantic fingerprint space to locate the target anchor point cluster most relevant to the current working condition.

[0014] S7: Perform a graph traversal operation along the evolved conditional rules within the target anchor cluster to generate attribution inference results containing the causal chain of error sources.

[0015] S8: Based on the attribution reasoning results, output specific error source diagnostic conclusions and update the initial anchor set to complete the incremental fusion closed loop of the knowledge graph.

[0016] The method for diagnosing error sources in the manufacturing process of precision micro connector molds provided in this application has the following beneficial effects: (1) To address the common problems of traditional knowledge graphs in precision manufacturing error diagnosis scenarios, such as strong reliance on artificial ontology design, static semantic modeling, and difficulty in cross-modal data fusion, this solution constructs an error semantic fingerprint system with measurable, reproducible, and comparable physical semantic features as its core. This achieves automated and process-consistent mapping from multi-source heterogeneous data to a unified semantic space. Compared to traditional methods that require pre-defining complex ontology structures or relying on large language models for text understanding, this solution abandons the end-to-end semantic embedding path and instead implements restricted entity recognition based on a mold process dictionary. It also combines physical feature extraction methods such as time-frequency joint entropy compression, differential geometric curvature sampling, and pulse energy distribution fitting to ensure that various error information is expressed in a standardized manner while retaining the key process significance. This mechanism significantly improves the accuracy and stability of knowledge extraction, especially when dealing with unstructured text such as handwritten notes. It avoids the misidentification problem caused by domain migration bias in general semantic models, while greatly reducing computational overhead. This allows the entire knowledge organization process to be completed on edge devices without cloud support, effectively ensuring the real-time performance and deployment flexibility of workshop applications.

[0017] (2) Furthermore, this scheme innovatively introduces a dynamic ontology anchoring mechanism based on a set of anchor points for typical error cases, breaking through the technical bottleneck that static graph structures are difficult to adapt to new failure modes. Unlike conventional approaches where explicit expansion of new categories may cause graph structure oscillations or logical breakage, this scheme adopts an "anchor point expansion cluster" strategy. Without adding new ontology nodes or relation types, it aggregates new fingerprints with low matching degrees and nearby existing anchor points into the same semantic granularity unit, and automatically derives lightweight process constraint rules between them and neighboring anchor points, such as temporal coupling, parameter association, or causal exclusion relationships. This local evolution approach maintains the stability of the overall graph topology and endows the system with the ability to continuously learn and adapt to changes in the field, solving the contradiction between knowledge sparsity in the cold start stage and incremental data accumulation in the later stage. The attribution analysis module utilizes this structure to quickly locate relevant anchor clusters by using semantic fingerprint spatial nearest neighbor retrieval, and then performs lightweight graph traversal along the constraint chain evolved within the cluster. This enables high-precision, low-latency causal attribution, and the reasoning path has clear process explanations, which is significantly better than the untraceability of black-box deep learning models, enhancing engineers' trust in and operability of the diagnostic results.

[0018] In summary, this solution constructs a closed-loop knowledge organization and reasoning system for diagnosing error sources in precision micro-connector mold manufacturing. Its core advantages lie in achieving physical semantic-driven knowledge acquisition, dynamic adaptive ontology structure, and lightweight and interpretable reasoning process. By completely avoiding dependence on external large models, cross-domain ontology, and complex parameter tuning mechanisms, the system exhibits high domain specificity, ease of deployment, and robust operation. It is particularly suitable for identifying and locating the root causes of frequent, minor but critical errors in high-precision mold processing environments, providing reliable technical support for improving mold life management, process optimization response speed, and intelligent manufacturing levels. Attached Figure Description

[0019] Figure 1 This is the main flowchart of a method for diagnosing error sources in the manufacturing process of precision micro connector molds.

[0020] Figure 2 This is a sub-flowchart of a method for diagnosing error sources in the manufacturing process of precision micro-connector molds.

[0021] Figure 3 This is another sub-flowchart of a method for diagnosing error sources in the manufacturing process of precision micro-connector molds. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention... Components and arrangements of specific examples are described. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplicity and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown, this application provides a method for diagnosing error sources in the manufacturing process of precision micro connector molds, specifically including: S1: Collect multi-source heterogeneous data during the manufacturing process of precision micro connector molds, including coordinate offset sequences, surface profile deviation cloud maps, abnormal discharge parameter records, and phenomenon description texts.

[0025] S2: Perform time-frequency joint entropy compression processing, curvature extreme point sampling and main direction projection processing, pulse energy and breakdown delay scatter distribution fitting processing, and entity recognition processing based on process dictionary on the multi-source heterogeneous data to generate error semantic fingerprints.

[0026] S3: Construct an initial anchor point set based on the pre-calibrated typical error case set, and uniformly map the error semantic fingerprint to a low-dimensional semantic fingerprint space, perform nearest neighbor matching operation, and generate fingerprint anchor point mapping relationship.

[0027] S4: Determine whether the matching degree value in the fingerprint anchor mapping relationship is lower than the preset evolution threshold. If the determination result is lower than the preset evolution threshold, trigger the local ontology evolution mechanism to generate an anchor extension cluster.

[0028] S5: Based on the fact that the new fingerprint in the anchor point extension cluster is included in the same semantic granularity level as the original anchor point, the process constraint relationship between the new fingerprint and the neighboring anchor points is automatically deduced, and condition rules are generated.

[0029] S6: Convert the real-time collected mold detection data into a real-time error semantic fingerprint, and perform a nearest neighbor search operation on the real-time error semantic fingerprint in the low-dimensional semantic fingerprint space to locate the target anchor point cluster most relevant to the current working condition.

[0030] S7: Perform a graph traversal operation along the evolved conditional rules within the target anchor cluster to generate attribution inference results containing the causal chain of error sources.

[0031] S8: Based on the attribution reasoning results, output specific error source diagnostic conclusions and update the initial anchor set to complete the incremental fusion closed loop of the knowledge graph.

[0032] Step S1: Collect multi-source heterogeneous data during the manufacturing process of precision micro-connector molds, including coordinate offset sequences, surface profile deviation cloud maps, discharge parameter anomaly records, and phenomenon description texts. Specifically, this includes: S1.1: Real-time monitoring of the CNC machining center is performed based on the industrial Ethernet protocol to obtain the raw data stream of CNC machining logs containing timestamps, axis command positions, and actual feedback positions. The raw data stream of CNC machining logs is then processed to extract coordinate offset sequences to generate a coordinate offset sequence dataset that characterizes the motion accuracy of the machine tool.

[0033] The real-time monitoring of a CNC machining center based on the Industrial Ethernet protocol targets a multi-axis CNC machining equipment equipped with an open data interface. The transmitted signal includes a real-time data stream containing timestamps, axis command positions, and actual axis feedback positions. The protocol parsing module is invoked on the monitored data stream to extract the command position and actual position fields of each axis according to the frame period, generating a raw position pair sequence. The sequence performs axial interpolation calculations to obtain the instantaneous coordinate offset of each axis at the corresponding timestamp, and constructs a time series structure using these offsets. Outlier removal is performed on the time series, using a set threshold detection rule to eliminate noisy data points, resulting in an interference-free offset sequence. Synchronous interpolation is then applied to the interference-free offset sequence to ensure that the offset data of each axis corresponds under a unified time reference, generating a multi-axis synchronous offset matrix. This matrix is ​​then segmented according to machine tool machining program segments to form a segmented coordinate offset sequence dataset. Through the above chained processing method, the raw log data obtained from industrial Ethernet monitoring is transformed into a standardized coordinate offset sequence that can be directly used for motion accuracy characterization, achieving the expected technical effect of accurately reflecting equipment position control errors in subsequent feature extraction.

[0034] For example, in a precision micro-connector mold manufacturing application, the CNC machining center used is a four-axis high-precision machine tool. The industrial Ethernet monitoring module is set to a sampling period of 1ms. The data frame structure includes a timestamp (accuracy 0.1ms), axis command position (unit μm), and actual feedback position. The parsing module maps the four-axis command position and actual position of each frame into position pairs. For example, if the X-axis command position is 2500μm and the actual feedback position is 2498μm, then the offset is 2μm. The offset is calculated using the formula: in, For position values, For instruction location, For feedback location, The offset is set as follows: An anomaly rejection threshold of 10 μm is set, and all instantaneous offsets exceeding this value are rejected. The resulting sequence undergoes linear synchronous interpolation to arrange the data for each axis according to a unified time reference. In this embodiment, the generated multi-axis synchronous offset matrix is ​​divided according to each tool call segment in the machine tool machining program, forming a coordinate offset sequence dataset containing several segments. This dataset can significantly improve the accuracy of machining trajectory stability feature extraction during subsequent time-frequency joint entropy feature compression.

[0035] S1.2: Establish a communication connection with the coordinate measuring machine based on the standard data exchange format interface, obtain the original file of the coordinate measuring report containing the deviation values ​​between the spatial coordinates of the measuring points and the theoretical model, and perform surface profile deviation cloud map reconstruction processing on the original file of the coordinate measuring report to generate a surface profile deviation cloud map data matrix characterizing the geometric deformation of the mold cavity.

[0036] When a coordinate measuring machine (CMM) performs data acquisition and processing tasks, an embedded control module loads a communication interface configuration file conforming to a standard data exchange format. This initialization includes connection parameters such as host address, port number, and handshake protocol version, establishing a stable data link at the physical layer to ensure real-time, lossless transmission of measurement point coordinates and deviation values. Based on this established communication link, the CMM's raw data output module is invoked to read the spatial coordinate vector of each measurement point and its deviation value from the theoretical model. The acquired raw file is stored in binary form in a cache queue to ensure data block integrity and processing continuity. A file parsing engine is then invoked to perform syntax tree parsing on the raw data blocks in the cache queue. Following a predefined CMM measurement report format, each field is mapped to the measurement point coordinate matrix and deviation matrix, completing the data structuring transformation and generating an intermediate dataset containing multi-point spatial location information and corresponding deviation values. The deviation matrix is ​​input into the surface reconstruction algorithm module, employing a combined processing strategy of mesh interpolation and normal vector estimation to map discrete measurement point coordinates and deviation values ​​to a continuous surface data domain. Curvature smoothing and boundary condition constraints are then applied to form a high-resolution deviation surface description. The deviation surface is encoded into a surface profile deviation cloud map according to a specified color mapping scheme. Using matrix storage, the deviation value corresponding to each sampling point of the surface is converted into the grayscale or color value of the cloud map pixels, outputting a surface profile deviation cloud map data matrix that characterizes the geometric deformation of the mold cavity. Through the aforementioned data exchange interface and surface reconstruction processing method, the output of the previous measuring machine step is transformed into a precise geometric deformation index that can be used for subsequent error semantic fingerprint extraction, achieving a high-precision digital expression of the mold cavity shape error.

[0037] For example, in the scenario of inspecting precision micro-connector molds, a coordinate measuring machine (CMM) supporting STEP format and XML data embedding is selected. The TCP / IP interface is configured with port number 5025, the handshake protocol uses version V2.1, and the initial refresh frequency of the communication link is set to 5Hz to ensure the acquisition of 5 complete sets of measurement point data per second. The CMM performs 500 evenly distributed measurement point sampling on the mold cavity, obtaining the position value of each measurement point in the three-dimensional Cartesian coordinate system and its deviation from the theoretical model, with the deviation ranging from -0.012mm to 0.010mm. This data is mapped into a 500×3 coordinate matrix and a 500×1 deviation matrix using a parsing engine. Bilinear interpolation and normal vector estimation are used to reconstruct the surface, and curvature smoothing is performed on the surface with a smoothing coefficient set to 0.85 to reduce the influence of local measurement noise. After surface reconstruction, color mapping is used to map the values ​​within the deviation range to a gradient band from blue to red to reflect both contraction and expansion deformation trends. The results are stored as a deviation contour map data matrix with a resolution of 1024×1024 pixels. The output deviation contour map data matrix can significantly improve the extraction accuracy of deformation modes in subsequent differential geometric curvature extreme point sampling processing and provide stable geometric reference data in the construction of error source causal chains.

[0038] S1.3: Based on a dedicated serial communication protocol, the status of the electrical discharge machining equipment is polled to obtain the original records of electrical discharge parameters including pulse width, peak current and gap voltage waveforms. Abnormal working condition segments are extracted from the original records of electrical discharge parameters to generate a set of abnormal records of electrical discharge parameters that characterize the stability of material removal.

[0039] The electrical discharge machining (EDM) equipment is polled based on a dedicated serial communication protocol. The system utilizes the real-time signal acquisition interface built into the equipment controller, setting the polling period and sampling resolution to obtain raw records of EDM discharge parameters, including pulse width, peak current, and gap voltage waveforms. Through serial communication checksums and frame format parsing, the acquired signal data stream is reassembled into a continuous processing state sequence in timestamp order. During this reassembly process, data integrity checks are performed to eliminate fragmented or duplicate data caused by communication errors. Based on the real-time change curves of pulse width and peak current in the acquired sequence, a threshold judgment strategy is used to detect pulse anomalies exceeding the process tolerance range, generating a preliminary label index for potential abnormal operating conditions. For the gap voltage waveform data, a waveform parsing module is invoked to perform window function segmentation analysis, extracting the mean, variance, and fluctuation amplitude of each segment. Segments with fluctuation amplitudes exceeding a preset upper limit are marked as abnormal waveform segments. Based on the intersection of the preliminary pulse anomaly label index and the abnormal waveform segments, an association matching operation is performed to filter out operating condition segments that simultaneously meet the abnormal characteristics of pulse width, peak current, and gap voltage. Continuous and interconnected segments are merged to form a set of abnormal operating condition segments. The set of abnormal operating condition segments is encapsulated into a standardized set of abnormal EDM discharge parameter records according to processing batches and time windows to characterize the dynamic changes in material removal stability and unstable states during processing. Through the above chain-like processing method, the raw discharge parameter data collected by polling is transformed into an abnormal record set that can be used for subsequent error semantic fingerprint extraction, thereby achieving a quantitative expression of the material removal stability characteristics of the EDM process.

[0040] For example, in the electrical discharge machining (EDM) process for manufacturing precision micro-connector molds, a dedicated serial communication protocol with a baud rate of 115200 is used to acquire four-channel parameters with the machining equipment. The polling period is set to 200ms, and the sampling resolution is 12-bit ADC. The average pulse width is set to a process tolerance of ±0.5μs, the peak current tolerance is set to ±0.2A, and the gap voltage fluctuation amplitude tolerance is set to ±2V. During the detection process, records with pulse widths exceeding the process tolerances and records with peak currents exceeding the tolerances are marked as abnormal. The variance of the gap voltage waveform is calculated using waveform analysis with a sliding window length of 256 points, using the following formula: in, The sampling gap voltage, The average voltage within the window. The window length is defined as follows: Window segments with fluctuation amplitudes exceeding the tolerance limit are calculated as abnormal waveform segments. These segments are then intersected with pulse width and peak current anomaly labels to obtain 17 abnormal operating condition segments, covering a maximum single-piece processing time of 3.4 seconds. These segments are encapsulated according to batch number and processing sequence. The resulting abnormal record set contains multi-dimensional features such as pulse width mean deviation, peak current overshoot, and gap voltage fluctuation amplitude. This significantly improves feature consistency and interpretability when used for subsequent material response fingerprint extraction, greatly enhancing the system's diagnostic ability to identify material removal stability.

[0041] S1.4: Based on optical character recognition technology and natural language preprocessing rules, the maintenance work order file is digitally scanned to obtain the original text of the maintenance engineer's handwritten fault notes, which contain descriptions of fault phenomena and records of maintenance measures. The original text of the maintenance engineer's handwritten fault notes is then subjected to unstructured corpus cleaning to generate a corpus of phenomenon description texts representing human experience judgment.

[0042] S1.5: Linearly interpolate the four types of data to a unified time axis according to their respective timestamps. Based on the multi-source data time alignment mechanism, perform spatiotemporal correlation fusion processing on the coordinate offset sequence dataset, the surface contour deviation cloud map data matrix, the electrical discharge parameter anomaly record set, and the phenomenon description text corpus to generate a multi-source heterogeneous data package for subsequent error semantic fingerprint extraction.

[0043] Step S2: The multi-source heterogeneous data undergoes time-frequency joint entropy compression, curvature extreme point sampling and principal direction projection, pulse energy and breakdown delay scatter point distribution fitting, and entity recognition based on a process dictionary to generate an error semantic fingerprint. Specifically, this includes: S2.1: Perform time-frequency joint entropy feature compression on the coordinate offset sequence in the CNC machining log, obtain the time-frequency energy distribution matrix using short-time Fourier transform and calculate the Shannon entropy value to generate a numerical vector fingerprint.

[0044] An offline cache is established for the coordinate offset sequence dataset extracted in step S1.1. The short-time Fourier transform operation module is called to divide each offset sequence into consecutive overlapping fixed-length data frames. The Hanning window function is used to perform window function weighting on each frame to alleviate the spectral leakage effect.

[0045] Perform a Fast Fourier Transform operation on each data frame weighted by a window function to generate the amplitude spectrum and phase spectrum of the data frame, and stack the amplitude spectra in time series to form a time-frequency energy distribution matrix.

[0046] The time-frequency energy distribution matrix is ​​normalized to map the energy values ​​of different frames and frequency bands to a uniform range, so as to ensure consistency in subsequent statistical analysis.

[0047] The Shannon entropy calculation module is invoked to calculate the entropy value for each frame energy vector of the time-frequency energy distribution matrix using the following formula: in, For the normalized first The relative energy percentage of the frequency band. This is the Shannon entropy value for the corresponding frame.

[0048] The Shannon entropy values ​​of each frame are arranged in time series to construct a time-series entropy curve that reflects the motion accuracy of the machine tool. Multi-scale smoothing filtering is then applied to this curve to eliminate the influence of random noise.

[0049] The smoothed entropy curve is divided into several segments according to a preset time window. Statistical indicators such as mean, standard deviation, and kurtosis of each segment are extracted and combined to generate a numerical vector fingerprint with fixed length and consistent dimensions.

[0050] Through the above processing method, the coordinate offset sequence of the previous step is transformed into a numerical vector fingerprint with time-frequency domain statistical properties, thereby realizing the quantitative feature expression of the stability of the processing trajectory.

[0051] For example, in the CNC machining process of precision micro-connector mold manufacturing, when the length of the acquired coordinate offset sequence is 1024 points and the sampling frequency is 200Hz, the short-time Fourier transform window length is set to 256 points and the frame shift is 128 points, and the Hanning window function parameters are taken in unnormalized form. After FFT transformation, the energy spectrum of each frame with 128 frequency bands is obtained, and normalization processing makes the energy values ​​of each frequency band in the range [0,1]. In the Shannon entropy calculation, assuming that the energy ratio of the first frequency band of the normalized energy vector is 0.05, then in the above formula, p(f1)=0.05, and the ln function is calculated using the natural logarithm. After all frequency bands p(fi) are calculated, the single-frame entropy value is approximately 4.12. The entropy curve is obtained by arranging all frame entropy values, and a five-point moving average smoothing is performed on it, dividing it into ten time windows. The mean, standard deviation, kurtosis, etc. of each segment are calculated, forming a 30-dimensional numerical vector fingerprint. When the machining center is operating stably, the fluctuation range of each field of this fingerprint is small. When trajectory oscillations or reverse gap anomalies are present, the standard deviation and kurtosis are significantly increased, verifying the sensitivity and practicality of this fingerprint in diagnosing the stability of the processing trajectory.

[0052] S2.2: Based on the original measurement object corresponding to the numerical vector fingerprint, perform differential geometric curvature extreme point sampling and principal direction projection processing on the surface profile deviation cloud map in the coordinate measuring machine report. Generate a geometric fingerprint by calculating the Gaussian curvature extreme point set and performing orthogonal projection in the principal component direction of the normal vector.

[0053] A normal vector field reconstruction operation is performed on the input surface profile deviation cloud map data matrix. Utilizing the correspondence between the known spatial coordinates of the measurement points and the surface of the theoretical model, the normal vector direction and radius of curvature at each measurement point are solved through local polynomial fitting. The Gaussian curvature calculation formula is applied to the reconstructed normal vector field and curvature field, with the curvature value at each measurement point used as an element of the two-dimensional curvature matrix. The formula is as follows: in, The value is the Gaussian curvature. With primary curvature one, The principal curvature is two.

[0054] An extremum point extraction operation is performed on the two-dimensional curvature matrix. Local maximum and minimum values ​​are selected as candidate extremum points by setting a curvature threshold. The spatial coordinates of these points are collected to form a preliminary extremum point set. Density clustering analysis is then performed on this preliminary set to remove isolated points and retain stable extremum point clusters that represent the overall deformation trend, forming a Gaussian curvature extremum point set. Principal component analysis (PCA) is then performed on this Gaussian curvature extremum point set. The principal direction is determined by constructing the covariance matrix and solving for the eigenvectors, as shown in the following formula: in Let covariance matrix be the variance matrix. It is the normal vector. The mean of the normal vector. The number of extreme points is given. Based on the obtained principal direction eigenvector, orthogonal projection is performed on all extreme points along this direction to obtain a one-dimensional projected coordinate set. The statistical characteristics of the projected distribution are then calculated, including indicators such as maximum displacement, mean shift, and standard deviation. The spatial location data and projection statistical characteristics are encapsulated into a structured geometric fingerprint data object, which serves as the output characterizing the deformation mode of the mold surface.

[0055] By sampling the extreme points of differential geometric curvature and projecting them onto the principal direction, the numerical vector fingerprint from the previous step is associated with specific deformation features, thereby transforming the surface profile deviation cloud map into a reasonable geometric fingerprint and improving the stability and interpretability of error pattern recognition.

[0056] For example, in the coordinate measuring machine (CMM) report of a precision micro-connector mold, the surface profile deviation cloud map data matrix has 1024 measuring points, with each point's coordinate accuracy within 0.001 mm. The normal vector is obtained using local second-order polynomial fitting, and the principal curvature values ​​k1 and k2 are calculated. The principal curvature values ​​at typical points are 0.025 and k2, respectively. 0.018, substitute into the formula The Gaussian curvature value is obtained. 0.00045. With a curvature threshold set to ±0.0004, 85 extreme points were extracted. After processing using the DBSCAN clustering algorithm (neighborhood radius 0.2 mm, minimum number of points 5), 4 stable extreme point clusters were retained. A covariance matrix was constructed for the normal vectors within each cluster, and the first eigenvector was obtained through principal component analysis, yielding the principal direction unit vector [0.87, 0.46, 0.15]. Orthogonal projection was performed along this principal direction to obtain the projected coordinate distribution. The maximum displacement was 0.012 mm, the mean offset was 0.005 mm, and the standard deviation was 0.002 mm. The extreme point cluster coordinates and projection statistical characteristics were encapsulated into a geometric fingerprint, which was verified to be useful in distinguishing between two typical errors: local cavity collapse and overall draft angle deviation. The reliability of the deformation mode identification results was significantly improved.

[0057] S2.3: Based on the process associated with the geometric fingerprint, the abnormal records of electrical discharge parameters of electrical discharge machining are fitted with a two-dimensional scatter distribution of pulse energy and breakdown delay. A two-dimensional scatter plot of pulse energy and breakdown delay is constructed, and the slope of the edge envelope and the dispersion statistics are extracted to generate a material response fingerprint.

[0058] Based on the process associated with the geometric fingerprint, when performing two-dimensional scatter distribution fitting processing on the abnormal discharge parameter record set of EDM, the input conditions are abnormal operating condition data segments containing pulse width, peak current, and gap voltage waveforms, and timestamp registration and noise removal have been completed. The waveform analysis module is then invoked to combine the pulse width and peak current to calculate the energy value of a single discharge pulse. The energy calculation formula is as follows: in, This is the peak current. This represents the peak value of the gap voltage. The pulse width is used. The energy value is paired with the corresponding breakdown delay to generate the original two-dimensional data point set. A curve fitting engine is invoked to perform least-squares fitting of the edge envelope on the two-dimensional data point set. The upper and lower envelopes are modeled as continuous function curves, and the slope of the local tangent line for each segment of the fitted curve is calculated using the slope statistical formula: in, The energy difference This represents the delay difference. The dispersion analysis module is invoked to calculate the variance of the horizontal and vertical distributions of the two-dimensional point set. The dispersion index uses the formula: in, For a single delayed data point, This is the average delay. The number of data points is specified. The calculated slope and dispersion results are normalized to ensure comparability within the material response fingerprint space. Normalized slope and dispersion indices are encapsulated to generate a material response fingerprint characterizing the electrode and workpiece material response properties. By fitting the edge envelope and extracting the slope and dispersion, the geometric fingerprint from the previous step is mapped to data indices reflecting material removal stability and discharge characteristics, achieving a structured characterization of the material response mode.

[0059] For example, in an EDM machining operation of a precision micro-connector mold, the peak pulse current is configured as 5A, the peak gap voltage as 120V, the pulse width is recorded between 0.8ms and 1.2ms, and the breakdown delay is distributed between 0.5ms and 1.0ms. The waveform analysis module calculates the energy value to be between 480μJ and 720μJ, and generates a two-dimensional scatter plot with the corresponding delay. The average slope of the upper envelope fitting curve is calculated to be 480, representing the rate of energy change with increasing delay. The dispersion analysis result shows a lateral distribution variance of 1.6 × 10⁻⁶ for the delay data. - The normalized slope index is 0.85, and the dispersion index is 0.12, corresponding to the material response fingerprint's position in the low-dimensional space being close to the "electrode loss-type discharge mode" anchor cluster. In subsequent attribution analysis, this fingerprint significantly improves the detection rate of EDM collapse-type faults, and in field verification, it has achieved a substantial improvement in processing stability.

[0060] S2.4: Based on the fault phenomena involved in the material response fingerprint, the phenomenon description text in the handwritten fault notes of the maintenance engineer is processed by restricted entity recognition based on the mold process dictionary. The predefined mold process dictionary containing terms such as locating pin wear and local cavity ablation is used to perform string matching and entity extraction to generate a standardized error phrase fingerprint.

[0061] S2.5: Based on the standardized error phrase fingerprint and the other three types of fingerprint data, perform error semantic fingerprinting operation to encapsulate the numerical vector fingerprint, geometric fingerprint, material response fingerprint and standardized error phrase fingerprint into a unified data structure to generate the error semantic fingerprint.

[0062] like Figure 2 As shown, step S3 involves constructing an initial anchor point set based on a pre-calibrated set of typical error cases, mapping the error semantic fingerprints to a low-dimensional semantic fingerprint space, performing nearest neighbor matching, and generating fingerprint anchor point mapping relationships. Specifically, this includes: S3.1: Construct an initial anchor point set based on a pre-calibrated set of typical error cases, and perform dimension alignment processing on the error semantic fingerprint set containing numerical vector fingerprints, geometric fingerprints, material response fingerprints, and standardized error phrase fingerprints to generate a fingerprint data package to be mapped with a unified input interface.

[0063] S3.2: Based on the high-dimensional feature matrix in the fingerprint data packet to be mapped, the principal component analysis algorithm is used to perform linear projection transformation processing to compress the high-dimensional feature data into a low-dimensional semantic fingerprint space coordinate system that is isomorphic to the initial anchor point set, so as to generate a real-time projected fingerprint vector group with spatial comparability.

[0064] Based on the high-dimensional feature matrix in the fingerprint data packet to be mapped, the feature centering processing module is called to perform a mean-zeroing operation on each column of the input matrix, generating a standardized high-dimensional feature matrix that maintains zero mean and whose covariance matrix is ​​computable.

[0065] The standardized matrix is ​​input into the covariance calculation unit. Based on the layout of horizontal samples as rows and vertical features as columns, the covariance matrix is ​​calculated to characterize the linear correlation strength between the features.

[0066] The eigenvalue decomposition algorithm is called on the covariance matrix data to solve for the corresponding set of eigenvalues ​​and eigenvectors. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors in the corresponding order are retained to establish the principal component priority sequence.

[0067] The contribution rate accumulation calculation module is invoked based on the preset cumulative contribution rate threshold to calculate the proportion of the total variance of the top few feature values ​​and filter out the number of principal components that can cover the predetermined contribution rate.

[0068] The selected feature vector set is combined column-wise into a mapping transformation matrix, and the high-dimensional feature matrix is ​​multiplied on the left by the mapping transformation matrix to perform a linear projection transformation, compressing the high-dimensional features into a low-dimensional semantic fingerprint space coordinate system.

[0069] By using the above processing method, the result of the previous step is transformed into a spatially comparable real-time projected fingerprint vector group, realizing the expected technical effect of measuring the distance between different modal fingerprints and the initial anchor point set in the same space.

[0070] S3.3: Based on the real-time projected fingerprint vector group and the initial anchor point set stored in the knowledge base, the Euclidean distance metric algorithm is called to perform point-by-point similarity calculation, and a similarity score matrix containing distance values ​​is constructed to quantitatively represent the spatial proximity between the current real-time operating conditions and historical typical error cases.

[0071] Based on the real-time projected fingerprint vector group and the initial anchor point set stored in the knowledge base, a pre-indexed structure containing anchor point spatial coordinate data is loaded as the calculation input condition.

[0072] The core function of the Euclidean distance metric algorithm is called to perform single-point difference vector calculation on each high-dimensional element in the real-time projected fingerprint vector group and its corresponding spatial coordinates in the initial anchor point set.

[0073] The elements of the difference vector are squared to eliminate directionality and preserve magnitude contribution, forming a squared difference matrix.

[0074] An accumulator is used to sum the squares along the rows of the squared difference matrix to obtain the scalar sum of squares between each fingerprint and each anchor point.

[0075] Perform square root operation on the sum of squares and scalars to generate Euclidean distance values ​​from each fingerprint to each anchor point, and construct a similarity score matrix containing the distance values.

[0076] The Euclidean distance calculation process is defined using the following mathematical formula: in, Represents Euclidean distance. Real-time projected fingerprint vector index For the initial anchor index, Indexed by vector dimension, For the dimensions of the projection space, To project fingerprint elements in real time, These are the initial anchor point coordinates.

[0077] Through the above point-by-point similarity calculation operation, the real-time projected fingerprint vector group from the previous step is transformed into a similarity scoring matrix that quantitatively represents the spatial proximity between the current working condition and historical typical error cases, thereby realizing the construction of a standardized distance metric that can be used for subsequent nearest neighbor matching and screening.

[0078] For example, in a precision micro-connector mold manufacturing error source diagnosis system, a real-time projected fingerprint vector group contains four fingerprints, each fingerprint having a projection space dimension of 6. The initial anchor point set contains 10 anchor points, all stored in the coordinate table of the knowledge base. When performing Euclidean distance calculation, the input conditions are the 6-dimensional feature values ​​of each fingerprint and the 6-dimensional coordinate values ​​of each anchor point. The process sequentially calculates the difference, square, sum, and square root operations according to the formula. For example, for the first real-time fingerprint x1 (0.012, 0.004, 0.006, 0.009, 0.011, 0.005) and the third initial anchor point a3 (0.010, 0.005, 0.006, 0.010, 0.012, 0.006), the difference vector is (-0.002, 0.001, 0.000, 0.001, 0.001), the squared difference is (4e-6, 1e-6, 0, 1e-6, 1e-6, 1e-6), the sum of squares is 8e-6, and the square root is 0.002828. This distance value is filled into the corresponding position in the similarity scoring matrix, and the remaining fingerprints and anchor points are calculated using the same algorithm. In this scenario, the scoring matrix is ​​used to subsequently filter the initial set of anchor points with the smallest Euclidean distance, ensuring that the spatial proximity of the matching is significantly improved, and achieving high-precision positioning for error source identification.

[0079] S3.4: Based on the similarity score matrix, perform nearest neighbor matching and filtering operation, select the K initial anchor points with the smallest Euclidean distance metric to form a candidate association set, and perform topological connectivity verification on the initial anchor points in the candidate association set to generate fingerprint anchor point mapping relationship.

[0080] It should be noted that the matching degree value is used to quantify the similarity between the new fingerprint and the initial anchor point, and it is converted based on the Euclidean distance calculated in the previous steps. The smaller the Euclidean distance, the closer the two are in the low-dimensional semantic fingerprint space, and the closer the matching degree value is to 1; the larger the Euclidean distance, the closer the matching degree value is to 0. The system compares the matching degree value with a preset evolution threshold. If the matching degree is lower than the threshold, a local ontology evolution mechanism is triggered.

[0081] Based on the similarity score matrix generated in the previous steps, the nearest neighbor matching and filtering algorithm module is called to sort the Euclidean distance values ​​between the new fingerprints corresponding to each row of the matrix and each initial anchor point in ascending order, forming a distance sorted list.

[0082] Extract the top K anchor points whose Euclidean distance metric values ​​are in the minimum interval from the distance sorting list, construct a candidate association set data structure, and map and register the unique identifier of each anchor point in the set with its corresponding new fingerprint index.

[0083] Based on the spatial location index of each anchor point in the candidate association set, its direct adjacent units are retrieved in the stored initial anchor point set topology table, and the retrieved adjacency relationship information is appended to the topology mapping description of the candidate set.

[0084] Connectivity verification is performed on the candidate set with adjacency relationships attached. Graph traversal is used to detect whether there is a path coverage relationship between all selected anchor points, and the connected component values ​​are calculated to determine the topological integrity within the set.

[0085] In the connectivity verification results, isolated anchor points or anchor points with missing topological relationships are removed. The retained anchor points and their corresponding new fingerprint indexes are encapsulated into a fingerprint anchor mapping relationship data packet in a bidirectional reference manner to realize the association and matching output between the new fingerprint and existing knowledge.

[0086] This screening and verification process transforms the similarity score matrix from the previous step into a fingerprint anchor mapping relationship that satisfies both spatial similarity and topological connectivity constraints, thereby achieving accurate association and highly reliable matching of the candidate anchor set.

[0087] For example, in a diagnostic scenario for manufacturing precision micro-connector molds, the similarity scoring matrix contains Euclidean distance metrics for 128 new fingerprints and 500 initial anchor points. The dimension of the new fingerprint vector is 12. The K value of the nearest neighbor matching algorithm is set to 5, and a candidate set is constructed by sorting and extracting the top 5 anchor points with the smallest Euclidean distance. The spatial index corresponding to each anchor point has 2 to 4 adjacent units in the topology table of the initial anchor point set. The connectivity check adopts a breadth-first traversal method, with a traversal depth limited to 3 layers. For a new fingerprint, 4 out of the 5 anchor points in its candidate set form a single connected component, and the other anchor point is an isolated node, which is removed after detection. A bidirectional mapping relationship is established between the retained 4 anchor points and the new fingerprint. In the test verification, this connectivity check and screening matching process can significantly improve the stability of error source identification, and the output fingerprint anchor point mapping relationship has efficient linkage and interpretability in subsequent attribution analysis.

[0088] S3.5: Based on the matching degree value in the fingerprint anchor mapping relationship and the preset evolution threshold, a logical comparison operation is performed to generate an evolution trigger signal that represents whether the ontology structure needs to be updated, and the evolution trigger signal is used as the execution basis for the subsequent local ontology evolution mechanism to complete the pre-judgment process of knowledge graph incremental fusion.

[0089] For the matching degree value in the fingerprint anchor mapping relationship, the evolution threshold constant stored in the system configuration file is loaded to ensure that the dimension of the input data is consistent with the dimension of the threshold to avoid logical comparison failure.

[0090] The comparison operation module is called to perform double-precision difference calculation on the matching degree value and the evolution threshold and obtain the absolute value of the difference to eliminate the influence of positive and negative direction differences on subsequent judgments.

[0091] A Boolean condition expression is constructed based on the absolute value of the difference. The condition result is parsed by the hard threshold decision function and the evolution trigger signal status bit is returned. A high level in this status bit indicates that the body structure needs to be updated, and a low level indicates that the body structure does not need to be updated.

[0092] The evolution trigger signal status bits are encapsulated into a judgment message containing a trigger timestamp, trigger source fingerprint identifier, and matching degree difference index, so that the subsequent local ontology evolution mechanism has complete context information.

[0093] By standardizing the enqueueing operation of the judgment message, it is synchronously written into the dynamic evolution task queue to form a pre-judgment result dataset for knowledge graph incremental fusion execution.

[0094] Through the above processing method, the fingerprint anchor mapping result of the previous step is transformed into an evolution trigger signal that can be directly parsed by the local ontology evolution module, so as to realize a fast, accurate and traceable pre-decision on whether the knowledge graph structure should be updated.

[0095] For example, in the process of identifying error sources in the manufacturing of a precision micro-connector mold, the matching degree between a real-time projected fingerprint vector and the initial anchor point set is 0.62, and the system's configured evolution threshold is constant at 0.75. The evolution determination module performs a difference calculation on the matching degree value and the evolution threshold, obtaining a difference of -0.13, and takes the absolute value of 0.13. A logical comparison operation is used to determine that the condition of the matching degree value being less than the evolution threshold is met, generating a high-level trigger signal. The trigger message records the matching degree value, threshold, difference, trigger time, and fingerprint ID. In this embodiment, the absolute value of the difference meets the preset trigger condition, and the trigger signal serves as the input to the local ontology evolution mechanism, generating task units in the dynamic evolution task queue. Through this processing, the evolution module introduces the new fingerprint to form an anchor point extension cluster during subsequent execution, and incrementally integrates new knowledge in the graph, achieving effective adaptation to new error patterns with semantic drift, significantly improving the generalization ability and self-learning efficiency of the diagnostic system.

[0096] like Figure 3 As shown, step S4 involves determining whether the matching degree value in the fingerprint anchor mapping relationship is lower than a preset evolution threshold. If the determination result is lower than the preset evolution threshold, a local ontology evolution mechanism is triggered to generate an anchor point expansion cluster. Specifically, this includes: S4.1: Based on the fingerprint anchor mapping relationship generated in the previous steps, extract the matching degree value that represents the similarity between the new fingerprint and the initial anchor set, and perform a logical comparison operation between the matching degree value and the evolution threshold preset in the system configuration to generate an evolution trigger signal.

[0097] Based on the fingerprint anchor point mapping relationship data structure output by the previous steps, the matching degree parsing algorithm is called to read the spatial distance metric value of each "new fingerprint-anchor point" association record and convert it into a matching degree value. The matching degree value comes from the calculation result of the inverse function of normalized Euclidean distance.

[0098] The extracted matching degree values ​​are stored in the matching degree sequence buffer according to the record index order. The system configuration module is called to read the preset evolution threshold, which is a fixed precision limit constant derived from the statistical analysis of the historical evolution process.

[0099] For each matching degree value in the matching degree sequence buffer, a logical comparison operation is performed. Relational operators are used to determine whether the matching degree is less than a preset evolution threshold. The comparison result is calculated using the following formula: in, As a difference judgment index, This represents the current matching score. This is a preset evolution threshold.

[0100] The logical comparison result is converted into a Boolean flag. If the flag is true, the trigger signal generator module is called to generate an evolution trigger signal object containing a new fingerprint identifier, a matching degree value, and a threshold information based on the matching record index and the corresponding matching degree value.

[0101] Write the evolution trigger signal object into the ontology evolution queue for consumption during the execution of subsequent local ontology evolution mechanisms.

[0102] Through the above-mentioned matching degree extraction, threshold reading, logical comparison and trigger signal generation processing methods, the mapping relationship results of the previous step are transformed into difference indicators and evolution trigger objects that can be judged by the local ontology evolution mechanism, so as to realize the quantitative control of the initiation conditions of new knowledge fusion.

[0103] S4.2: In response to the determination result in the evolution trigger signal indicating that the matching degree value is lower than the preset evolution threshold, the cluster generation module in the dynamic ontology anchoring engine is invoked to initialize a temporary semantic space region to be expanded, using the unmatched error semantic fingerprint as the core seed node, so as to generate a candidate anchor point expansion cluster prototype with an independent identifier.

[0104] For the evolution trigger signal generated by S4.1, whose matching degree value has been determined to be lower than the preset evolution threshold, the interface function of the cluster generation module in the dynamic ontology anchoring engine is called to input the error semantic fingerprint of the previous nearest neighbor matching that failed to be associated as the core seed node into the initialization routine of the cluster generation module.

[0105] The core seed node is used to calculate the center vector in the low-dimensional semantic fingerprint space. The weighted mean method is used to sum and average the error semantic fingerprints according to the weight coefficients of their respective feature domains to form the spatial geometric center of the seed node. This center vector is used to define the boundary of the subsequent temporary semantic space region.

[0106] Based on the spatial coordinates of the geometric center of the seed node, a radius parameter for the temporary semantic space region is set. This radius parameter is calculated based on the relationship between the evolution threshold and the seed node matching degree residual, and the calculation expression is as follows: in, For the temporary area radius, This is the proportionality coefficient. The average Euclidean distance between the seed nodes is given. This represents the evolution threshold.

[0107] The spatial partitioning subroutine of the cluster generation module is invoked to register the temporary semantic space region as a region object with an independent spatial index in the low-dimensional fingerprint space, and a unique identifier is assigned to the region object to ensure global recognizability in the graph meta-layer.

[0108] After the region object registration is completed, the core seed node and its spatial radius parameters are recorded in the data structure of the candidate anchor point extended cluster prototype, and its member list is initialized to be empty to be filled in subsequent neighboring domain searches.

[0109] Through the above spatial partitioning and identifier registration operations, the new knowledge seed nodes corresponding to the evolution trigger signals are transformed into candidate anchor point extension cluster prototypes with independent identifiers and boundary parameters, thereby realizing the regional foundation construction of local ontology evolution.

[0110] For example, consider a set of low-dimensional semantic fingerprint space coordinates (1.5, 2). (1.8,2. (1.6,2. A novel numerical vector fingerprint, geometric fingerprint, material response fingerprint, and error phrase fingerprint (1.7, 2.5) were calculated using weighted averages with weighting coefficients of 0.3, 0.3, 0.2, and 0.2, respectively, yielding a geometric center coordinate of (1.675, 2.35). The matching degree residual value D... τ is 0.18, and the proportionality coefficient α is set to 5. The formula calculates the temporary region radius R to be 0.9. This radius and its geometric center are registered in the dynamic ontology anchoring engine to form a temporary semantic space region, and an identifier ID_AEC_20240601 is assigned. This region object serves as a prototype for candidate anchor point expansion clusters. Subsequent neighborhood searches will retrieve uncovered historical fingerprints based on this radius range, ultimately significantly improving the boundary accuracy and index uniqueness of new knowledge inclusion during ontology evolution.

[0111] S4.3: Based on the candidate anchor point extended cluster prototype, perform a neighborhood search operation in the low-dimensional semantic fingerprint space to retrieve discrete historical fingerprint data that are within the Euclidean distance range of the core seed node and are not covered by the existing initial anchor point set. Aggregate the retrieved discrete historical fingerprint data into the candidate anchor point extended cluster prototype to generate a complete anchor point extended cluster entity containing information on the association between new and old knowledge.

[0112] Based on the initialized candidate anchor point extended cluster prototype, the coordinate system of the low-dimensional semantic fingerprint space and the spatial position of the core seed node are used as input conditions. The neighborhood search module is called to set the Euclidean distance search radius and load the existing initial anchor point set coverage index to generate a set of all candidate points whose distance from the core seed node does not exceed the set search radius. Coverage filtering is performed on the candidate point set, using the initial anchor point set coverage index to remove historical fingerprint data already included in the existing initial anchor points, retaining only uncovered discrete historical fingerprint data as qualified candidate points. Based on the data index of qualified candidate points, the corresponding original semantic content and fingerprint feature vectors are retrieved from the historical knowledge base, and their source, collection time, and process category are confirmed through a multi-source data synchronization verification mechanism to ensure the consistency of the aggregated historical fingerprint data with the semantic space of the current extended cluster. The verified discrete historical fingerprint data and the core seed node are subjected to feature vector stacking operation and merged into a set of cluster member feature matrices. The mean and variance of the Euclidean distance between cluster members are calculated to confirm that the spatial adjacency and knowledge association between cluster members meet the preset threshold conditions. Through aggregation processing, the core seed nodes and the verified discrete historical fingerprint data are combined to form anchor point extended cluster entities with complete information on the relationship between old and new knowledge, thereby realizing the organic integration of incremental data in the process of knowledge evolution.

[0113] Through the aforementioned neighborhood search, coverage filtering, data verification, and feature stacking processing methods, the candidate anchor point extension cluster prototype generated in the previous step is transformed into a complete anchor point extension cluster entity containing the relationship between new knowledge and historical knowledge, thereby achieving the expected technical effects of spatial positioning accuracy and knowledge fusion interpretability in local ontology evolution.

[0114] S4.4: Perform semantic consistency verification on the member fingerprints within the complete anchor point extended cluster entity, calculate the variance distribution index of all member fingerprints within the cluster relative to the cluster center vector, and if the variance distribution index is within a preset convergence interval, lock the topology of the complete anchor point extended cluster entity to generate a state-stable standardized anchor point extended cluster.

[0115] For the complete anchor-extended cluster entity after neighborhood historical fingerprint aggregation, the semantic consistency verification module is invoked to extract the vector coordinates of all member fingerprints in the low-dimensional semantic fingerprint space and assemble them into a feature matrix in index order. Based on the feature matrix, the cluster center vector is calculated, defined as the arithmetic mean of all member fingerprint vectors, using the following formula: in, Let be the cluster center vector. For member fingerprint vectors, The number of members is used as a reference. The difference vector between the fingerprint vector of each member and the cluster center vector is calculated, and the difference vector is squared to obtain the variance distribution index. The variance distribution index is calculated using the following formula: in, This is an index of the variance distribution within a cluster. The result is the Euclidean distance between the member fingerprint and the center vector. The system compares the values ​​with the upper and lower limits of the convergence interval preset in the system configuration, and the execution logic control module determines whether it falls within the convergence interval range. If it does, the topology locking subroutine is called to write the member connection relationship, center vector parameters, and edge weight coefficients of the current cluster into the standardized anchor point extended cluster structure storage unit, prohibiting subsequent temporary node insertion or deletion operations. Through this semantic consistency verification and variance convergence determination processing method, the complete anchor point extended cluster entity generated in the previous step is transformed into a state-stable standardized anchor point extended cluster that can directly participate in causal inference, achieving the goal of structural stabilization of local ontology evolution.

[0116] For example, in the scenario of constructing a knowledge graph of error sources in the manufacturing of precision micro-connector molds, a complete anchor point extended cluster entity contains 12 member fingerprints, among which the numerical vector fingerprint, geometric fingerprint, material response fingerprint, and standardized error phrase fingerprint have all been uniformly mapped to a low-dimensional semantic fingerprint space. When calculating the cluster center vector, the average of the member vectors' coordinates in each dimension of the space is used to obtain the values ​​of the center vector C in each dimension as 1.25, 0.87, and 0.87, respectively. The difference vector squares are 0.34 and 2.14; the sum of these squares yields σ², which is 0.056. The system convergence interval is set to [0.05, 0.08]. If σ² falls within this interval, the topology locking subroutine is triggered. After locking, the response time of this standardized anchor point extension cluster in the map query is reduced to less than 30 milliseconds, and no structural drift occurs during attribution inference, significantly improving the stability and traceability of the error source identification results.

[0117] S4.5: Register the state-stable standardized anchor extension cluster to the dynamic evolution index table of the graph meta-layer, and establish a weakly supervised mapping link between the state-stable standardized anchor extension cluster and the original initial anchor set to complete the execution of the local ontology evolution mechanism and output the generated anchor extension cluster.

[0118] Step S5: Based on the fact that the new fingerprint within the anchor point extension cluster and the original anchor point are both incorporated into the same semantic granularity level, the process constraint relationship between the new fingerprint and neighboring anchor points is automatically derived, and conditional rules are generated. Specifically, this includes: S5.1: Based on the new fingerprint within the anchor point extension cluster and the original anchor point, perform semantic granularity alignment processing to generate a standardized set of semantic nodes with a unified description dimension, ensuring the hierarchical consistency between new knowledge and existing knowledge in the graph structure.

[0119] Based on a structured data packet of a state-stable, standardized anchor cluster, the semantic granularity alignment module loads the attribute vectors and relational label information of the new fingerprints and original anchors within the cluster, serving as the input source for semantic mapping. Within the alignment module, a multi-dimensional attribute matrix is ​​constructed, arranging the numerical features, geometric features, material response features, and standardized error phrases of each fingerprint node in a predefined field order to form a unified feature column vector. A dimension alignment operator ensures the dimensional consistency and field position stability of all node vectors. During the alignment process, semantic mapping weights are calculated, using the product of attribute similarity scores and frequency statistics as mapping weights to form a mapping weight matrix, which guides subsequent accuracy optimization of the alignment transformation. The mapping weight matrix... The minimum distortion alignment operator minimizes the semantic difference between nodes by solving the transformation matrix, and calculates the alignment error function value using the following formula: in, For the alignment error function value, For node feature vectors, To align the target feature vector, The number of nodes is [number]. Iterative optimization is performed to minimize the alignment error function value until convergence. The optimized node feature vectors and the original anchor feature vectors are then uniformly encapsulated into a standardized semantic node set. This process transforms the anchor expansion clusters and their member fingerprints from the previous step into a standardized semantic node set with a unified description dimension and consistent hierarchical positioning, achieving semantic granularity consistency between new and existing knowledge in the graph structure.

[0120] For example, in the error source diagnosis scenario of precision micro-connector mold manufacturing, the stable normalized anchor point extension cluster contains the material response feature vector [0.85, 0.62, 0.41, 0.33] and geometric feature vector [0.79, 0.58, 0.47, 0.28] of the new fingerprint, while the original anchor point contains the numerical feature vector [0.81, 0.59, 0.43, 0.30] and geometric feature vector [0.78, 0.57, 0.46, 0.27]. All feature vectors are uniformly arranged into a four-dimensional column vector, and the attribute similarity scores are calculated to be 0.92 and 0.88, respectively, with frequency statistics of 0.97 and 0.94. Therefore, the mapping weights are 0.8924 and 0.8272, respectively. The mapping weight matrix is ​​input into the minimum distortion alignment operator. Under the condition of 15 iteration steps, the alignment error function value E is calculated. Convergence is achieved when E=0.0008, and the aligned node feature vector [0.80,0.59,0.43,0.29] is output. This is encapsulated into a standardized semantic node set. It has been verified that this set significantly improves the accuracy and stability of process constraint derivation in subsequent temporal correlation analysis and causal strength calculation.

[0121] S5.2: Perform process timing correlation analysis on the standardized semantic node set, and use the sliding time window algorithm to extract the sequential features of the new fingerprint and the original anchor point in the processing flow, so as to generate a temporal coupling relationship vector representing the error propagation path.

[0122] The system calls the processing procedure timestamp retrieval module to retrieve the precise timestamp of each semantic node in the manufacturing execution system and its corresponding procedure number.

[0123] The time stamps are grouped by process number, and then arranged in ascending order of time within each group to establish a node sequence, thus constructing an ordered node time series matrix for time series analysis.

[0124] A sliding time window algorithm is applied to the matrix, setting the window width and step size parameters, extracting the time difference vector between nodes based on the time index of the nodes within the window, and generating a multi-window time difference sequence.

[0125] For each time difference sequence, perform time series feature extraction, calculate the node sequence identifier, mean time interval, maximum value and standard deviation, and encapsulate these features into a time series feature vector.

[0126] Error propagation path mapping is constructed based on temporal feature vectors. By labeling the sequential relationship and time interval of each pair of nodes, it is transformed into a temporal coupling relationship vector to reflect the error flow direction in the processing process.

[0127] By using a sliding time window algorithm and feature extraction processing, the standardized semantic node set from the previous step is transformed into a quantitative temporal coupling relationship vector, thereby enabling the accurate capture of error propagation paths and the construction of temporal constraints for causal chains.

[0128] For example, in the machining scenario of precision micro-connector molds, the standardized semantic node set includes numerical vector fingerprint nodes in the "CNC finishing" process, geometric fingerprint nodes in the "EDM cavity machining" process, and material response fingerprint nodes in the "polishing" process. The timestamps obtained from the manufacturing execution system are 2024-06-01 08:30:12, 2024-06-01 10:15:45, and 2024-06-01 11:05:20, respectively. Setting the sliding time window width to 7200 seconds and the step size to 1800 seconds, the above timestamps are mapped to second-level time stamps and grouped by process to obtain the sequence. The time difference between adjacent nodes is calculated within a sliding window, yielding the time difference of 36945 for the first pair of nodes. 30612 = 6333 seconds, the time difference between the second pair of nodes is 39920. 36945 = 2975 seconds. Perform feature statistics on the time difference sequence, calculating the mean (6333 + 2975) / 2 = 4654 seconds, the maximum value (6333), and the standard deviation. Use these statistics to establish a temporal feature vector [4654, 6333, 2975, σ]. Label the relationship between node 1 and node 2 as "predecessor-successor" with a time interval of 6333 seconds, and the relationship between node 2 and node 3 as "predecessor-successor" with a time interval of 2975 seconds. This forms a temporal coupling relationship vector, which is stored in the graph meta-layer. Validation results show that this vector significantly improves the accuracy of error propagation path identification and supports subsequent causal strength calculations.

[0129] S5.3: Perform physical causal strength calculation based on the temporal coupling relationship vector, evaluate the statistical dependence between the error phenomenon represented by the new fingerprint and the error source represented by the original anchor point through mutual information entropy, and generate a confidence score that quantifies the strength of causal association.

[0130] Based on the temporal coupling relationship vector generated in the previous steps, the physical causal strength calculation module is called to use the vector as input and establish a joint distribution model of the error phenomenon corresponding to the new fingerprint and the error source corresponding to the original anchor point.

[0131] A discrete time series window is constructed, and the joint probability density function of the new fingerprint feature value and the original anchor point feature value in each window is calculated by the kernel density estimation method to ensure that the probability structure of process constraints is captured in the low-dimensional semantic fingerprint space.

[0132] Based on the obtained joint probability density function and their respective marginal probability density functions, the mutual information entropy calculation formula is called: in, Let the probability be the joint probability of the two. and Each represents its own marginal probability.

[0133] The above mutual information entropy formula is summed over the entire domain using numerical integration to obtain a single scalar value as the temporary causal correlation strength.

[0134] The mutual information values ​​are normalized, and a standardization coefficient is introduced to eliminate the influence of different data lengths and sampling frequencies on the mutual information results. The final confidence score matrix is ​​generated by averaging the standardized causal strength values ​​of all sliding windows.

[0135] The matrix is ​​used to extract the single-value confidence scores of the new fingerprint and the original anchor point, which are then used as quantitative input parameters for constructing the causal chain to achieve a strength assessment of the process constraint relationship.

[0136] By using the above-mentioned physical causal strength calculation and mutual information entropy evaluation processing methods, the temporal coupling relationship vector is transformed into a quantifiable confidence score, thereby achieving precise quantification of the causal relationship between error sources and error phenomena.

[0137] For example, in a precision micro-connector mold manufacturing scenario, the new fingerprint represents "local cavity ablation," and the original anchor point represents "electrode wear during electrical discharge machining." The temporal coupling vector is sampled within a 0.5-second sliding window to obtain a time series of length 200. The kernel density estimation bandwidth is set to 0.05, and the probability densities of p(x,y), p(x), and p(y) are calculated based on the data within the window. Substituting the probability density values ​​into the mutual information formula yields a mutual information value of approximately 0.87 per window. After adjustment with a standardization coefficient of 0.92, the normalized mutual information value is 0.80. Within the 200 sliding windows, the average confidence score matrix is ​​approximately 0.79. This value is written into the causal relationship index as the basis for selecting strong causal chains in subsequent process constraint rule screening. Performance verification shows that the confidence score in this scenario is significantly higher than the control group without causal association (approximately 0.15), significantly improving the accuracy of causal inference and enabling stable identification of the causal chain between electrode wear and cavity ablation in the diagnostic system.

[0138] S5.4: Based on the confidence score, perform threshold screening and logical verification on potential process constraint relationships, eliminate spurious correlations and retain strong causal links to generate a lightweight candidate set of process constraint rules with interpretability.

[0139] Based on the numerical results of the confidence score matrix, the threshold screening algorithm module is invoked to perform a quantitative comparison operation on the confidence scores of each process constraint relationship. A preset strong causal threshold is used as the screening boundary, and process constraint relationships below the threshold are marked as objects to be eliminated.

[0140] The object to be removed is input into the logic verification submodule. The logic consistency of the predecessor and successor nodes of the process constraint relationship is matched and checked using the verified causal chain template in the process knowledge base. If the relationship has the characteristics of step reversal or physical unreachability in the process flow, it is determined to be a pseudo-related connection and logic removal is performed.

[0141] The remaining process constraints that satisfy the logical consistency condition are reordered and weighted based on the coupling strength coefficient. A strong causal link score is generated using the following formula: in, The link score for the i-th process constraint relationship. This is the coupling strength coefficient of the relationship. Assign a confidence score to the relationship. Compare the weighted result with a preset interpretability threshold, and remove weak links with scores below the interpretability threshold.

[0142] The remaining links are used to form a lightweight candidate set of interpretable process constraint rules, which is then output to the rule generation module for ontology incremental fusion.

[0143] By using threshold filtering and logical verification, the confidence score results from the previous step are transformed into a set of rule candidates that meet the conditions of strong causal links, thereby achieving the expected technical effects of eliminating pseudo-correlation links, enhancing the interpretability of the rule set, and improving the reliability of the inference chain.

[0144] For example, in the error source identification scenario of precision micro-connector mold manufacturing, for a confidence score matrix containing 15 potential process constraint relationships, a strong causal threshold of 0.65, an interpretability threshold of 0.55, and a coupling strength coefficient ranging from 0.8 to 1.2 are set. During threshold filtering, 7 relationships with confidence scores below 0.65 are removed, leaving 8 relationships to enter the logic verification submodule. In the logic verification stage, using a knowledge base containing 38 processing step templates, the consistency of the processing order of each relationship with the template is compared, and 2 relationships are determined to have reversed steps and are removed. A weighted calculation is performed on the remaining 6 relationships. For example, if a relationship has a coupling strength coefficient of 0.95 and a confidence score of 0.72, the link score is calculated using the following formula: 0.95 × 0.72 = 0.684, this score is higher than the interpretability threshold of 0.55 and is therefore retained. After the above processing, the final lightweight process constraint rule candidate set contains 5 strong causal links, which are used in actual attribution analysis to quickly construct the causal chain of error sources, significantly improving the accuracy of the inference results and greatly shortening the error source localization time.

[0145] S5.5: Encapsulate the lightweight process constraint rule candidate set into a parsable condition rule object of the graph meta-layer, and write it into the dynamic evolution index area of ​​the knowledge graph to generate condition rules that support real-time reasoning calls to complete the incremental fusion of the ontology.

[0146] Step S6: The real-time acquired mold detection data is converted into a real-time error semantic fingerprint, and a nearest neighbor search operation is performed on the real-time error semantic fingerprint in the low-dimensional semantic fingerprint space to locate the target anchor point cluster most relevant to the current working condition. Specifically, this includes: S6.1: Acquire real-time multi-source heterogeneous detection data stream from the precision micro-connector mold manufacturing site. The real-time multi-source heterogeneous detection data stream includes the coordinate offset sequence of the CNC machining center, the surface profile deviation cloud map of the coordinate measuring machine, the abnormal discharge parameter records of the EDM equipment, and the phenomenon description text entered by the maintenance terminal. Perform timestamp alignment and noise filtering preprocessing on the real-time multi-source heterogeneous detection data stream to generate a standardized real-time raw data set.

[0147] S6.2: Based on the standardized real-time raw data set, the pre-calibrated feature extraction algorithm group is called for parallel processing: time-frequency joint entropy feature compression operation is performed on the coordinate offset sequence to generate a real-time numerical vector; differential geometric curvature extreme point sampling and principal direction projection transformation are performed on the surface profile deviation cloud map to generate a real-time geometric fingerprint; pulse energy and breakdown delay two-dimensional scatter distribution fitting is performed on the abnormal discharge parameter records to generate a real-time material response fingerprint; restricted entity recognition based on mold process dictionary is performed on the phenomenon description text to generate a real-time standardized error phrase; the above four types of output results are uniformly defined as the real-time error semantic fingerprint set to be mapped.

[0148] Based on a standardized real-time raw data set, the coordinate offset sequence is input into a pre-calibrated time-frequency joint entropy feature compression algorithm. The time-frequency energy distribution matrix is ​​generated using short-time Fourier transform, and the Shannon entropy value is calculated in each time window of the matrix to quantify the trajectory stability. The signal amplitude difference is eliminated by matrix row and column normalization, and the processed entropy value sequence is encapsulated into a real-time numerical vector with fixed length and dimension.

[0149] The surface profile deviation cloud map is input into the preset differential geometric curvature extreme point sampling module. Gaussian curvature is calculated for the local curvature of each measurement point, the extreme point set is filtered, and these extreme points are orthogonally projected on the principal component direction of the normal vector to form a real-time geometric fingerprint that represents the local deformation mode. The fingerprint distribution is normalized in the projection space to facilitate subsequent feature fusion across data sources.

[0150] The abnormal discharge parameters are recorded and input into the two-dimensional scatter distribution fitting module for pulse energy and breakdown delay. The pulse energy value and breakdown delay of each discharge event are calculated. The material response feature vector is constructed by fitting the slope of the edge envelope and the dispersion index. The standard deviation normalization is used to reduce the feature scale difference between different processing batches.

[0151] The text describing the phenomenon is input into the restricted entity recognition processing module based on the mold process dictionary. String matching and entity extraction are performed, and only standardized error phrases consistent with the mold experience base are retained. Text fingerprint data in a unified format is generated, and the phrase index number is hashed to ensure consistency between storage and retrieval.

[0152] The multimodal fingerprint integration interface encapsulates the real-time numerical vector, real-time geometric fingerprint, real-time material response fingerprint, and real-time standardized error phrase into a set of real-time error semantic fingerprints to be mapped, thereby achieving unified feature input across modalities and sources.

[0153] By using the above processing method, the result of the previous step is transformed into structured and projectable real-time error semantic fingerprint data, realizing a robust transformation of real-time detection data from raw signals to semantic features, and ensuring the accuracy of subsequent mapping and retrieval.

[0154] For example, in the manufacturing site of precision micro-connector molds, the sampling frequency of the real-time coordinate offset sequence is 100Hz, and the short-time Fourier transform uses the Hanning window function with a window length of 256 points and a window shift of 128 points. The Shannon entropy value is calculated for the time-frequency energy distribution matrix within each window. ,in The normalized spectral energy yields an entropy vector of length 64 and dimension 32. The surface profile deviation cloud map, output from a coordinate measuring machine, contains 20,000 point cloud data points. Ten Gaussian curvature extrema points are sampled, and a 12-dimensional geometric fingerprint is generated by orthogonal projection onto the principal normal direction. The discharge parameter anomaly record is sampled at a frequency of 10 kHz, with pulse energy ranging from 0.02 to 0.04 J and breakdown delay ranging from 2 to 5 μs. A 2-dimensional material response fingerprint is constructed by fitting the edge envelope, resulting in a slope of 0.85 and a dispersion of 0.12. The phenomenon description text is recognized by OCR as... Localized ablation of the cavity The corresponding phrase index number in the process dictionary is 005, and the hash encoding result is... d93a1f The fingerprint integration interface encapsulates the above four types of features into a unified structure input in a fixed order. The verification results show that the real-time error semantic fingerprint set can accurately correspond to the high neighbor region of the similarity score matrix in historical cases when mapped in low-dimensional space, thus significantly improving the accuracy of subsequent nearest neighbor retrieval and anchor point clusters.

[0155] S6.3: Using the low-dimensional semantic fingerprint space mapping matrix constructed in the previous step, perform a linear projection transformation on each fingerprint element in the real-time error semantic fingerprint set to be mapped, compress the high-dimensional feature data into a low-dimensional semantic fingerprint space coordinate system isomorphic to the initial anchor point set, so as to generate a real-time projected fingerprint vector set with spatial comparability.

[0156] For each fingerprint element in the real-time error semantic fingerprint set to be mapped, the low-dimensional semantic fingerprint space mapping matrix is ​​called, the transformation coefficient group of the matrix is ​​read and loaded into the projection operation core to ensure that the input features are consistent with the mapping basis of the target space coordinate system.

[0157] Centering and standardization are performed on the real-time numerical vector, geometric fingerprint, material response fingerprint, and standardized error phrase, respectively. The mean and standard deviation of each feature component are calculated, and the mean of each component is subtracted from the standard deviation to eliminate the influence of scale differences in different feature domains on projection accuracy.

[0158] The normalized high-dimensional feature tensor and the mapping matrix are multiplied together to generate the corresponding low-dimensional projected coordinate vector. The matrix multiplication is performed according to the following formula: in, Represents a low-dimensional projection vector. Represents the mapping matrix, This represents a high-dimensional fingerprint vector.

[0159] An abnormal peak value in the projected coordinate vector is smoothed by applying a median filter. The local window value distribution of the projected coordinate sequence is statistically analyzed and the window center value is replaced to reduce the interference of noise on spatial matching.

[0160] The smoothed low-dimensional projection vector set is sorted and stored according to the index of the initial anchor point set spatial coordinate system, and the source category and timestamp of each vector are marked to form a real-time projection fingerprint vector set with spatial comparability.

[0161] Through the above processing method, the real-time error semantic fingerprint set to be mapped in the previous step is transformed into a structured, low-dimensional real-time projected fingerprint vector, realizing the spatial comparability matching between real-time operating conditions and historical knowledge.

[0162] For example, in the manufacturing process of precision micro-connector molds, the mapping matrix W is configured as 4×12 dimensions, covering all feature dimensions of real-time numerical vectors, geometric fingerprints, material response fingerprints, and text phrase fingerprints. The high-dimensional feature x acquired in real time contains 12 items, where the numerical vector component is normalized to 0.67 after normalization with a mean of 25.0 and a standard deviation of 3.0, and the geometric fingerprint component is normalized to [value missing]. The material response fingerprint component normalization value is 0.53, the phrase fingerprint component normalization value is 1.12, and the phrase fingerprint component normalization value is 0.53. 0.21. Performing matrix multiplication y=W×x yields a 4-dimensional projection vector y=[0.45, [0.12, 0.63, 0.08], after processing this vector using a 3-point window mid-range filter, a smoothed result [0.44, [0.11, 0.61, 0.09]. The vector is stored according to the spatial coordinate system of the initial anchor point set and labeled with the category label "geometric pattern deviation" and the timestamp "2024-06-18T10:15:32Z". This results in a real-time projected fingerprint vector group with spatial comparability, which significantly improves matching stability and diagnostic accuracy in subsequent Euclidean distance calculations.

[0163] S6.4: Based on the real-time projected fingerprint vector group and the initial anchor point set stored in the knowledge base, calculate the Euclidean distance metric between each real-time projected fingerprint vector and all initial anchor points, and construct a similarity scoring matrix containing distance values ​​to quantitatively characterize the spatial proximity between the current real-time operating conditions and historical typical error cases.

[0164] S6.5: Perform a nearest neighbor search operation based on the similarity scoring matrix, select the K initial anchor points with the smallest Euclidean distance to form a candidate set, and perform topological connectivity verification on the initial anchor points in the candidate set. Merge adjacent anchor points with strong process constraints to form the final target anchor point cluster, thereby completing the process of finding the most relevant anchor points for the current working condition. Precise positioning of subgraphs.

[0165] The input condition is the similarity score matrix generated in the previous step S6.4. This matrix contains the Euclidean distance metric between each real-time projected fingerprint vector and all initial anchor points, and the matrix row and column indices have been mapped to the low-dimensional semantic fingerprint space coordinate system and the knowledge base anchor point index table.

[0166] The similarity score matrix is ​​subjected to a sorting and retrieval process based on the K value. In each row of the matrix, a sorting algorithm is called to sort the Euclidean distance metric values ​​in ascending order, and the initial anchor indexes of the top K distance values ​​are extracted to form a candidate set.

[0167] The topology connectivity verification module is called on the formed candidate set. The pre-stored adjacency matrix of process constraint relationship between anchor points is used to determine whether there are strong constraint links between anchor points in the set. During the judgment process, anchor point pairs with direct connectivity or multi-hop connectivity paths and path weights higher than preset thresholds are marked.

[0168] Perform set merging operations on the marked strongly constrained anchor point pairs, aggregating all anchor points that satisfy the connectivity conditions into a continuous domain, forming several anchor point subsets, and calculating the average distance value within each subset as a representative index within the domain.

[0169] All anchor point subsets are sorted at the domain level according to representative indicators. The top-level domain with the best indicators is selected as the final target anchor point cluster. This target anchor point cluster is then output to the causal reasoning module of the main step S7 to complete the accurate positioning of the most relevant knowledge subgraph of the current working condition.

[0170] By combining sorting retrieval with topological connectivity verification, the similarity score matrix from the previous step is transformed into a cluster of target anchor points with consistent process constraints and spatial proximity features, thereby improving the accuracy and interpretability of the input in the causal reasoning stage.

[0171] Step S7: Perform a graph traversal operation along the evolved conditional rules within the target anchor cluster to generate attribution inference results containing the causal chain of error sources. Specifically, this includes: S7.1: Based on the initial error semantic fingerprint within the target anchor cluster, extract the set of conditional rules directly associated with the fingerprint, and perform Boolean logic verification on the predecessor node constraints in the set of conditional rules to generate an activation rule list containing all candidate process constraint relationships that satisfy the triggering conditions.

[0172] Based on the target anchor cluster that has been located, the knowledge base retrieval interface is called to load a set of conditional rules that are directly associated with the initial error semantic fingerprint within the cluster. The set of rules includes various types of trigger definitions such as node attribute constraints, process parameter dependencies, and deformation pattern matching conditions.

[0173] For each rule, the constraints of the predecessor node are mapped to fields. A one-to-one correspondence is established between the feature values ​​of each dimension in the real-time error semantic fingerprint vector and the corresponding parameter indexes in the rule constraints, so as to ensure the consistency of the input data for matching judgment.

[0174] A Boolean logic verification module is used to perform conditional operations on the mapped feature value set. If the combination conditions defined in the rule are multiple intersection and union relationships, the decision path that does not meet the conditions is terminated first through short-circuit logic during the operation, thereby improving the efficiency of rule filtering.

[0175] For numerical constraints, a range comparison operation is used to determine whether the real-time fingerprint feature value is within the threshold range specified by the rules. The range comparison formula is as follows: ,in Real-time fingerprint feature values, and These are the lower and upper limits of the threshold defined in the rule, respectively. If the condition is met, the flag is set to true.

[0176] For enumerated constraints, the set inclusion operation is used to determine whether the standardized error phrase corresponding to the real-time fingerprint belongs to the pre-defined enumerated set of the rule. If the set matches successfully, the sub-condition is marked as true in the multi-condition rule by Boolean AND operation.

[0177] All verification results are merged to generate an activation rule list containing all candidate process constraint relationships that meet the triggering conditions, along with the rule ID, trigger timestamp, and associated node index, to support subsequent timing determination and causal traversal operations.

[0178] By using a multi-source rule verification and merging process, the target anchor cluster and knowledge base rule structure from the previous step are transformed into an executable list of activation rules, enabling accurate identification of triggerable process constraint relationships under real-time operating conditions and providing a reliable data foundation for time series consistency judgment.

[0179] For example, in a precision micro-connector mold manufacturing process, during a cause analysis task, the target anchor point cluster contains two types of initial error semantic fingerprints: "local cavity ablation" and "electrode loss." The corresponding numerical constraints in the rule set define the discharge peak current range [15.0, 18.5] Amps, and the enumerated constraints include the phrase "local cavity ablation." The measured discharge peak current in the real-time fingerprint feature value is 16.8 Amps, satisfying the range condition, and the standardized phrase matching is successful. The Boolean logic verification result is true, and a new rule ID#R103 is added to the rule list, with the trigger timestamp recorded as 2024-06-18 10:35:27, and the associated node index as (node ​​5 → node 8). This rule is used in the subsequent S7.2 step's timing consistency judgment to confirm the process constraint relationship that "electrode loss" precedes "local cavity ablation," ultimately significantly improving the accuracy of error propagation path location during causal chain traversal.

[0180] S7.2: Perform temporal consistency discrimination processing on each candidate process constraint relationship in the activation rule list, and use the timestamp sequence of mold processing steps to compare the order dependency defined in the rules to filter out the legitimate causal jump path that conforms to the actual manufacturing process and generate an ordered causal edge set.

[0181] S7.3: Perform a depth-first graph traversal operation based on the ordered causal edge set, pass the error confidence weight of the current node level by level along the traversal direction, and perform weighted cumulative calculation in combination with the coupling strength coefficient in the condition rules to generate a dynamic causal chain score vector that characterizes the reliability of the error propagation link.

[0182] Based on each legitimate causal jump path in the ordered causal edge set, the depth-first graph traversal algorithm is called as the traversal control core. The traversal stack structure is initialized and the error confidence weight of the starting node of the path is set as the original score value output by the previous causal discrimination step.

[0183] During the traversal, for each current node, the coupling strength coefficient in its corresponding condition rule is extracted, and the coefficient is multiplied with the confidence weight of the current node to obtain the confidence gain value passed along the current edge to the successor node.

[0184] When performing weight transfer, the confidence gain value is combined with the existing cumulative confidence value of the successor node in a weighted accumulation manner. To ensure the reliability of the traversal path, a node state update operation is performed after the accumulation calculation. The updated confidence value is written into the node attribute table and the traversal depth is recorded for subsequent path pruning conditions.

[0185] When recursively calling depth-first traversal, when the traversal reaches the end node of the path or meets the preset traversal termination condition (such as the traversal depth exceeding the maximum level), the confidence values ​​of all nodes in the current path are combined in order to form a dynamic causal chain score vector.

[0186] Through the above processing method, the ordered causal edge set result of the previous step is transformed into a dynamic causal chain score vector containing information on the step-by-step weight transfer and coupling strength accumulation of nodes, thereby realizing a quantitative characterization of the reliability of the error propagation link.

[0187] S7.4: Perform threshold truncation and redundancy pruning on each traversal path in the dynamic causal chain scoring vector, remove branches with scores lower than the preset confidence threshold, and merge duplicate causal sub-chains pointing to the same endpoint to generate a simplified and logically complete main error source causal topology graph.

[0188] The input for each traversal path in the dynamic causal chain scoring vector is an ordered causal edge set generated by depth-first traversal and a corresponding link reliability scoring data matrix. The scoring data matrix contains the node order of each causal path and a confidence score calculated by weighted accumulation using coupling strength coefficients. Based on this input, a threshold discrimination module is invoked to perform logical comparison operations on the link reliability score of each path in the scoring data matrix. The score value is compared one by one with a preset confidence threshold in the system configuration to generate a set of judgment flags. Paths with scores below the confidence threshold are filtered out using the judgment flag set, and the corresponding rows are removed from the scoring data matrix to form a candidate set of remaining paths. Endpoint node aggregation analysis is performed on the remaining candidate set of paths to extract the endpoint node identifiers of all paths and construct an endpoint similarity index table. By comparing the difference index of paths with the same endpoint node identifiers on the node sequence, duplicate causal sub-chains pointing to the same endpoint are identified. Redundancy pruning is performed on the duplicate causal sub-chains, retaining the path with the highest link reliability score and the most complete node connection logic, while deleting the remaining paths with the same endpoint. The topology reconstruction module maps the pruned path set back to the semantic node structure within the target anchor cluster, and rearranges the edge index order according to the process constraints between nodes, generating a simplified and logically complete causal topology graph of the main error source. Through the threshold truncation and redundancy pruning methods described above, the dynamic causal chain score vector from the previous step is transformed into causal topology data with high reliability and low redundancy, thereby improving the interpretability and inference efficiency of the attribution analysis module.

[0189] For example, in error source identification at the precision micro-connector mold manufacturing site, the dynamic causal chain scoring vector contains 20 traversal paths, with each path corresponding to a score value ranging from 0.35 to 0.92. In this embodiment, the confidence threshold is set to 0.60. The threshold discrimination module compares the score value with the threshold one by one, filtering out 9 paths with scores below 0.60, forming a candidate set of 11 remaining paths. Endpoint node aggregation analysis shows that there are 3 sets of duplicate endpoint node identifiers in the remaining paths, corresponding to "EDM collapse". The errors identified were "local cavity ablation" and "locating pin wear." Redundancy pruning was used to calculate the difference index for each group of repeated paths. Taking two paths ending at "EDM collapse" as an example, their scores were 0.78 and 0.82 respectively. The path with a score of 0.82 was retained, and the path with the lower score was deleted. After pruning, eight paths remained. When mapping these paths back to the semantic node topology, a process constraint rearrangement algorithm was called to rearrange the edge indices according to the processing time sequence, generating a causal topology graph of the main error sources. In this topology graph, all links satisfied a score greater than 0.60 and had no redundant sub-chains, significantly improving attribution inference efficiency. The processing latency of the inference engine in field testing was reduced to half that of the original scheme, ensuring the timely and reliable output of error source conclusions.

[0190] S7.5: Based on the mapping relationship of the endpoint nodes of the main error source causal topology diagram, trace back and connect the process parameter deviation descriptions and geometric deformation pattern characteristics of all intermediate nodes on the path to generate an error source causal chain attribution reasoning result containing complete evolutionary logic and key evidence support.

[0191] Based on the mapping relationship of the endpoint nodes in the causal topology graph of the main error source, the endpoint nodes of the topology graph are bound to the fingerprint elements of the target working condition to form a location data index for the terminal error source. This index is traversed in reverse order along the causal chain, and the path backtracking module is called to access all intermediate nodes level by level. The process parameter deviation description data stored in the matching and retrieval stage of each node is extracted sequentially, and its associated geometric deformation pattern feature matrix is ​​read synchronously. The process parameter deviation description data and the geometric deformation pattern feature matrix are linked together according to the node time series order to establish a composite data set containing node indexes, process deviation vectors, and deformation features. Consistency verification is performed on the composite data set by calculating the collaborative change index between the process deviation vector and the deformation feature matrix. Node data that passes consistency verification is concatenated into a recursive link to form a complete evolutionary logical sequence, with the causal confidence score generated by each node during the inference phase added. The evolutionary logical sequence is then structurally formatted, encapsulating the logical chain, key evidence items, and corresponding causal strength indicators into attribution reasoning result data objects. Through reverse tracing and concatenation processing, the results of the previous step are transformed into a complete chain of attribution reasoning reports, achieving interpretable error source diagnosis based on the main causal topology.

[0192] Step S8: Based on the attribution reasoning results, output specific error source diagnostic conclusions and update the initial anchor set to complete the incremental fusion closed loop of the knowledge graph. Specifically, this includes: S8.1: Perform confidence-weighted aggregation processing on the attribution reasoning results containing the causal chain of error sources to generate specific error source diagnostic conclusions with clear probability distributions.

[0193] S8.2: Extract novel error semantic fingerprints under the current operating conditions based on the diagnostic conclusions of the specific error sources, so as to construct new knowledge candidates to be added to the database.

[0194] S8.3: Calculate the topological distance between the new error semantic fingerprint in the newly added knowledge candidate and the existing initial anchor set to determine whether the dynamic expansion condition of the initial anchor set is met.

[0195] S8.4: Perform a local topology reconstruction operation on the initial anchor point set based on the determination result of the dynamic expansion condition to generate an updated initial anchor point set that incorporates the new error mode.

[0196] For the initial anchor point set based on the dynamic expansion condition judgment result, the local topology reconstruction module is called to load the new error pattern nodes that need to be merged, and the low-dimensional semantic fingerprint vector of the node is used as the starting element for merging the anchor point cluster.

[0197] The adjacency matrix of the existing anchor points in the current initial anchor point set and the new error mode nodes that have an Euclidean distance of less than a preset proximity threshold is updated. The edge weights of the new anchor points and the new nodes are set to a weighted value based on the confidence score of the process constraint, so as to enhance the correlation strength between the new mode and the existing mode.

[0198] The updated adjacency matrix is ​​subjected to local subgraph extraction operation to identify the shortest path set containing all nodes with novel error patterns. Within the subgraph, process constraint rules are synchronized and node attribute inheritance is performed to maintain the consistency of topology and semantic rules.

[0199] A dual-index optimization strategy based on node degree and path betweenness centrality is adopted to reset the node connectivity order within a local subgraph, using the following formula: in, For the node To the node The total number of shortest paths, For the nodes The number of shortest paths.

[0200] Using the product of the betweenness centrality value and the node degree value as weights, the edges of nodes in the new error pattern are sorted, and high-weight edges are retained first to improve the stability of the ontology structure.

[0201] The optimized local subgraph is merged with the global initial anchor set to generate an updated initial anchor set containing nodes with new error patterns and their weighted optimized edges. The changed edge set and node attributes are recorded in the topology index table.

[0202] By using an adjacency matrix update and node centrality optimization approach, the newly added knowledge candidates from the previous step are transformed into an updated initial anchor set that is structurally stable, semantically consistent, and has high connection efficiency, thus achieving the effect of local topology optimization in the closed loop of graph fusion.

[0203] S8.5: Based on the updated initial anchor set, trigger the knowledge graph metadata version iteration mechanism to complete the incremental fusion closed loop of the knowledge graph and output the final diagnostic report.

[0204] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0205] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The term "first" as used in the patent application specification and claims herein... second The terms "third" and similar expressions do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "a" or "one" and similar expressions do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar expressions mean that the elements or objects preceding "comprising" or "including" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. In the embodiments of this application, "multiple" refers to two or more. "A and / or B" indicates three possibilities: A; B; and A and B.

[0206] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for diagnosing error sources in the manufacturing process of precision micro-connector molds, specifically including: S1: Collect multi-source heterogeneous data during the manufacturing process of precision micro connector molds, including coordinate offset sequence, surface profile deviation cloud map, discharge parameter anomaly record and phenomenon description text; S2: Perform time-frequency joint entropy compression processing, curvature extreme point sampling and main direction projection processing, pulse energy and breakdown delay scatter point distribution fitting processing, and entity recognition processing based on process dictionary on the multi-source heterogeneous data to generate error semantic fingerprints. S3: Construct an initial anchor point set based on the pre-calibrated typical error case set, and uniformly map the error semantic fingerprint to a low-dimensional semantic fingerprint space, perform nearest neighbor matching operation, and generate fingerprint anchor point mapping relationship; S4: Determine whether the matching degree value in the fingerprint anchor mapping relationship is lower than the preset evolution threshold. If the determination result is lower than the preset evolution threshold, trigger the local ontology evolution mechanism to generate an anchor extension cluster. S5: Based on the fact that the new fingerprint in the anchor point extension cluster is included in the same semantic granularity level as the original anchor point, the process constraint relationship between the new fingerprint and the neighboring anchor points is automatically deduced, and conditional rules are generated. S6: Convert the real-time collected mold detection data into a real-time error semantic fingerprint, and perform a nearest neighbor search operation on the real-time error semantic fingerprint in the low-dimensional semantic fingerprint space to locate the target anchor point cluster most relevant to the current working condition. S7: Perform a graph traversal operation along the evolved conditional rules within the target anchor cluster to generate attribution inference results containing the causal chain of error sources.

2. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that, Step S7 is followed by step S8, which specifically includes: S8: Based on the attribution reasoning results, output specific error source diagnostic conclusions and update the initial anchor set to complete the incremental fusion closed loop of the knowledge graph.

3. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that, Step S2 specifically includes: The coordinate offset sequence in the CNC machining log is compressed using time-frequency joint entropy features. The time-frequency energy distribution matrix is ​​obtained by using short-time Fourier transform and the Shannon entropy value is calculated to generate a numerical vector fingerprint. Based on the original measurement object corresponding to the numerical vector fingerprint, the surface profile deviation cloud map in the coordinate measuring machine report is sampled for differential geometric curvature extreme points and projected in the principal direction. By calculating the Gaussian curvature extreme point set and performing orthogonal projection in the principal component direction of the normal vector, a geometric fingerprint is generated. Based on the process associated with the geometric fingerprint, the abnormal records of electrical discharge parameters of electrical discharge machining are fitted with a two-dimensional scatter distribution of pulse energy and breakdown delay. A two-dimensional scatter plot of pulse energy and breakdown delay is constructed, and the slope of the edge envelope and the dispersion statistics are extracted to generate a material response fingerprint. Based on the fault phenomena involved in the material response fingerprint, the phenomenon description text in the handwritten fault notes of the maintenance engineer is subjected to restricted entity recognition processing based on the mold process dictionary. The predefined mold process dictionary containing terms such as locating pin wear and local cavity ablation is used to perform string matching and entity extraction to generate standardized error phrase fingerprints. Based on the standardized error phrase fingerprint and the other three types of fingerprint data, an error semantic fingerprint operation is performed to encapsulate the numerical vector fingerprint, geometric fingerprint, material response fingerprint, and standardized error phrase fingerprint into a unified data structure to generate the error semantic fingerprint.

4. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that, Step S4 specifically includes: Based on the fingerprint anchor mapping relationship generated in the previous steps, the matching degree value representing the similarity between the new fingerprint and the initial anchor set is extracted, and the matching degree value is logically compared with the evolution threshold preset in the system configuration to generate an evolution trigger signal. In response to the determination result in the evolution trigger signal indicating that the matching degree value is lower than the preset evolution threshold, the cluster generation module in the dynamic ontology anchoring engine is invoked to initialize a temporary semantic space region to be expanded, using the unmatched error semantic fingerprint as the core seed node, so as to generate a candidate anchor point expansion cluster prototype with an independent identifier. Based on the candidate anchor point extended cluster prototype, a neighborhood search operation is performed in the low-dimensional semantic fingerprint space to retrieve discrete historical fingerprint data that are within the Euclidean distance range of the core seed node and are not covered by the existing initial anchor point set. The retrieved discrete historical fingerprint data is aggregated into the candidate anchor point extended cluster prototype to generate a complete anchor point extended cluster entity containing information on the association between new and old knowledge. Semantic consistency verification is performed on the member fingerprints within the complete anchor point extended cluster entity. The variance distribution index of all member fingerprints within the cluster relative to the cluster center vector is calculated. If the variance distribution index is within a preset convergence interval, the topology of the complete anchor point extended cluster entity is locked to generate a state-stable standardized anchor point extended cluster. The state-stable standardized anchor point extension clusters are registered in the dynamic evolution index table of the graph meta-layer, and a weakly supervised mapping link is established between the state-stable standardized anchor point extension clusters and the original initial anchor point set to complete the execution of the local ontology evolution mechanism and output the generated anchor point extension clusters.

5. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that: The multi-source heterogeneous data is organized into a comprehensive package after time alignment. The input feature extraction module performs time-frequency joint entropy compression processing, curvature extreme point sampling and principal direction projection processing, pulse energy and breakdown delay scatter distribution fitting processing, and entity recognition processing based on process dictionary in parallel. This generates an error semantic fingerprint containing numerical vectors, geometric fingerprints, material response fingerprints, and standardized error phrases. The fingerprint is then packaged into a unified format through an integrated interface to facilitate mapping to a low-dimensional semantic fingerprint space.

6. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that: The low-dimensional semantic fingerprint space is reduced in dimensionality through principal component analysis. The length of the feature vector in the space after dimensionality reduction is set according to the information retention requirements. The principal component contribution rate is set to meet the requirement that information retention is dominated by process differences. It also supports distance calculation under the same spatial projection of semantic fingerprints with different errors.

7. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that: When the local ontology evolution mechanism is triggered, the anchor point expansion cluster is dynamically grown with the new fingerprint as the center point and the radius is adaptively expanded according to the distance residual. The cluster members include the new fingerprint and historical uncovered fingerprints within the radius. After the expansion is completed, the semantic consistency is checked. If the variance of the fingerprint center within the cluster is within the set convergence interval, the topology is locked and registered as the anchor point expansion cluster.

8. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that: Different anchor points are extended to clusters, and the initial anchor point and the extended anchor point are connected by a weakly supervised mapping relationship to build a chain fusion of new and old knowledge. The dimension alignment between nodes in each cluster is optimized by the attribute weight matrix for adaptive minimum distortion, and a unified semantic node set is formed.

9. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that: The attribution reasoning results include causal logic chains, process parameter deviations, deformation characteristics, and diagnostic confidence indices, and based on these results, continuously drive the incremental evolution closed loop of knowledge graph metadata.

10. The method for diagnosing error sources in the manufacturing process of precision micro connector molds according to claim 1, characterized in that: The processing steps S2 to S7 can be distributed and support parallel computation. The mapping of the error semantic fingerprint to the low-dimensional semantic fingerprint space, the triggering of the local ontology evolution mechanism to generate the anchor point expansion cluster, the derivation of the condition rules, and the execution of the graph traversal operation can all be implemented independently on different edge computing nodes.

Images