Electric spark forming machine fault data analysis method and system based on cross-equipment migration

Abstract

The invention discloses an electric spark forming machine fault data analysis method and system based on cross-equipment migration, and relates to the technical field of manufacturing equipment data analysis, and the method comprises the steps of multi-source fault data collection, multi-modal data preprocessing and alignment, discharge feature and fault factor extraction, cross-machine knowledge migration analysis, causal chain inference and root cause positioning. According to the method, multi-source data are collected, multi-dimensional fault factors are uniformly preprocessed and extracted, the multi-dimensional fault factors are stored in a feature library, cross-machine knowledge migration and comparative learning are combined, low-frequency and novel fault recognition is achieved, a fault evolution path is revealed by using a causal relationship graph, and a root cause is positioned; and the system generates early warning, issues pulse width, energy and flushing period optimization instructions, feeds back and updates a knowledge base, and realizes closed-loop optimization.

Description

Technical Field

[0001] This invention relates to the field of manufacturing equipment data analysis technology, and in particular to a method and system for fault data analysis of electrical discharge machining (EDM) machines based on cross-equipment migration. Background Technology

[0002] Forming machines are widely used in precision machining, and their processing principle relies on the discharge effect between the electrode and the workpiece. However, in actual production, due to factors such as discharge current fluctuations, unstable discharge voltage, dielectric contamination, electrode wear, and changes in ambient temperature, forming machines are prone to processing defects such as short circuits, open circuits, burns, collapses, or microcracks.

[0003] Currently, Chinese invention patent application number CN202410963512.1 discloses an intelligent fault diagnosis system for molding machines, including a behavioral feature analysis module, a state difference quantification module, a strategy generation and execution module, a fault mode matching module, an abnormal behavior identification module, a response strategy optimization module, and a knowledge base update and optimization module. Through the behavioral feature analysis module and the state difference quantification module, the system can more accurately identify subtle changes and potential deviations in equipment operation, quickly adapting to complex and variable operating environments. This ensures the equipment operates in optimal condition while significantly advancing fault warning times, reducing the risk of unexpected downtime. The fault mode matching module and the abnormal behavior identification module enable the system to utilize a historical case database for deep learning, improving the accuracy and efficiency of fault diagnosis. Through the analysis and application of historical fault data using intelligent algorithms, the system can quickly match and identify current fault modes.

[0004] The aforementioned technologies are insufficient for unified processing of multi-source data and cross-machine knowledge transfer, making it difficult to identify low-frequency and novel faults in a timely manner and to complete interpretable root cause localization and closed-loop optimization. Summary of the Invention

[0005] The technical problem solved by this invention is that existing technologies are unable to achieve unified processing of multi-source data and cross-machine knowledge transfer, thus making it difficult to identify low-frequency faults and new types of faults in a timely manner and to complete interpretable root cause localization and closed-loop optimization.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: The method for fault data analysis of EDM machines based on cross-device migration includes the following steps: Step S1: Collect the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log of the molding machine; Step S2 involves preprocessing and aligning the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs. Step S3: Extract the discharge characteristics and fault factors from the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control log. Step S4: Establish a cross-machine operation knowledge base and perform transfer analysis on discharge characteristics and fault factors to output fault prediction probability. Step S5: Construct a causal relationship graph and establish causal dependency chains, identify evolutionary paths, and output root cause localization conclusions; Step S6: Based on the fault prediction probability, perform early warning and closed-loop optimization.

[0007] Preferably, step S1 includes the following sub-steps: Step S101: During the molding process, simultaneously collect discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log. Step S102: Add the machine number, workpiece number, and acquisition timestamp to the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log.

[0008] Preferably, step S2 includes the following sub-steps: Step S201: Denoise and amplitude standardization are performed on the discharge current waveform and discharge voltage waveform; Step S202: Perform outlier removal on the dielectric conductivity data and temperature data; Step S203: Perform edge enhancement and texture extraction processing on the electrode wear image; Step S204: Synchronize the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log using a unified timestamp alignment mechanism.

[0009] Preferably, step S3 includes the following sub-steps: Step S301: Based on the hybrid depth model, feature recognition is performed on the discharge current waveform and voltage waveform to obtain the distribution of single discharge duration; Step S302: Calculate the proportion of abnormal discharge events in the signal to form an abnormal spark ratio index; Step S303: Extract plasma channel stability indicators from the discharge current waveform and voltage waveform. Step S304: Analyze the pit texture and edge morphology of the electrode wear image and output the uniformity and defect degree. Step S305: Extract defect features from the processed surface image, the defect features including burns, collapses and microcracks; Step S306: Combine the single discharge duration distribution, abnormal spark ratio index, plasma channel stability index, uniformity and defect degree, and defect characteristics to form a multidimensional fault factor vector.

[0010] Preferably, step S4 includes the following sub-steps: Step S401: Establish a cross-machine operation knowledge base, which includes historical fault factor vectors and labeled fault category information of different machines under various operating conditions; The predicted failure probability is obtained by weighted voting based on the similarity of Top-K similar historical samples and then calibrated for probability. At the same time, similar sample identifiers are output to form traceable evidence. Step S403: Combine the comparison results with the contrastive learning model to extract common features and difference features, identify and classify rare fault types outside of common patterns. When the embedding distance between a sample and any known fault category exceeds the threshold, it is judged as a rare fault. After forming a new category through clustering and manual confirmation, it is written back to the knowledge base.

[0011] Preferably, step S5 includes the following sub-steps: Step S501: Construct a causal relationship graph and establish a causal dependency chain by combining the multidimensional fault factor vector, dielectric conductivity, temperature data and control logs; Step S502: Identify the evolution path of the fault from the initial cause to the final manifestation based on the causal dependency chain; Step S503: Output the root cause localization conclusion, which includes the fault cause, evolution chain and impact result, and write the conclusion into the diagnostic record.

[0012] Preferably, step S6 includes the following sub-steps: Step S601: When the predicted failure probability exceeds a preset threshold, generate real-time warning information and input the real-time warning information into the operation interface. Step S602: Based on the multi-dimensional fault factor vector generated in step S3 and the root cause localization conclusion output in step S5, an adaptive optimization command is sent to the machine control terminal. The optimization command includes: The pulse width parameter is fine-tuned based on the discharge current waveform and discharge voltage waveform acquired in step S1 and preprocessed in step S2. The energy required to perform the operation is reduced based on the stability index of the plasma channel. Extend the flushing cycle based on the dielectric conductivity and temperature data processed in step S2; Step S603: After the optimization instruction is implemented, compare the pre-optimization running data formed in steps S1 to S3 with the new data after optimization, and feed the difference results and the control log record of step S1 back to the knowledge base.

[0013] Preferably, the multidimensional fault factor vectors are uniformly stored in a fault feature library after generation. The fault feature library saves the original vectors and the corresponding acquisition conditions, machine number and time tag.

[0014] Preferably, the multidimensional fault factor vector generated in step S3 is compared with the cross-machine knowledge base in step S4 to output a fault factor similarity matrix. Dynamic optimization is achieved by continuously iterating and updating the fault factor similarity matrix, recording the similarity of fault factors between different machines, and automatically adjusting the matching weight each time a new fault factor vector is added.

[0015] The fault data analysis system for EDM machines based on cross-device migration includes a data acquisition module, a data processing module, a feature extraction module, a knowledge analysis module, a chain inference module, and an early warning and optimization module. The data acquisition module is used to acquire the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log of the molding machine. The data processing module is used to preprocess and align the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs. The feature extraction module is used to extract discharge characteristics and fault factors from the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs. The knowledge analysis module is used to establish a cross-machine operation knowledge base and perform transfer analysis on discharge characteristics and fault factors, and output fault prediction probability. The chain inference module is used to construct a causal relationship graph and establish a causal dependency chain, identify evolutionary paths, and output root cause location conclusions; The early warning optimization module is used to perform early warning and closed-loop optimization based on the fault prediction probability.

[0016] The beneficial effects of this invention are as follows: This invention collects multi-source data, performs unified preprocessing and extracts multi-dimensional fault factors, stores them in a feature library, and combines cross-machine knowledge transfer and comparative learning to achieve low-frequency and novel fault identification. It uses causal relationship graphs to reveal the fault evolution path and locate the root cause. When the predicted fault probability exceeds the threshold, the system generates an early warning and issues pulse width, energy and flushing cycle optimization instructions, and feeds back to update the knowledge base to achieve closed-loop optimization. Attached Figure Description

[0017] Figure 1A flowchart illustrating the steps of an EDM machine fault data analysis method based on cross-device migration provided in an embodiment of the present invention; Figure 2 This is a basic flowchart of an EDM machine fault data analysis system based on cross-device migration, provided as an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1, referring to Figure 1 This paper provides a method for fault data analysis of EDM machines based on cross-device migration, including the following steps: Step S1: Collect the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log of the molding machine.

[0020] Step S2 involves preprocessing and aligning the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs.

[0021] Step S3: Extract the discharge characteristics and fault factors from the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control log.

[0022] Step S4: Establish a cross-machine operation knowledge base and perform transfer analysis on discharge characteristics and fault factors to output the fault prediction probability.

[0023] Step S5: Construct a causal relationship graph and establish causal dependency chains, identify evolutionary paths, and output root cause localization conclusions.

[0024] Step S6: Based on the fault prediction probability, perform early warning and closed-loop optimization.

[0025] This invention collects multi-source data, performs unified preprocessing, extracts multi-dimensional fault factors, stores them in a feature library, and combines cross-machine knowledge transfer and comparative learning to identify low-frequency and novel faults. It uses a causal relationship graph to reveal the fault evolution path and locate the root cause. When the predicted fault probability exceeds the threshold, the system generates an early warning and issues pulse width, energy, and flushing cycle optimization instructions, which are fed back to update the knowledge base to achieve closed-loop optimization.

[0026] Step S1 includes the following sub-steps: Step S101: During the molding process, simultaneously collect discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log.

[0027] Step S101 enables multi-channel synchronous acquisition of discharge current waveform, discharge voltage waveform, dielectric conductivity, temperature, electrode wear image and control log, ensuring comprehensive coverage of machine operation information.

[0028] Step S102: Add the machine number, workpiece number, and acquisition timestamp to the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log.

[0029] The discharge current waveform is used to reflect the energy and stability of a single discharge, the discharge voltage waveform is used to determine the characteristics of the gap breakdown process, the dielectric conductivity data is used to characterize the degree of contamination and ion concentration changes of the processing medium, the temperature data is used to monitor the thermal state of the discharge zone and the machine environment, the electrode wear image is used to present the morphology of pits, erosion and defects on the electrode end face, and the control log is used to record the processing parameter adjustment and operation instructions.

[0030] Step S102 adds machine number, workpiece number and collection timestamp to the collected multi-source data to form a data stream with source fingerprint, ensuring data traceability and consistency in subsequent analysis.

[0031] Step S1, by simultaneously collecting and identifying multi-source operating data during the molding machine processing, can comprehensively reflect the discharge behavior, dielectric state, temperature environment, and electrode wear, and correspond to the operating instructions, thereby establishing a complete data foundation and providing reliable support for subsequent preprocessing, feature extraction, and fault tracing.

[0032] Step S2 includes the following sub-steps: Step S201, analyze the discharge current waveform. and discharge voltage waveform Denoising and amplitude standardization are performed.

[0033] Step S201 effectively filters out electromagnetic interference and fluctuations caused by equipment jitter by denoising and amplitude standardizing the discharge current waveform and discharge voltage waveform, ensuring the comparability of amplitudes of signals from different batches.

[0034] Step S202: Perform outlier removal on the dielectric conductivity data and temperature data.

[0035] Step S202 performs outlier removal on the dielectric conductivity data and temperature data, eliminating abnormal points caused by sensor drift, instantaneous impact or measurement error, so that the data can truly reflect the dielectric contamination and temperature change trend of the discharge zone.

[0036] Step S203: Perform edge enhancement and texture extraction processing on the electrode wear image.

[0037] Step S203 performs edge enhancement and texture extraction processing on the electrode wear image to make the erosion pits and fine textures on the electrode end face clearer, which facilitates the accurate capture of subsequent wear and defect features.

[0038] Step S204: Synchronize the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log using a unified timestamp alignment mechanism.

[0039] Step S204 aligns all processed data using a unified timestamp to achieve a one-to-one correspondence between discharge events, environmental states, image features, and operation logs in the time dimension, forming a multimodal synchronous data system.

[0040] Step S2 performs denoising, standardization, outlier removal, image texture enhancement, and time alignment on different types of fault-related data. This eliminates interference and errors during the acquisition process and ensures that various types of data establish corresponding relationships in the same time dimension. This provides a high-quality, comparable, and consistent data foundation for subsequent feature extraction and causal inference.

[0041] Step S3 includes the following sub-steps: Step S301: Based on the hybrid depth model, feature recognition is performed on the discharge current waveform and voltage waveform to obtain the distribution of single discharge duration.

[0042] Discharge current waveform after preprocessing and discharge voltage waveform The process is handled by setting a sliding window (e.g., 1s–5s), and event detection is performed within each sliding window.

[0043] Define the start and end points of candidate discharge events: starting point : The current exceeds the threshold And the first sampling point where the voltage drops significantly from the open circuit level (or the first-order difference exceeds the threshold); end Current drops back to The following and continue for at least One sampling point.

[0044] Get the first Sub-event fragment: Duration .

[0045] Eliminate obvious false events: for example or (duration threshold) and (Given by the machine specifications).

[0046] The structure and inference process of the hybrid deep model are as follows: Input format: For each candidate event segment, extract a segment of length [length missing]. The sequence (with zeros padded or interpolated if the length is insufficient) is used to form a two-dimensional sequence. .

[0047] The logic of the hybrid depth model is as follows: 1D-CNN extracts local waveform morphological features (peaks, plateaus, and drops); BiLSTM models the internal temporal dependencies of events; The attention layer weights critical moments (breakdown instant, sustaining phase, and arc-extinguishing phase); The output header is the probability of event validity. and the duration of direct regression .

[0048] Summarize all valid events within the preset statistics window. Output: Histogram / kernel density distribution vector (divided according to preset bins); It also outputs the distribution statistics: mean, variance, P50 / P90, and skewness.

[0049] The output yields the distribution of the duration of a single discharge (distribution vector and distribution statistics).

[0050] Step S301 involves feature recognition of the discharge current waveform and voltage waveform to obtain the duration distribution of a single discharge, which is used to measure the stability and breakdown law of the discharge process.

[0051] Step S302: Calculate the proportion of abnormal discharge events in the signal to form an abnormal spark ratio index.

[0052] For each valid event Calculate the characteristics of basic events: Peak current Minimum voltage ; Event energy ; Sustaining fluctuations: or The standard deviation of the event midway.

[0053] Alternatively, anomaly detection can be implemented in two equivalent ways: Rule-based: Mark an exception if any condition is met. Exceeding the normal range; or Exceeding the limit; or Too low accompanied by Too high; The event involved multiple instances of secondary breaches of the spike.

[0054] Learning classifiers: use Input a lightweight classification model (GBDT / MLP are both acceptable) and output anomaly probabilities. Anomalies are identified using threshold values, where... for standard deviation for The standard deviation.

[0055] Define within the statistics window: Total number of valid events Number of abnormal events ; Abnormal spark ratio index .

[0056] It can also output the proportion of the anomaly type (short circuit, arcing, energy overload) to facilitate subsequent root cause analysis.

[0057] Step S302 calculates the proportion of abnormal discharge events to form an abnormal spark ratio index, which is used to determine whether there are abnormal situations such as short circuits or open circuits during the discharge process.

[0058] Step S303: Extract the plasma channel stability index from the discharge current waveform and voltage waveform.

[0059] For each event segment, the initial breakdown segment and the final arc-extinguishing segment are removed, and only the intermediate stable segment is taken: The stable segment can be defined as the sampling points in the middle of the event, from 20% to 80%, or automatically located by the interval where the rate of change of current or voltage is below a threshold.

[0060] Calculate for each stable segment of the event: Dynamic impedance (Ratio of mean to stable segment) Within the stable phase of a single discharge event average value, Within the stable phase of a single discharge event The average value; Voltage-current coupling correlation coefficient (Stability segment correlation); Energy stability: power in the steady-state range coefficient of variation .

[0061] Summarize event-level quantities into window-level stability metrics: ; ; .

[0062] Provide a stability index that is 0–1 normalized:

[0063] in, , , Calibrated based on historical normal operating conditions.

[0064] Step S303 involves the combined extraction of plasma channel stability indicators to reflect energy fluctuations during the formation and maintenance of the discharge channel.

[0065] Step S304: Analyze the pit texture and edge morphology of the electrode wear image and output the uniformity and defect degree.

[0066] The electrode wear image processed in step S203 is scaled, and the image is divided into blocks according to a fixed grid (e.g., M×N tiles). Each block is calculated independently, and finally the results are summarized into a global index.

[0067] Optionally, perform pit region segmentation: Traditional segmentation: Adaptive thresholding and morphological opening / closing operations are used to obtain the pit mask; Learning segmentation: Use U-Net or DeepLab to perform pixel-level segmentation of the pits, with the training set derived from manually drawn images.

[0068] Edge morphology analysis: Calculate for each tile: pit density Average pit area, maximum pit area, and area distribution quantiles; Texture features: GLCM contrast, LBP histogram; Edge morphology: Perform Canny edge detection on the electrode edge region to calculate edge roughness, curvature statistics, and number of notches.

[0069] Output uniformity and defect level: Uniformity Based on the degree of dispersion of pit density or texture energy between tiles: ; Defect level : Weighted summation of the proportion of large pits, edge gaps, and roughness.

[0070] Step S304 performs texture and edge analysis on the electrode wear image and outputs the uniformity of wear and defect features of the electrode end face.

[0071] Step S305: Extract defect features from the processed surface image. The defect features include burns, collapses, and microcracks.

[0072] Perform illumination correction (flat field correction / white balance) on the processed surface image and establish benchmark features with normal samples (same material, same process parameters).

[0073] Define the ROI (Relevant Area of ​​Interest) to avoid mistaking fixtures or reflections for defects.

[0074] The extraction logic for burns, collapses, and microcracks is as follows: Burns: manifested as areas of abnormal brightness, color, or texture; Optional: Traditional method: Calculate the grayscale mean, contrast, and local entropy within the ROI, and identify areas whose difference from the baseline exceeds a threshold as burn candidates; Learning method: Lightweight segmentation network outputs burn mask.

[0075] Output features: burn area ratio, maximum connected region area, burn intensity grading (based on mean deviation).

[0076] Collapse: manifested as abnormal boundary shape or localized pit and depression outline; If only a 2D image is available: use shadow / reflection changes + morphological detection to detect concave areas (commonly used in engineering). If a low-cost enhancement could be added: take photos from multiple angles at the same location, and perform a simple brightness consistency estimate to obtain the relative depth level.

[0077] Output features: number of collapsed areas, area ratio, and maximum collapse scale (equivalent diameter).

[0078] Microcracks: fine, elongated linear textures; Optional: Traditional method: Gabor filtering or Frangi line structure enhancement, followed by binarization and thinning (skeleton) processing, and finally connected component filtering; Learning method: The crack detection network outputs a crack mask and a skeleton.

[0079] Output characteristics: number of cracks, total length, maximum length, number of branches, and directional distribution.

[0080] Step S305 extracts defect features from the processed surface image and identifies processing abnormalities such as burns, collapses, and microcracks.

[0081] Step S306: Combine the single discharge duration distribution, abnormal spark ratio index, plasma channel stability index, uniformity and defect degree, and defect characteristics to form a multidimensional fault factor vector.

[0082] After being generated, the multidimensional fault factor vectors are uniformly stored in the fault feature library, which saves the original vectors and the corresponding acquisition conditions, machine number and time tag.

[0083] Concatenate and normalize the following results in a fixed order: Single discharge duration distribution vector ; Abnormal spark ratio index ; Plasma channel stability index ; Uniformity and Defect Level ; Defect Feature Set .

[0084] For each feature dimension, use the mean and variance of historical normal operating conditions to perform a z-score, or perform a 0–1 min-max (parameters are from the cross-machine knowledge base).

[0085] If a window is missing an image (not captured or blurry), then: Preserve the features from the last valid image (forward padding) and add missing bits; Alternatively, set the dimension to its default value and append an image availability flag to the end of the vector.

[0086] Output a multi-dimensional fault factor vector, along with: machine number, workpiece number, timestamp, and operating condition segment identifier (from control log parsing).

[0087] Step S306 merges all the above features to form a multi-dimensional fault factor vector, realizing a comprehensive multi-angle representation of the fault state, and stores it uniformly in the fault feature library to ensure traceability and consistency of cross-machine comparison.

[0088] Step S3, through deep feature recognition and fusion of multi-source data, extracts multi-dimensional discharge features and fault factors from current, voltage signals, and electrode and machining surface images, comprehensively characterizing the stability, abnormality, and defect status during the machining process. The resulting multi-dimensional fault factor vector is uniformly stored in the fault feature library, providing high-quality basic data for similarity comparison, causal inference, and cross-machine knowledge transfer.

[0089] Step S4 includes the following sub-steps: Step S401: Establish a cross-machine operation knowledge base, which includes historical fault factor vectors and labeled fault category information of different machines under various operating conditions.

[0090] Each historical sample is defined as a fault sample record, which must contain at least: Multidimensional failure factor vector; Fault category labels: such as short circuit, arcing, unstable discharge, abnormal flushing, abnormal electrode wear (existing maintenance codes of the enterprise can be used); Machine and operating condition metadata: machine number, control parameters (pulse width, peak current setting, gap servo parameters, flushing cycle), workpiece material, processing stage, timestamp, operator shift; Evidence chain pointer: corresponds to the original waveform segment ID, image ID, and log segment ID.

[0091] Each sample is mapped to the same statistical window (discharge events are counted once every 30 seconds) to ensure that the dimensions of the fault factor vector are consistent.

[0092] First, normalize the mean and variance estimated within each machine using historical normal segments, and then normalize it a second time using global statistics.

[0093] Only samples that are confirmed to be repaired, shut down, or manually verified are considered strong labels, while uncertain samples are labeled as weak labels, and their weights are reduced during training.

[0094] Establish a tag index (binding by fault category tags) and a vector index (establish a nearest neighbor retrieval structure (such as an ANN vector index) for the normalized multidimensional fault factor vectors, supporting Top-K similarity retrieval).

[0095] Record the knowledge base version number and the version number currently used for online diagnostics, and output the knowledge base for cross-machine operation.

[0096] Step S401 establishes a cross-machine operation knowledge base, which collects historical fault factor vectors and labeled fault categories of different machines under various operating conditions, providing a systematic data foundation for comparison and migration.

[0097] The predicted failure probability is obtained by weighted voting based on the similarity of Top-K similar historical samples and then calibrated for probability. At the same time, similar sample identifiers are output to form traceable evidence.

[0098] Obtain the normalized multidimensional fault factor vector of the current machine in a certain statistical window, and check whether the dimensions are complete, whether the missing flag exists, and whether it is in a diagnosable operating condition (which can be determined by the control log: processing start / stop, fluid flushing start, and parameter switching).

[0099] Use vector indexing to retrieve the Top-K historical vectors in the knowledge base that are closest to the normalized multidimensional fault factor vector; Alternatively, a specific similarity metric can be defined: Cosine similarity (suitable for normalized vectors), or Mahalanobis distance (considering feature correlation).

[0100] Alternatively, convert the Top-K label information into probabilities, using at least one feasible method: Method 1: For each fault category Calculate its score : ; in, This represents the number of similar samples retrieved from the cross-machine knowledge base. This represents the sequence number of the similar samples. This is the multidimensional fault factor vector generated for the current machine in the current window. For the first in the knowledge base A multidimensional failure factor vector of historical records. For the first The fault category labels of each historical sample are stored in the knowledge base.

[0101] Normalization yields the predicted failure probability : ; in, For all candidate fault categories.

[0102] Method 2: Use historical validation sets for Platt calibration, map the scores to calibration probabilities, and avoid overestimating the predicted failure probability.

[0103] The output should include at least: the predicted fault category (with the highest probability), the predicted fault probability, the Top-N candidate categories and their probabilities, and supporting evidence (Top-K similar sample IDs, similarity, and corresponding operating conditions).

[0104] Step S402 compares the similarity of the multidimensional fault factor vector generated by the current machine with the historical factors in the knowledge base, outputs the predicted fault probability, and realizes real-time diagnosis of the operating status.

[0105] Step S403: Combine the comparison results with the contrastive learning model to extract common features and difference features, identify and classify rare fault types outside of common patterns. When the embedding distance between a sample and any known fault category exceeds the threshold, it is judged as a rare fault. After forming a new category through clustering and manual confirmation, it is written back to the knowledge base.

[0106] Using samples from the knowledge base as the training set, for each anchor sample structure: Positive samples Samples with the same fault category and similar operating conditions (same material, similar parameter levels, or similar processing stages); negative samples Samples of different fault categories, or samples of the same type but with vastly different operating conditions.

[0107] If the labels are incomplete, weakly labeled samples only participate in positive samples of the same working condition cluster and are not used as strong supervision.

[0108] Using a representation network Mapping the multidimensional failure factor vector to an embedding vector .

[0109] Training objective: To make the embeddings of similar samples closer together and those of dissimilar samples further apart (using triplet loss).

[0110] Output the embedding vectors and the interpretation results of common and differential features: Common characteristics: Stable and consistent dimensions among the nearest neighbors of the same type; Distinctive features: The dimension in which the current sample deviates the most from its nearest neighbors of the same type.

[0111] Perform feature difference analysis on the Top-K nearest neighbors and sort them by absolute difference. Take the top M dimensions as the list of differential features.

[0112] For each known category, the class center is calculated in the embedding space. The distance from the current sample to the nearest class center is calculated. If the distance from the current sample to the nearest class center is greater than a preset threshold (the threshold is determined by the historical intra-class distance quantile), it is judged as a rare fault type.

[0113] When the fault type is determined to be rare, execute: Temporary clustering: Based on DBSCAN, cluster samples corresponding to rare fault types in the embedding space in the recent period to form temporary cluster IDs, and output the representative samples of the cluster and their list of differential features, so that maintenance personnel can add rare class labels.

[0114] Write the confirmed category, cluster ID, and evidence chain pointer back to the cross-machine running knowledge base, and trigger the next round of comparative learning incremental training.

[0115] Step S403 combines the comparative learning model to conduct in-depth analysis of the results, extract common and differential features, identify rare fault types outside of common patterns and classify them reasonably, thereby improving the ability to detect and recognize abnormal patterns.

[0116] Step S4 establishes a cross-machine operation knowledge base and compares and analyzes the current machine fault factor vector, which enables knowledge transfer and sharing between different machines, thereby improving the ability to identify low-frequency faults and new faults, and providing cross-machine experience support for fault prediction.

[0117] Step S5 includes the following sub-steps: Step S501: Construct a causal relationship graph by establishing a causal dependency chain from the multidimensional fault factor vector, dielectric conductivity, temperature data, and control logs.

[0118] The input data for step S501 are a multidimensional fault factor vector, dielectric conductivity, temperature, and control log (time series of control parameters obtained by parsing, including pulse width, peak current, gap servo setting, and flushing cycle). The above data are numbered according to the same time axis window to generate a feature table.

[0119] The nodes of the causal relationship graph should be defined into at least three categories: Fault factor nodes: Key dimensions derived from multidimensional fault factor vectors, including the proportion of abnormal sparks, plasma channel stability, average discharge duration, burn area ratio, and total crack length; Environmental nodes and dielectric nodes: temperature, dielectric conductivity, and corresponding rate of change; Control node: Parameters and action events obtained from log parsing. Parameters include control parameter time series, and events include parameter switching, start / stop, and alarm reset (which can be encoded as 0 / 1).

[0120] Establish temporal causal constraints and structural prior constraints: Temporal causality constraint: Only points from past windows to the current window are allowed, avoiding cross-referencing at the same time.

[0121] Structural prior constraints: Control parameters and medium conditions are usually considered upstream causes, prohibiting defects from causing unreasonable edges such as control parameters in the reverse direction.

[0122] The method for determining the edges is as follows: For variables A and B, calculate the correlation or regression gain under lag: Compare the error decrease of predicting the ratio of variable B within the time window using only variable B with the error decrease of predicting the ratio of variable B within the time window after adding variable A. If the error decrease exceeds a preset threshold, add the candidate causal edge A→B.

[0123] For each edge A→B, calculate the edge weight (using the absolute value of the regression coefficient or mutual information), and represent the causal relationship graph as a set of nodes (the above variables), a set of edges (direction + edge weight), and a causal dependency chain (a set of directed paths from the upstream node to the fault manifestation node).

[0124] Step S501 establishes a causal dependency chain from the multidimensional fault factor vector, dielectric conductivity, temperature data, and control logs to form a complete causal model, ensuring that the interaction relationships between different features are expressed.

[0125] Step S502: Identify the evolution path of the fault from the initial cause to the final manifestation based on the causal dependency chain.

[0126] Define the target node for the final performance and the initial trigger: Final performance node: can be set as the key performance indicator corresponding to the predicted fault category; Initial trigger candidate nodes: sudden change in control parameters, change in flushing cycle, increase in medium conductivity, and temperature rise.

[0127] Select the current diagnosis time In a causal relationship graph, the final node is taken as the endpoint, starting from... Backtracking For each candidate causal path, a path score is calculated within a window. : ; in, For edge weights, For the upstream node of this edge The degree of abnormality (e.g., number of z-score overthresholds, slope overthresholds, and mutation point detection results). The top-N paths with the highest scores are selected as the evolutionary paths.

[0128] For each node, an anomaly index is calculated using CUSUM within the backtracking window. When the change exceeds the threshold, it is recorded as 1. The anomaly level is normalized to 0–1 and used as a component of the anomaly level.

[0129] Output the evolution path (node ​​sequence, time range, and path score) and the degree of anomalousness of each node.

[0130] Step S502 identifies the evolution path of the fault from the initial cause to the final manifestation based on the causal dependency chain, clearly presents the key nodes and turning points, and reveals the logical sequence of the occurrence and development of the fault.

[0131] Step S503: Output the root cause localization conclusion, which includes the fault cause, evolution chain, and impact results, and write the conclusion into the diagnostic record.

[0132] For Top-N evolution paths, nodes that satisfy the following conditions are selected as root cause candidates: Upstream: Located near the starting point of the path (the foremost medium node); Interpretability: Strongest evidence of anomalousness (longest duration of anomalousness or most significant mutation); Interventional factors: Prioritize control parameters or media conditions as operable reasons.

[0133] Root cause output can be a single root cause or a primary root cause plus a secondary root cause, such as: The primary cause is the increased conductivity of the medium due to the prolonged flushing cycle.

[0134] Secondary cause: A larger pulse width amplifies energy fluctuations, leading to decreased stability.

[0135] Structured fields are generated based on the cause of the failure, the evolution chain, and the impact on the output: Fault causes: root cause node name, anomaly type, and anomaly timestamp; Evolutionary chain: node sequence (including edge directions), weight of each edge segment, and key anomaly evidence; Impact on results: final performance and its corresponding window.

[0136] Write the following into the diagnostic log: Machine number, workpiece number, diagnosis time, knowledge base version number, root cause localization conclusion (fault causes, evolution chain and impact results), evidence pointers (referenced original waveform segment ID, image ID, control log segment ID, and Top-K similar sample ID).

[0137] The causal relationship graph is constructed under time lag constraints and process prior constraints, and the evolution path is sorted by combining edge weights and anomalous evidence, thereby achieving traceable root cause localization.

[0138] Step S503 outputs the root cause localization conclusion, which includes the fault cause, evolution chain and impact results, and writes it into the diagnostic record to achieve the interpretability and traceability of the results, providing a basis for subsequent optimization and knowledge base updates.

[0139] Step S5, by constructing a causal relationship diagram and reasoning about the fault evolution process, can reveal the causal chain between multidimensional fault factors and media state, environmental parameters and control operations, clarify the evolution logic of the fault from the initial cause to the final manifestation, thereby achieving interpretable root cause localization and forming a traceable diagnostic record.

[0140] Step S6 includes the following sub-steps: Step S601: When the predicted failure probability exceeds the preset threshold, generate real-time warning information and input the real-time warning information into the operation interface.

[0141] Step S601: When the predicted failure probability exceeds the preset threshold, the system automatically generates real-time early warning information and prompts the operator through the operation interface to ensure that potential risks can be detected and responded to in a timely manner.

[0142] Step S602: Based on the multi-dimensional fault factor vector generated in step S3 and the root cause localization conclusion output in step S5, an adaptive optimization command is sent to the machine control terminal. The optimization command includes: The pulse width parameter is fine-tuned based on the discharge current waveform and discharge voltage waveform acquired in step S1 and preprocessed in step S2.

[0143] The energy is reduced based on the stability index of the plasma channel.

[0144] The flushing cycle is extended based on the dielectric conductivity and temperature data processed in step S2.

[0145] Step S602, based on the multidimensional fault factor vector in step S3 and the root cause localization conclusion in step S5, sends optimization instructions to the machine control terminal. Specifically, it includes: reducing discharge fluctuations by fine-tuning the pulse width of the discharge current and voltage waveforms; reducing energy based on the stability index of the plasma channel to avoid excessive ablation; and extending the flushing cycle based on the dielectric conductivity and temperature data to improve the impurity removal efficiency, thereby improving the processing environment.

[0146] Step S603: After the optimization instruction is implemented, compare the pre-optimization running data formed in steps S1 to S3 with the new data after optimization, and feed the difference results and the control log record of step S1 back to the knowledge base.

[0147] Step S603 involves comparing the operating data before and after optimization after implementing the optimization instructions, and feeding the difference results back to the knowledge base in conjunction with the control log. This enables dynamic updates of fault modes and optimization strategies, forming a closed-loop optimization mechanism shared across machines.

[0148] Step S6 uses a real-time early warning and closed-loop optimization mechanism. The system can not only promptly alert operators when potential risks occur, but also issue adaptive optimization instructions to the machine based on the results of the preceding analysis, and provide feedback and updates on the optimization effect, thereby achieving continuous iterative fault prevention and operation optimization.

[0149] The multidimensional fault factor vector generated in step S3 is compared with the cross-machine knowledge base in step S4 to output the fault factor similarity matrix. Dynamic optimization is achieved by continuously iterating and updating the fault factor similarity matrix, recording the similarity of fault factors between different machines, and automatically adjusting the matching weight each time a new fault factor vector is added.

[0150] By comparing the multidimensional fault factor vector generated in step S3 with the cross-machine knowledge base in step S4 and outputting a fault factor similarity matrix, the similarity relationship of fault factors between different machines can be dynamically reflected. This matrix is ​​continuously iterated and updated during operation to ensure the real-time performance and accuracy of the comparison results. By continuously adjusting the matching weights, not only can newly emerging fault modes be captured in a timely manner, but the consistency and applicability of cross-machine diagnostic results can also be maintained, thereby improving the detection and migration capabilities for low-frequency and novel faults.

[0151] Example 2, refer to Figure 2It provides a fault data analysis system for EDM machines based on cross-device migration, including a data acquisition module, a data processing module, a feature extraction module, a knowledge analysis module, a chain inference module, and an early warning and optimization module.

[0152] The data acquisition module is used to collect the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs of the molding machine.

[0153] The data processing module is used to preprocess and align the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs.

[0154] The feature extraction module is used to extract discharge characteristics and fault factors from the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs.

[0155] The knowledge analysis module is used to establish a cross-machine operation knowledge base and perform transfer analysis on discharge characteristics and fault factors, and output the fault prediction probability.

[0156] The chain inference module is used to construct a causal relationship graph and establish causal dependency chains, identify evolutionary paths, and output root cause location conclusions.

[0157] The early warning optimization module is used to perform early warning and closed-loop optimization based on the probability of fault prediction.

[0158] The intelligent fault data analysis method for molding machines proposed in this invention can overcome the shortcomings of existing technologies. It achieves comprehensive coverage of multi-source data by simultaneously collecting discharge current waveforms, discharge voltage waveforms, dielectric conductivity, temperature data, electrode wear images, and control logs, along with machine number, workpiece number, and timestamp. Through denoising, standardization, outlier removal, and image texture extraction, and by establishing a unified time alignment mechanism, it ensures that various data types have corresponding relationships within the same time dimension. A hybrid depth model is used to identify discharge duration distribution, abnormal spark ratio, plasma channel stability indicators, and electrode end wear characteristics, forming a multi-dimensional fault factor vector for subsequent analysis. The analysis provides high-quality input, enabling experience sharing among different machines based on similarity comparison and contrastive learning models. It effectively identifies rare or newly emerging fault modes, constructs causal chains of discharge characteristics, medium status, and control logs, reveals fault evolution paths, and outputs clear fault causes and key nodes, facilitating operators to quickly locate problems. When the predicted fault probability exceeds the threshold, the system automatically generates an early warning and issues optimization instructions, specifically adjusting pulse width, energy, and flushing cycle. The optimization results are fed back to the knowledge base, forming a dynamically updated iterative mechanism. Through continuous iteration and updating of the fault factor similarity matrix, it achieves real-time comparison and expansion of fault modes across machines, ensuring the accuracy and applicability of the diagnosis.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fault data analysis method for EDM machines based on cross-device migration, characterized in that, Includes the following steps: Step S1: Collect the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log of the molding machine; Step S2 involves preprocessing and aligning the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs. Step S3: Extract the discharge characteristics and fault factors from the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control log. Step S4: Establish a cross-machine operation knowledge base and perform transfer analysis on discharge characteristics and fault factors to output fault prediction probability. Step S5: Construct a causal relationship graph and establish causal dependency chains, identify evolutionary paths, and output root cause localization conclusions; Step S6: Based on the fault prediction probability, perform early warning and closed-loop optimization.

2. The method for analyzing fault data of EDM machines based on cross-device migration as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: During the molding process, simultaneously collect discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log. Step S102: Add the machine number, workpiece number, and acquisition timestamp to the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log.

3. The method for analyzing fault data of EDM machines based on cross-device migration as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S201: Denoise and amplitude standardization are performed on the discharge current waveform and discharge voltage waveform; Step S202: Perform outlier removal on the dielectric conductivity data and temperature data; Step S203: Perform edge enhancement and texture extraction processing on the electrode wear image; Step S204: Synchronize the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log using a unified timestamp alignment mechanism.

4. The method for analyzing fault data of an EDM machine based on cross-device migration as described in claim 3, characterized in that, Step S3 includes the following sub-steps: Step S301: Based on the hybrid depth model, feature recognition is performed on the discharge current waveform and voltage waveform to obtain the distribution of single discharge duration; Step S302: Calculate the proportion of abnormal discharge events in the signal to form an abnormal spark ratio index; Step S303: Extract plasma channel stability indicators from the discharge current waveform and voltage waveform. Step S304: Analyze the pit texture and edge morphology of the electrode wear image and output the uniformity and defect degree. Step S305: Extract defect features from the processed surface image, the defect features including burns, collapses and microcracks; Step S306: Combine the single discharge duration distribution, abnormal spark ratio index, plasma channel stability index, uniformity and defect degree, and defect characteristics to form a multidimensional fault factor vector.

5. The fault data analysis method for EDM machines based on cross-device migration as described in claim 4, characterized in that, Step S4 includes the following sub-steps: Step S401: Establish a cross-machine operation knowledge base, which includes historical fault factor vectors and labeled fault category information of different machines under various operating conditions; The predicted failure probability is obtained by weighted voting based on the similarity of Top-K similar historical samples and then calibrated for probability. At the same time, similar sample identifiers are output to form traceable evidence. Step S403: Combine the comparison results with the contrastive learning model to extract common features and difference features, identify and classify rare fault types outside of common patterns. When the embedding distance between a sample and any known fault category exceeds the threshold, it is judged as a rare fault. After forming a new category through clustering and manual confirmation, it is written back to the knowledge base.

6. The method for analyzing fault data of an EDM machine based on cross-device migration as described in claim 5, characterized in that, Step S5 includes the following sub-steps: Step S501: Construct a causal relationship graph and establish a causal dependency chain by combining the multidimensional fault factor vector, dielectric conductivity, temperature data and control logs; Step S502: Identify the evolution path of the fault from the initial cause to the final manifestation based on the causal dependency chain; Step S503: Output the root cause localization conclusion, which includes the fault cause, evolution chain and impact result, and write the conclusion into the diagnostic record.

7. The method for analyzing fault data of an EDM machine based on cross-device migration as described in claim 6, characterized in that, Step S6 includes the following sub-steps: Step S601: When the predicted failure probability exceeds a preset threshold, generate real-time warning information and input the real-time warning information into the operation interface. Step S602: Based on the multi-dimensional fault factor vector generated in step S3 and the root cause localization conclusion output in step S5, an adaptive optimization command is sent to the machine control terminal. The optimization command includes: The pulse width parameter is fine-tuned based on the discharge current waveform and discharge voltage waveform acquired in step S1 and preprocessed in step S2. The energy required to perform the operation is reduced based on the stability index of the plasma channel. Extend the flushing cycle based on the dielectric conductivity and temperature data processed in step S2; Step S603: After the optimization instruction is implemented, compare the pre-optimization running data formed in steps S1 to S3 with the new data after optimization, and feed the difference results and the control log record of step S1 back to the knowledge base.

8. The method for analyzing fault data of an EDM machine based on cross-device migration as described in claim 7, characterized in that, After being generated, the multidimensional fault factor vectors are uniformly stored in the fault feature library, which saves the original vectors and the corresponding acquisition conditions, machine numbers and time tags.

9. The method for fault data analysis of EDM machines based on cross-device migration as described in claim 8, characterized in that, The multidimensional fault factor vector generated in step S3 is compared with the cross-machine knowledge base in step S4 to output the fault factor similarity matrix. Dynamic optimization is achieved by continuously iterating and updating the fault factor similarity matrix, recording the similarity of fault factors between different machines, and automatically adjusting the matching weight each time a new fault factor vector is added.

10. A fault data analysis system for EDM machines based on cross-device migration, which is applied to the fault data analysis method for EDM machines based on cross-device migration as described in any one of claims 1-9, characterized in that, It includes a data acquisition module, a data processing module, a feature extraction module, a knowledge analysis module, a chain inference module, and an early warning and optimization module; The data acquisition module is used to acquire the discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear image, and control log of the molding machine. The data processing module is used to preprocess and align the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs. The feature extraction module is used to extract discharge characteristics and fault factors from the collected discharge current waveform, discharge voltage waveform, dielectric conductivity data, temperature data, electrode wear images, and control logs. The knowledge analysis module is used to establish a cross-machine operation knowledge base and perform transfer analysis on discharge characteristics and fault factors, and output fault prediction probability. The chain inference module is used to construct a causal relationship graph and establish a causal dependency chain, identify evolutionary paths, and output root cause location conclusions; The early warning optimization module is used to perform early warning and closed-loop optimization based on the fault prediction probability.

Images