Intelligent multi-head wire cutting processing control method and system based on internet of things

Abstract

The application discloses a kind of based on Internet of Things intelligent multi-head wire cutting processing control method and system, it is related to numerical control processing and Internet of Things field, including the following steps: step one: through Internet of Things gateway real-time acquisition multi-dimensional running data of multi-head wire cutting processing equipment cluster;The kind of based on Internet of Things intelligent multi-head wire cutting processing control method and system, through global multimodal perception data acquisition, edge-cloud collaborative computing architecture, high-precision multi-head collaborative motion closed-loop control, discharge state real-time identification and self-adaptive adjustment, full-dimension dynamic error compensation, digital twin three-dimensional visualization mapping, predictive maintenance, global production scheduling and process library iterative optimization, completely solve traditional multi-head wire cutting equipment collaborative precision low, discharge control lag, process experience-dependent, Internet of Things fusion is insufficient, lack of intelligent optimization and predictive maintenance and other core problems.

Description

Technical Field

[0001] This invention relates to CNC machining and Internet of Things (IoT) technology, specifically to an intelligent multi-head wire EDM machining control method and system based on IoT. Background Technology

[0002] Electrical discharge wire cutting (EDM) is a non-contact special machining process that relies on high-frequency pulse discharge between the electrode wire and the workpiece. The instantaneous high temperature melts and vaporizes the local material of the workpiece, which is then flushed out by the working fluid, ultimately achieving the forming of the workpiece. This process has unique advantages such as no macroscopic cutting force, high machining accuracy, ability to machine ultra-hard, brittle, tough, and difficult-to-machine materials, suitability for machining complex irregular curved surfaces and microstructures, and no need to manufacture forming tools. It is an indispensable key core process in modern advanced manufacturing and widely supports the production of precision parts in pillar industries such as medical, traditional Chinese medicine, high-end equipment manufacturing, semiconductor aerospace, automotive, and electronic information.

[0003] Multi-head wire EDM equipment includes multiple sets of electrode wire running mechanisms, multi-axis servo motion systems, and worktable feeding mechanisms. There are problems such as mechanical coupling, vibration interference, motion lag, transmission error, and thermal deformation coupling between the various motion units. Traditional control systems adopt independent motion control or simple master-slave follow mode, which lack multi-agent collaborative planning, real-time error correction, and coupling dynamics compensation mechanisms. This easily leads to problems such as multi-head trajectory deviation, interpolation error accumulation, synchronization inconsistency, and positioning deviation. These problems directly result in excessive dimensional deviations in multi-head processing, uneven surface texture, and poor workpiece consistency, making it difficult to meet the high-precision processing requirements of fields such as traditional Chinese medicine decoction pieces, aerospace, and semiconductors. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent multi-head wire EDM machining control method and system based on the Internet of Things, in order to solve the problems of traditional control systems that adopt independent motion control or simple master-slave follow mode, lack multi-agent collaborative planning, real-time error correction, and coupled dynamics compensation mechanisms, which are prone to problems such as multi-head trajectory deviation, interpolation error accumulation, synchronization inconsistency, and positioning deviation. These problems directly lead to excessive dimensional deviations in multi-head machining, uneven surface texture, and poor workpiece consistency. It can be mainly applied to solve batch-to-batch problems in the preparation of traditional Chinese medicine decoction pieces.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent multi-head wire EDM machining control method based on the Internet of Things, comprising the following steps:

[0006] Step 1: Collect multi-dimensional operating data of the multi-head wire EDM equipment cluster in real time through the IoT gateway. The multi-dimensional operating data includes equipment status data, process parameter data, sensor detection data, processing data and energy consumption data. Perform protocol parsing, spatiotemporal alignment and standardization preprocessing on the multi-dimensional operating data to obtain a synchronous time-series operating dataset.

[0007] Step 2: Based on the pre-trained AI processing parameter adaptive model, multi-head collaborative scheduling rules and fault diagnosis rule base, the synchronous time-series running dataset is parsed and fused in parallel to generate the time-series processing instruction-state sequence of each cutting unit;

[0008] Step 3: Using the processing actions, state events, and control instructions in each time-series processing instruction-state sequence as nodes, the causal relationship of the nodes is analyzed based on the causal discovery algorithm of time constraints and conditional independence test, and a multi-station processing behavior causal graph is constructed; the multi-station processing behavior causal graph represents the dynamic influence and constraint relationship between each cutting unit, execution mechanism, and processing process;

[0009] Step 4: Input the time-series snapshot sequence of the causal graph of multi-station processing behavior into a pre-trained graph neural network and time-series network combination model. Extract the dynamic structural features of multi-head collaborative processing through message passing aggregation and graph-level representation learning. Combine the preset collaborative processing mode to classify and output multi-head collaborative control dimension labels.

[0010] Step 5: Based on the timing-based machining instructions and state sequence, match and compare them with the standard machining process library to calculate machining accuracy, machining efficiency, wire breakage probability, energy consumption utilization rate and surface quality indicators, and generate single-station process optimization dimension labels;

[0011] Step Six: Integrate the multi-head collaborative control dimension label with the single-station process optimization dimension label, associate key causal evidence in the multi-station processing behavior causal graph, generate intelligent multi-head wire EDM full-domain control instructions and processing optimization reports, and send them to the corresponding execution units through the Internet of Things to complete closed-loop control.

[0012] Further, in step one, the following steps are included: using OPCUA, Modbus, and MQTT industrial communication protocols to uniformly access multiple multi-head wire EDM machines, sensor modules, execution drive modules, and cooling systems; collecting raw data such as processing voltage, processing current, wire feed speed, tension value, coolant concentration, temperature, pulse width, pulse interval, station coordinates, and processing progress at a high-frequency sampling rate of 1MHz; performing outlier removal, missing value interpolation, and noise filtering on the raw data; completing spatiotemporal alignment based on a unified time reference; and converting heterogeneous data into time-series data in a unified format; and performing frame segmentation and feature extraction on the time-series data according to a preset time window to form a synchronous time-series running dataset containing equipment identifiers, timestamps, station numbers, and feature vectors.

[0013] Further, in step two, the following steps are included: assigning a continuous and unique identifier to each cutting unit using a target tracking model on the video stream and image sensing data; extracting the continuous position and posture sequences of the electrode wire, worktable, cooling nozzle, and workpiece; inputting the position and posture sequences into a pre-trained temporal action classification model to identify the atomic action types of processing feed, idle travel, reversal, cutting, pause, and emergency stop within a fixed short temporal window, generating a visual modal processing action stream; performing voice activity detection and sound source separation on audio and discharge sound signals, associating them with the corresponding cutting unit, and extracting abnormal discharge, collision, and abnormal vibration features through automatic speech recognition and semantic analysis to generate an audio modal event stream; parsing current, voltage, tension, temperature, and liquid level sensor data streams based on a rule engine, mapping them to normal processing, abnormal discharge, wire breakage precursors, over-temperature, over-pressure, and coolant abnormal operation events, generating a sensor modal event stream; and performing spatiotemporal correlation and evidence fusion on the visual modal processing action stream, audio modal event stream, and sensor modal event stream based on timestamps and spatial locations to generate a temporal processing instruction-state sequence for each cutting unit.

[0014] Further, step three includes: treating processing actions, state events, and control instructions in the time-series processing instruction-state sequence as nodes, and treating equipment number, workstation number, action type, event type, timestamp, and parameter value as node attributes; performing time-series filtering on nodes based on timestamps, and identifying node pairs whose time differences fall within a preset causal time window as candidate causal node pairs; employing a constraint-based causal discovery algorithm on the candidate causal node pairs, calculating independence statistics through kernel conditional independence tests when given a set of conditions, determining the existence and direction of candidate causal edges, and obtaining a preliminary causal structure; based on the preliminary causal structure, calculating causal edge weights through historical co-occurrence frequency, sequence dependency strength, and conditional mutual information to obtain an estimated causal strength value; and constructing a causal graph of multi-workstation processing behavior based on nodes, node attributes, causal edges, and the estimated causal strength value.

[0015] Step four includes: slicing the causal graph of multi-station processing behavior into fixed-duration slices to obtain a sequence of time-series snapshots; learning node representations for each time-series snapshot using a graph attention network, combining initial node features, neighbor features, and causal strength estimates, and aggregating node embedding vectors containing local structural information through an attention mechanism; aggregating all node embedding vectors using a graph-level readout function to generate a graph global representation vector representing the current collaborative state; inputting the graph global representation vectors of continuous time-series snapshots into a long short-term memory network to capture the temporal evolution features of collaborative patterns, obtaining a team collaborative state encoding vector; inputting the collaborative state encoding vector into a pre-trained multi-label collaborative pattern classifier to output collaborative pattern labels and confidence scores; the collaborative pattern labels include efficient synchronous processing, predictive assisted feeding, workstation conflict, resource competition, communication redundancy, and chain reaction risk of broken wires; when the confidence score exceeds a set threshold, generating corresponding multi-head collaborative control dimension labels, and backtracking to locate the key subgraphs and key nodes in the causal graph of multi-station processing behavior that cause anomalies.

[0016] Furthermore, step five includes: comparing the timing-based machining instructions-state sequence with the standard SOP process, standard process parameter library, and standard machining trajectory frame by frame and action by action; calculating the action deviation, parameter deviation, timing deviation, and trajectory deviation between the actual machining and standard machining; calculating the machining accuracy compliance rate, machining efficiency, theoretical wire breakage probability, energy consumption utilization rate, and surface roughness prediction value based on the deviation values; and generating single-station process optimization dimension labels based on the calculation results, including labels such as reasonable parameters, excessively high parameters, excessively low parameters, excessively fast feed, excessively slow feed, insufficient cooling, excessive cooling, abnormal tension, easy wire breakage, high precision, low efficiency, and high energy consumption.

[0017] Step six includes: weighted fusion of multi-head collaborative control dimension labels and single-station process optimization dimension labels to form a comprehensive evaluation label system; mapping the comprehensive evaluation labels to a causal graph of multi-station processing behavior to locate key causal paths, key instructions, key parameters, and key stations leading to anomalies or inefficiencies; generating global control instructions based on key causal evidence, including synchronous start / stop, group scheduling, automatic parameter correction, feed rate adjustment, tension compensation, coolant ratio adjustment, fault isolation, and early warning prompts; distributing global control instructions to corresponding equipment, stations, and actuators via the Internet of Things to achieve real-time closed-loop control; and generating a processing optimization report that includes a processing overview, collaborative evaluation, single-station evaluation, key issues, optimization suggestions, and control logs, supporting visualization and report export.

[0018] An intelligent multi-head wire EDM machining control system based on the Internet of Things includes:

[0019] The IoT data acquisition module is used to collect multi-dimensional operational data of the multi-head wire cutting equipment cluster in real time through the IoT gateway. It performs protocol parsing, spatiotemporal alignment and standardization preprocessing on the multi-dimensional operational data to obtain a synchronous time-series operational dataset.

[0020] The multimodal time series parsing module is used to perform parallel parsing and fusion processing on synchronous time series running datasets based on pre-trained AI processing parameter adaptive models, multi-head collaborative scheduling rules and fault diagnosis rule bases, to generate time series processing instruction-state sequences for each cutting unit;

[0021] The causal graph construction module is used to take the processing actions, state events and control instructions in the time-series processing instructions-state sequence as nodes, and analyze the causal relationship of the nodes based on the causal discovery algorithm of time constraints and conditional independence test to construct a causal graph of multi-station processing behavior.

[0022] The multi-head collaborative labeling module is used to input the time-series snapshot sequence of the causal graph of multi-station processing behavior into a pre-trained graph neural network and time-series network combination model, extract the dynamic structural features of multi-head collaborative processing and classify them, and output multi-head collaborative control dimension labels.

[0023] The single-station process label module is used to match and compare the timing-processing instruction-state sequence with the standard processing technology library, calculate processing accuracy, efficiency, wire breakage probability, energy consumption and surface quality indicators, and generate single-station process optimization dimension labels.

[0024] The full-domain closed-loop control module is used to integrate multi-head collaborative control dimension labels with single-station process optimization dimension labels, associate key evidence of the cause-and-effect graph of multi-station processing behavior, generate full-domain control instructions and processing optimization reports, and send them to the execution unit through the Internet of Things to complete closed-loop control.

[0025] The AI ​​current adaptive adjustment module, based on the trial algorithm and fuzzy comparison prediction algorithm, realizes self-optimization of pulse width, pulse interval, peak current, processing waveform and servo feed speed, and self-evolving correction parameter library;

[0026] The intelligent coolant mixing module monitors the coolant status in real time, dynamically calculates the optimal mixing ratio, and automatically completes replenishment, purification, and circulation control.

[0027] The multi-head synchronous control module enables synchronous start-up, synchronous pause, synchronous reversal, synchronous feed, and synchronous parameter adjustment across multiple workstations, ensuring consistent processing.

[0028] Furthermore, the IoT data acquisition module includes:

[0029] The multi-protocol access unit supports OPCUA, Modbus, and MQTT protocols, enabling unified access for devices, sensors, and actuators; the high-frequency data sampling unit collects processing current, voltage, wire feed speed, tension, temperature, density, and coordinate data at a sampling rate of 1MHz.

[0030] The data preprocessing unit is used for outlier removal, missing value interpolation, noise filtering, spatiotemporal alignment, and normalization transformation; the time-series data management unit is used for frame segmentation by time window, feature extraction, and synchronous time-series dataset storage and management.

[0031] The multimodal timing analysis module includes:

[0032] The visual motion recognition unit is used for target tracking, extraction of Chinese herbal medicine slices and basic substances, classification of temporal motions, and generation of visual modal processing motion flow.

[0033] The audio and discharge detection unit is used for separating, recognizing and extracting abnormal features of human voice and discharge sound, and generating an audio modal event stream;

[0034] The sensor event parsing unit maps sensor data into operation events based on the rule engine, generating a sensor modal event stream.

[0035] The spatiotemporal fusion unit completes multimodal data stream association and evidence fusion based on timestamps and spatial locations, and outputs time-series processing instructions - state sequences.

[0036] Furthermore, the causal graph construction module includes:

[0037] The node and attribute definition unit is used to define processing actions, events, and instructions as nodes, and to define equipment, workstations, time, and parameters as node attributes;

[0038] The temporal filtering unit is used to filter candidate causal node pairs according to causal time windows; the conditional independence test unit is used to determine the existence and direction of causal edges based on the kernel conditional independence test.

[0039] The causal strength calculation unit is used to calculate the causal edge weights based on co-occurrence frequency, sequence dependence, and mutual information; the graph storage and visualization unit is used to store, update, display, and query the causal graph of multi-station processing behavior.

[0040] Furthermore, the multi-head collaborative tagging module includes:

[0041] The temporal slicing unit is used to divide the causal graph into a sequence of temporal snapshots;

[0042] The graph attention learning unit is used to learn node embedding vectors and aggregate local structural information; the graph-level representation unit is used to generate global graph representation vectors.

[0043] Temporal feature extraction unit, used to capture co-evolutionary features using long short-term memory network;

[0044] The collaborative pattern classification unit is used to output collaborative pattern labels and confidence scores, and generate multi-head collaborative control dimension labels.

[0045] Furthermore, the single-station process label module includes:

[0046] The standard process library unit stores standard operating procedures (SOPs), standard parameters, standard trajectories, and standard quality indicators.

[0047] The deviation calculation unit is used to calculate the deviation by comparing each action, each parameter, and each time sequence.

[0048] The performance evaluation unit is used to calculate processing accuracy, efficiency, wire breakage probability, energy consumption, and surface quality.

[0049] The tag generation unit is used to output single-station process optimization dimension tags.

[0050] Compared with existing technologies, this invention provides an intelligent multi-head wire EDM machining control method and system based on the Internet of Things (IoT). Through full-domain multimodal perception data acquisition, edge-cloud collaborative computing architecture, high-precision multi-head collaborative motion closed-loop control, real-time discharge state identification and adaptive adjustment, full-dimensional dynamic error compensation, digital twin 3D visualization mapping, predictive maintenance, global production scheduling and process library iterative optimization, it completely solves the core problems of traditional multi-head wire EDM equipment, such as low collaborative accuracy, lagging discharge control, process dependence on experience, insufficient IoT integration, and lack of intelligent optimization and predictive maintenance. It achieves high-precision, high-efficiency, high-reliability, intelligent, networked, and unmanned multi-head wire EDM machining, meeting the core needs of high-end manufacturing for special processing equipment. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0052] Figure 1 A flowchart illustrating the steps of the intelligent multi-head wire EDM machining control method provided in this embodiment of the invention;

[0053] Figure 2 This is a schematic diagram of the overall control flow provided in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the sensing data acquisition and spatiotemporal alignment process provided in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of the sensing data acquisition and spatiotemporal alignment process provided in an embodiment of the present invention;

[0056] Figure 5 This is a schematic diagram of the architecture module of the intelligent multi-head wire cutting control system provided in an embodiment of the present invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0058] As attached Figure 1 To be continued Figure 5 As shown:

[0059] This invention provides an intelligent multi-head wire EDM machining control method based on the Internet of Things, comprising the following steps:

[0060] Step 1: Collect multi-dimensional operating data of the multi-head wire EDM equipment cluster in real time through the IoT gateway. The multi-dimensional operating data includes equipment status data, process parameter data, sensor detection data, processing data and energy consumption data. Perform protocol parsing, spatiotemporal alignment and standardization preprocessing on the multi-dimensional operating data to obtain a synchronous time-series operating dataset.

[0061] Step 2: Based on the pre-trained AI processing parameter adaptive model, multi-head collaborative scheduling rules and fault diagnosis rule base, the synchronous time-series running dataset is parsed and fused in parallel to generate the time-series processing instruction-state sequence of each cutting unit;

[0062] Step 3: Using the processing actions, state events, and control instructions in each time-series processing instruction-state sequence as nodes, the causal relationship of the nodes is analyzed based on the causal discovery algorithm of time constraints and conditional independence test, and a multi-station processing behavior causal graph is constructed; the multi-station processing behavior causal graph represents the dynamic influence and constraint relationship between each cutting unit, execution mechanism, and processing process;

[0063] Step 4: Input the time-series snapshot sequence of the causal graph of multi-station processing behavior into a pre-trained graph neural network and time-series network combination model. Extract the dynamic structural features of multi-head collaborative processing through message passing aggregation and graph-level representation learning. Combine the preset collaborative processing mode to classify and output multi-head collaborative control dimension labels.

[0064] Step 5: Based on the timing-based machining instructions and state sequence, match and compare them with the standard machining process library to calculate machining accuracy, machining efficiency, wire breakage probability, energy consumption utilization rate and surface quality indicators, and generate single-station process optimization dimension labels;

[0065] Step Six: Integrate the multi-head collaborative control dimension label with the single-station process optimization dimension label, associate key causal evidence in the multi-station processing behavior causal graph, generate intelligent multi-head wire EDM full-domain control instructions and processing optimization reports, and send them to the corresponding execution units through the Internet of Things to complete closed-loop control.

[0066] Step one includes: using OPCUA, Modbus, and MQTT industrial communication protocols to uniformly connect multiple multi-head wire EDM machines, sensor modules, execution drive modules, and cooling systems; collecting raw data such as processing voltage, processing current, wire feed speed, tension value, coolant concentration, temperature, pulse width, pulse interval, station coordinates, and processing progress at a high-frequency sampling rate of 1MHz; performing outlier removal, missing value interpolation, and noise filtering on the raw data; completing spatiotemporal alignment based on a unified time reference; and converting heterogeneous data into time-series data in a unified format; and performing frame segmentation and feature extraction on the time-series data according to a preset time window to form a synchronous time-series running dataset containing equipment identifiers, timestamps, station numbers, and feature vectors.

[0067] Step two includes: assigning a continuous and unique identifier to each cutting unit using a target tracking model on the video stream and image sensing data; extracting the continuous position and posture sequences of the electrode wire, worktable, cooling nozzle, and workpiece; inputting the position and posture sequences into a pre-trained temporal action classification model to identify the atomic action types of processing feed, idle travel, reversal, cutting, pause, and emergency stop within a fixed short temporal window, generating a visual modal processing action stream; performing voice activity detection and sound source separation on audio and discharge sound signals, associating them with the corresponding cutting unit, and extracting abnormal discharge, collision, and abnormal vibration features through automatic speech recognition and semantic analysis to generate an audio modal event stream; parsing current, voltage, tension, temperature, and liquid level sensor data streams based on a rule engine, mapping them to normal processing, abnormal discharge, wire breakage precursors, over-temperature, over-pressure, and coolant abnormal operation events, generating a sensor modal event stream; and performing spatiotemporal correlation and evidence fusion on the visual modal processing action stream, audio modal event stream, and sensor modal event stream based on timestamps and spatial locations to generate a temporal processing instruction-state sequence for each cutting unit.

[0068] Step three includes: treating processing actions, state events, and control instructions in the time-series processing instruction-state sequence as nodes, and treating equipment number, workstation number, action type, event type, timestamp, and parameter value as node attributes; performing time-series filtering on nodes based on timestamps, and identifying node pairs whose time differences fall within a preset causal time window as candidate causal node pairs; employing a constraint-based causal discovery algorithm on the candidate causal node pairs, calculating independence statistics through kernel conditional independence tests when given a set of conditions, determining the existence and direction of candidate causal edges, and obtaining a preliminary causal structure; based on the preliminary causal structure, calculating causal edge weights through historical co-occurrence frequency, sequence dependency strength, and conditional mutual information to obtain an estimated causal strength value; and constructing a multi-workstation processing behavior causal graph based on nodes, node attributes, causal edges, and the estimated causal strength value.

[0069] Step four includes: slicing the causal graph of multi-station processing behavior into fixed-duration slices to obtain a sequence of time-series snapshots; learning node representations for each time-series snapshot using a graph attention network, combining initial node features, neighbor features, and causal strength estimates, and aggregating node embedding vectors containing local structural information through an attention mechanism; aggregating all node embedding vectors using a graph-level readout function to generate a graph global representation vector representing the current collaborative state; inputting the graph global representation vectors of continuous time-series snapshots into a long short-term memory network to capture the temporal evolution features of collaborative patterns, obtaining a team collaborative state encoding vector; inputting the collaborative state encoding vector into a pre-trained multi-label collaborative pattern classifier to output collaborative pattern labels and confidence scores; the collaborative pattern labels include efficient synchronous processing, predictive assisted feeding, workstation conflict, resource competition, communication redundancy, and chain reaction risk of broken wires; when the confidence score exceeds a set threshold, generating corresponding multi-head collaborative control dimension labels, and backtracking to locate the key subgraphs and key nodes in the causal graph of multi-station processing behavior that cause anomalies.

[0070] Step five includes: comparing the timing-based machining instructions – state sequence – with the standard SOP process, standard process parameter library, and standard machining trajectory frame by frame and action by action; calculating the action deviation, parameter deviation, timing deviation, and trajectory deviation between the actual machining and standard machining; calculating the machining accuracy compliance rate, machining efficiency, theoretical wire breakage probability, energy consumption utilization rate, and surface roughness prediction value based on the deviation values; and generating single-station process optimization dimension labels based on the calculation results. The labels include reasonable parameters, high parameters, low parameters, excessively fast feed, excessively slow feed, insufficient cooling, excessive cooling, abnormal tension, easy wire breakage, high precision, low efficiency, and high energy consumption.

[0071] Step six includes: weighted fusion of multi-head collaborative control dimension labels and single-station process optimization dimension labels to form a comprehensive evaluation label system; mapping the comprehensive evaluation labels to a causal graph of multi-station processing behavior to locate key causal paths, key instructions, key parameters, and key stations leading to anomalies or inefficiencies; generating full-domain control instructions based on key causal evidence, including synchronous start / stop, group scheduling, automatic parameter correction, feed rate adjustment, tension compensation, coolant ratio adjustment, fault isolation, and early warning prompts; distributing full-domain control instructions to corresponding equipment, stations, and actuators via the Internet of Things to achieve real-time closed-loop control; and generating a processing optimization report that includes a processing overview, collaborative evaluation, single-station evaluation, key issues, optimization suggestions, and control logs, supporting visualization and report export.

[0072] The overall process of multi-head wire EDM machining control includes the following steps:

[0073] S101: IoT Multi-Source Sensing Data Acquisition and High-Precision Spatiotemporal Alignment Deployment of Multi-Dimensional High-Speed ​​Synchronous Sensing Units to collect motion, discharge, filament, liquid, environment, and equipment health data; achieves microsecond-level time synchronization and spatial coordinate alignment through IEEE1588PTP, and outputs standardized synchronous multimodal time-series data streams;

[0074] This step specifically includes:

[0075] S201: Initialize the sensing unit, start the motion sensing, discharge sampling, tension monitoring, wire status detection, environmental sensing, and equipment health sensing units, and perform self-test, zero-point calibration, range calibration, and communication verification;

[0076] S202: Multi-channel data synchronous acquisition, data is acquired at the configured frequency: motion data 1kHz, discharge data 10kHz, tension / wire speed data 5kHz, environmental / health data 1Hz.

[0077] S203: High-precision time synchronization: All sensing units are synchronized using the IEEE1588PTP protocol, with a time deviation of ≤1μs, ensuring that all data have a unified timestamp.

[0078] S204: Spatial coordinate alignment, completing the calibration of the grating ruler, sensor, and processing coordinate system to achieve unified data spatial mapping; S205 Preliminary data verification: performing integrity verification, outlier removal, missing value completion, and deduplication on the collected data; S206 Output synchronous multimodal data stream: generating a unified timestamp, unified spatial coordinates, and standardized time-series dataset, which is then sent to the edge computing unit;

[0079] The above steps collect a total of 32 analog signals, 16 digital signals, and 6 discharge waveform data, with a data integrity rate of ≥99.9% and a synchronization accuracy that meets the requirements of high-precision control.

[0080] S102: Edge-end multimodal data preprocessing and real-time discharge state identification: Filtering, normalizing, and feature extraction of time-series data; Real-time identification of discharge states such as normal, no-load, short-circuit, arc, and wire breakage precursor based on CNN-LSTM hybrid model, outputting abnormality level and location information;

[0081] This step specifically includes:

[0082] S301: Multi-level data filtering, using wavelet transform to remove high-frequency noise, Kalman filtering to suppress random interference, and moving average filtering to smooth data;

[0083] S302: Feature extraction, extracting discharge peak value, mean value, duty cycle, short circuit rate, arc duration, fluctuation, tension change rate, and wire vibration amplitude;

[0084] S303: Model inference, inputting features into a lightweight CNN-LSTM hybrid model, with an inference latency of ≤3ms;

[0085] S304: Discharge status classification, with 6 states: normal output, no load, short circuit, micro arc, unstable arc, and wire breakage precursor.

[0086] S305: Confidence determination, calculate the confidence of each state, and determine the current state if it exceeds the threshold;

[0087] S306: Anomaly level classification: minor, moderate, severe, and fatal; S307: Output status results: send the status, confidence level, anomaly level, and location of occurrence to the control unit.

[0088] The above steps achieve a discharge state identification accuracy of ≥98% and an anomaly identification coverage of ≥99.5%.

[0089] S103: Multi-head cooperative motion trajectory planning and multi-dimensional online error compensation. Generate multi-head machining trajectory according to drawings and process requirements; establish coupled dynamic model, and adopt NURBS interpolation and flexible velocity planning; integrate multi-source feedback data to complete online compensation for thermal error, clearance error, and deflection error.

[0090] This step specifically includes:

[0091] S401: Machining task analysis, import DXF drawings, analyze contour, dimensions, accuracy, material, thickness, and machining requirements;

[0092] S402: Processing area division, automatically dividing the processing area according to the number of multiple heads to avoid interference;

[0093] S403: Trajectory generation, using NURBS curve interpolation to generate smooth trajectories, supporting arcs, straight lines, and spline curves;

[0094] S404: Cooperative control strategy selection, adopting full synchronization mode;

[0095] S405: Speed ​​and acceleration / deceleration planning, flexible acceleration / deceleration, avoiding impact and vibration;

[0096] S406: Real-time position feedback, obtaining the actual position of multiple axes through a grating ruler;

[0097] S407: Error calculation, calculate synchronization error, trajectory error, and positioning error;

[0098] S408: Multi-dimensional error compensation, including thermal deformation compensation, gap compensation, wire deflection compensation, and return error compensation;

[0099] S409: Synchronous calibration, real-time calibration of multi-head positions to ensure synchronization accuracy;

[0100] S410: Collision protection, real-time calculation of distance between multiple targets, and immediate deceleration and stop if the distance is lower than the safety threshold;

[0101] In the above steps, the multi-head synchronization error is ≤ ±0.5μm, and the trajectory tracking error is ≤ ±0.3μm;

[0102] S104: Multi-electrode parameter adaptive optimization and closed-loop adjustment aims at efficiency, accuracy, surface quality and low loss. It adopts improved NSGA-II and deep reinforcement learning to adaptively adjust parameters such as pulse, wire speed, tension, servo and flushing fluid according to the processing stage and real-time discharge status.

[0103] This step specifically includes:

[0104] S501: Processing stage judgment: Automatically identify roughing, intermediate processing, finishing, and ultra-finishing;

[0105] S502: Objective function setting: Roughing prioritizes efficiency, while finishing prioritizes surface quality and precision;

[0106] S503: Multi-objective optimization calculation: Optimal parameters are calculated using the improved NSGA-II+DDQN algorithm;

[0107] S504: Parameter output: output pulse width, pulse interval, peak current, wire speed, tension, servo speed;

[0108] S505: Execution parameter adjustment: The discharge power supply and execution unit update parameters in real time;

[0109] S506: Status Feedback: Collects discharge status and processing quality to form a closed loop;

[0110] S507: Model Iteration: Update the reinforcement learning model based on feedback and continuously optimize it;

[0111] In the above steps, the parameter adjustment response time is ≤10ms, and the processing efficiency is 180% higher than that of the traditional single head.

[0112] S105: Discharge abnormality graded protection and wire breakage emergency handling: Perform reversal, speed reduction, pause, and emergency stop according to the minor / general / severe / fatal abnormality grade; cut off pulse power supply, lock axis movement, trigger alarm, and record all abnormal data and upload it to the cloud;

[0113] S601: Receives abnormal status signals;

[0114] S602: Grading judgment: Minor → Fine adjustment; Moderate → Rollback; Severe → Pause; Fatal → Shutdown;

[0115] S603: Performs protection actions: short circuit back-off, arc suppression, and wire breakage emergency stop;

[0116] S604: Alarm output: Audible and visual alarms + cloud push;

[0117] S605: Data Recording: Saves data throughout the entire process of an anomaly;

[0118] S606: Anomaly Recovery: After automatic troubleshooting, you can choose automatic recovery or manual recovery;

[0119] In the above steps, the wire breakage protection response time is ≤1ms, and the workpiece damage rate is reduced by ≥95%;

[0120] S106: IoT encrypted transmission and real-time cloud-based digital twin mapping: Data is uploaded via 5G / Industrial Ethernet with MQTT / OPCUA encryption; 3D digital twins of equipment are built in the cloud to achieve real-time visualization and remote monitoring of location, trajectory, discharge, process, and health status.

[0121] S701: Edge data encapsulation and encryption;

[0122] S702: Uploads data to the cloud via 5G / Ethernet;

[0123] S703: Cloud data parsing and verification;

[0124] S704: Digital twin model updated in real time;

[0125] S705: Remote visual display;

[0126] In the above steps, the cloud mapping latency is ≤20ms, supporting remote real-time monitoring via PC / mobile phone.

[0127] S107: Cloud-based device health assessment and predictive maintenance are based on LSTM / Transformer timing models and fault knowledge bases to predict the remaining lifespan of key components; output maintenance level, location, content and time recommendations to achieve predictive maintenance;

[0128] S801: Receives health data via the cloud;

[0129] S802: Predicting remaining lifetime based on LSTM model;

[0130] S803: Generate maintenance recommendations;

[0131] S804: Push maintenance reminder;

[0132] S805: Maintenance record archiving;

[0133] In the above steps, the prediction accuracy is ≥92%, and the reduction of sudden failures is ≥70%;

[0134] S108: Cloud-based global production scheduling and process library selection optimization intelligently allocates tasks based on work order priority, equipment status, and capacity load; it also accumulates processing data and sends it to optimize the process library, reinforcement learning model, and discharge recognition model.

[0135] S901: Receive production work order;

[0136] S902: Equipment Status Assessment;

[0137] S903: Intelligent task allocation;

[0138] S904: Processing data storage;

[0139] S905: Process library updated and optimized;

[0140] S109: Automatic processing quality assessment and full-process report generation. Automatically calculates dimensional accuracy, roughness, consistency, efficiency, loss, and energy consumption; generates assessment reports containing equipment, process, status, anomalies, quality, and improvement suggestions, which are saved locally and archived in the cloud.

[0141] S1001: Collect processing result data;

[0142] S1002: Calculate quality indicators;

[0143] S1003: Generate a complete report;

[0144] S1004: Local storage + cloud archiving;

[0145] S110: Multi-level safety protection and abnormal log management provide protection against collisions, limit switches, emergency stops, overcurrent and overvoltage, overheating, and insulation failure; it records operation, parameter, alarm, maintenance, and processing logs throughout the process, and supports querying and exporting audit logs;

[0146] S1101: Real-time security monitoring;

[0147] S1102: Multi-level protection trigger;

[0148] S1103: Log recording throughout the entire process;

[0149] S1104: Log query and auditing;

[0150] An intelligent multi-head wire EDM machining control system based on the Internet of Things includes:

[0151] The IoT data acquisition module is used to collect multi-dimensional operational data of the multi-head wire cutting equipment cluster in real time through the IoT gateway. It performs protocol parsing, spatiotemporal alignment and standardization preprocessing on the multi-dimensional operational data to obtain a synchronous time-series operational dataset.

[0152] The multimodal time series parsing module is used to perform parallel parsing and fusion processing on synchronous time series running datasets based on pre-trained AI processing parameter adaptive models, multi-head collaborative scheduling rules and fault diagnosis rule bases, to generate time series processing instruction-state sequences for each cutting unit;

[0153] The causal graph construction module is used to take the processing actions, state events and control instructions in the time-series processing instructions-state sequence as nodes, and analyze the causal relationship of the nodes based on the causal discovery algorithm of time constraints and conditional independence test to construct a causal graph of multi-station processing behavior.

[0154] The multi-head collaborative labeling module is used to input the time-series snapshot sequence of the causal graph of multi-station processing behavior into a pre-trained graph neural network and time-series network combination model, extract the dynamic structural features of multi-head collaborative processing and classify them, and output multi-head collaborative control dimension labels.

[0155] The single-station process label module is used to match and compare the timing-processing instruction-state sequence with the standard processing technology library, calculate processing accuracy, efficiency, wire breakage probability, energy consumption and surface quality indicators, and generate single-station process optimization dimension labels.

[0156] The full-domain closed-loop control module is used to integrate multi-head collaborative control dimension labels with single-station process optimization dimension labels, associate key evidence of the cause-and-effect graph of multi-station processing behavior, generate full-domain control instructions and processing optimization reports, and send them to the execution unit through the Internet of Things to complete closed-loop control.

[0157] The AI ​​current adaptive adjustment module, based on the trial algorithm and fuzzy comparison prediction algorithm, realizes self-optimization of pulse width, pulse interval, peak current, processing waveform and servo feed speed, and self-evolving correction parameter library;

[0158] The intelligent coolant mixing module monitors the coolant status in real time, dynamically calculates the optimal mixing ratio, and automatically completes replenishment, purification, and circulation control.

[0159] The multi-head synchronous control module enables synchronous start-up, synchronous pause, synchronous reversal, synchronous feed, and synchronous parameter adjustment across multiple workstations, ensuring consistent processing.

[0160] The IoT data acquisition module includes:

[0161] The multi-protocol access unit supports OPCUA, Modbus, and MQTT protocols, enabling unified access for devices, sensors, and actuators; the high-frequency data sampling unit collects processing current, voltage, wire feed speed, tension, temperature, density, and coordinate data at a sampling rate of 1MHz.

[0162] The data preprocessing unit is used for outlier removal, missing value interpolation, noise filtering, spatiotemporal alignment, and normalization transformation; the time-series data management unit is used for frame segmentation by time window, feature extraction, and synchronous time-series dataset storage and management.

[0163] The multimodal timing analysis module includes:

[0164] The visual motion recognition unit is used for target tracking, skeleton and pose extraction, temporal motion classification, and to generate visual modality processing motion streams.

[0165] The audio and discharge detection unit is used for separating, recognizing and extracting abnormal features of human voice and discharge sound, and generating an audio modal event stream;

[0166] The sensor event parsing unit maps sensor data into operation events based on the rule engine, generating a sensor modal event stream.

[0167] The spatiotemporal fusion unit completes multimodal data stream association and evidence fusion based on timestamps and spatial locations, and outputs time-series processing instructions - state sequences.

[0168] The causal mapping construction module includes:

[0169] The node and attribute definition unit is used to define processing actions, events, and instructions as nodes, and to define equipment, workstations, time, and parameters as node attributes;

[0170] The temporal filtering unit is used to filter candidate causal node pairs according to causal time windows; the conditional independence test unit is used to determine the existence and direction of causal edges based on the kernel conditional independence test.

[0171] The causal strength calculation unit is used to calculate the causal edge weights based on co-occurrence frequency, sequence dependence, and mutual information; the graph storage and visualization unit is used to store, update, display, and query the causal graph of multi-station processing behavior.

[0172] The multi-head collaborative tagging module includes:

[0173] The temporal slicing unit is used to divide the causal graph into a sequence of temporal snapshots;

[0174] The graph attention learning unit is used to learn node embedding vectors and aggregate local structural information; the graph-level representation unit is used to generate global graph representation vectors.

[0175] Temporal feature extraction unit, used to capture co-evolutionary features using long short-term memory network;

[0176] The collaborative pattern classification unit is used to output collaborative pattern labels and confidence scores, and generate multi-head collaborative control dimension labels.

[0177] The single-station process label module includes:

[0178] The standard process library unit stores standard operating procedures (SOPs), standard parameters, standard trajectories, and standard quality indicators.

[0179] The deviation calculation unit is used to calculate the deviation by comparing each action, each parameter, and each time sequence.

[0180] The performance evaluation unit is used to calculate processing accuracy, efficiency, wire breakage probability, energy consumption, and surface quality.

[0181] The tag generation unit is used to output single-station process optimization dimension tags.

[0182] Furthermore, the technical solution of the present invention can be extended to the pre-processing stage of traditional Chinese medicine research and production: In response to the batch-to-batch difference in the extraction of effective components caused by the diverse origins and uneven texture of Chinese medicinal materials, the invention utilizes the characteristics of multi-head parallel processing and dynamic parameter adaptive processing to achieve micron-level homogeneous slicing of Chinese medicinal decoction pieces and medicinal plant samples, thus meeting the standardization requirements for sample preparation in the study of the pharmacodynamic material basis of traditional Chinese medicine.

[0183] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A smart multi-head wire EDM machining control method based on the Internet of Things, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional operating data of the multi-head wire EDM equipment cluster in real time through the IoT gateway. The multi-dimensional operating data includes equipment status data, process parameter data, sensor detection data, processing data and energy consumption data. Perform protocol parsing, spatiotemporal alignment and standardization preprocessing on the multi-dimensional operating data to obtain a synchronous time-series operating dataset. Step 2: Based on the pre-trained AI processing parameter adaptive model, multi-head collaborative scheduling rules and fault diagnosis rule base, the synchronous time-series running dataset is parsed and fused in parallel to generate the time-series processing instruction-state sequence of each cutting unit; Step 3: Using the processing actions, state events, and control instructions in each time-series processing instruction-state sequence as nodes, the causal relationship of the nodes is analyzed based on the causal discovery algorithm of time constraints and conditional independence test, and a multi-station processing behavior causal graph is constructed; the multi-station processing behavior causal graph represents the dynamic influence and constraint relationship between each cutting unit, execution mechanism, and processing process; Step 4: Input the time-series snapshot sequence of the causal graph of multi-station processing behavior into a pre-trained graph neural network and time-series network combination model. Extract the dynamic structural features of multi-head collaborative processing through message passing aggregation and graph-level representation learning. Combine the preset collaborative processing mode to classify and output multi-head collaborative control dimension labels. Step 5: Based on the timing-based machining instructions and state sequence, match and compare them with the standard machining process library to calculate machining accuracy, machining efficiency, wire breakage probability, energy consumption utilization rate and surface quality indicators, and generate single-station process optimization dimension labels; Step Six: Integrate the multi-head collaborative control dimension label with the single-station process optimization dimension label, associate key causal evidence in the multi-station processing behavior causal graph, generate intelligent multi-head wire EDM full-domain control instructions and processing optimization reports, and send them to the corresponding execution units through the Internet of Things to complete closed-loop control.

2. The intelligent multi-head wire EDM machining control method based on the Internet of Things according to claim 1, characterized in that, Step one includes: using OPCUA, Modbus, and MQTT industrial communication protocols to uniformly connect multiple multi-head wire EDM machines, sensor modules, execution drive modules, and cooling systems; collecting raw data such as processing voltage, processing current, wire feed speed, tension value, coolant concentration, temperature, pulse width, pulse interval, station coordinates, and processing progress at a high-frequency sampling rate of 1MHz; performing outlier removal, missing value interpolation, and noise filtering on the raw data; completing spatiotemporal alignment based on a unified time reference; and converting heterogeneous data into time-series data in a unified format; and performing frame segmentation and feature extraction on the time-series data according to a preset time window to form a synchronous time-series running dataset containing equipment identifiers, timestamps, station numbers, and feature vectors.

3. The intelligent multi-head wire EDM machining control method based on the Internet of Things according to claim 1, characterized in that, Step two includes: assigning a continuous and unique identifier to each cutting unit using a target tracking model on the video stream and image sensing data; extracting the continuous position and posture sequences of the electrode wire, worktable, cooling nozzle, and workpiece; inputting the position and posture sequences into a pre-trained temporal action classification model to identify the atomic action types of processing feed, idle travel, reversal, cutting, pause, and emergency stop within a fixed short temporal window, generating a visual modal processing action stream; performing voice activity detection and sound source separation on audio and discharge sound signals, associating them with the corresponding cutting unit, and extracting abnormal discharge, collision, and abnormal vibration features through automatic speech recognition and semantic analysis to generate an audio modal event stream; parsing current, voltage, tension, temperature, and liquid level sensor data streams based on a rule engine, mapping them to normal processing, abnormal discharge, wire breakage precursors, over-temperature, over-pressure, and coolant abnormal operation events, generating a sensor modal event stream; and performing spatiotemporal correlation and evidence fusion on the visual modal processing action stream, audio modal event stream, and sensor modal event stream based on timestamps and spatial locations to generate a temporal processing instruction-state sequence for each cutting unit.

4. The intelligent multi-head wire EDM machining control method based on the Internet of Things according to claim 1, characterized in that, Step three includes: treating processing actions, state events, and control instructions in the time-series processing instruction-state sequence as nodes, and treating equipment number, workstation number, action type, event type, timestamp, and parameter value as node attributes; performing time-series filtering on nodes based on timestamps, and identifying node pairs whose time differences fall within a preset causal time window as candidate causal node pairs; employing a constraint-based causal discovery algorithm on the candidate causal node pairs, calculating independence statistics through kernel conditional independence tests when given a set of conditions, determining the existence and direction of candidate causal edges, and obtaining a preliminary causal structure; based on the preliminary causal structure, calculating causal edge weights through historical co-occurrence frequency, sequence dependency strength, and conditional mutual information to obtain an estimated causal strength value; and constructing a multi-workstation processing behavior causal graph based on nodes, node attributes, causal edges, and the estimated causal strength value. Step four includes: slicing the causal graph of multi-station processing behavior into fixed-duration slices to obtain a sequence of time-series snapshots; learning node representations for each time-series snapshot using a graph attention network, combining initial node features, neighbor features, and causal strength estimates, and aggregating node embedding vectors containing local structural information through an attention mechanism; aggregating all node embedding vectors using a graph-level readout function to generate a graph global representation vector representing the current collaborative state; inputting the graph global representation vectors of continuous time-series snapshots into a long short-term memory network to capture the temporal evolution features of collaborative patterns, obtaining a team collaborative state encoding vector; inputting the collaborative state encoding vector into a pre-trained multi-label collaborative pattern classifier to output collaborative pattern labels and confidence scores; the collaborative pattern labels include efficient synchronous processing, predictive assisted feeding, workstation conflict, resource competition, communication redundancy, and chain reaction risk of broken wires; when the confidence score exceeds a set threshold, generating corresponding multi-head collaborative control dimension labels, and backtracking to locate the key subgraphs and key nodes in the causal graph of multi-station processing behavior that cause anomalies.

5. The intelligent multi-head wire EDM machining control method based on the Internet of Things according to claim 1, characterized in that, Step five includes: comparing the timing-based machining instructions – state sequence – with the standard SOP process, standard process parameter library, and standard machining trajectory frame by frame and action by action; calculating the action deviation, parameter deviation, timing deviation, and trajectory deviation between the actual machining and standard machining; calculating the machining accuracy compliance rate, machining efficiency, theoretical wire breakage probability, energy consumption utilization rate, and surface roughness prediction value based on the deviation values; and generating single-station process optimization dimension labels based on the calculation results. The labels include reasonable parameters, high parameters, low parameters, excessively fast feed, excessively slow feed, insufficient cooling, excessive cooling, abnormal tension, easy wire breakage, high precision, low efficiency, and high energy consumption. Step six includes: weighted fusion of multi-head collaborative control dimension labels and single-station process optimization dimension labels to form a comprehensive evaluation label system; mapping the comprehensive evaluation labels to a causal graph of multi-station processing behavior to locate key causal paths, key instructions, key parameters, and key stations leading to anomalies or inefficiencies; generating global control instructions based on key causal evidence, including synchronous start / stop, group scheduling, automatic parameter correction, feed rate adjustment, tension compensation, coolant ratio adjustment, fault isolation, and early warning prompts; distributing global control instructions to corresponding equipment, stations, and actuators via the Internet of Things to achieve real-time closed-loop control; and generating a processing optimization report that includes a processing overview, collaborative evaluation, single-station evaluation, key issues, optimization suggestions, and control logs, supporting visualization and report export.

6. An intelligent multi-head wire EDM machining control system based on the Internet of Things, characterized in that, include: The IoT data acquisition module is used to collect multi-dimensional operational data of the multi-head wire cutting equipment cluster in real time through the IoT gateway. It performs protocol parsing, spatiotemporal alignment and standardization preprocessing on the multi-dimensional operational data to obtain a synchronous time-series operational dataset. The multimodal time series parsing module is used to perform parallel parsing and fusion processing on synchronous time series running datasets based on pre-trained AI processing parameter adaptive models, multi-head collaborative scheduling rules and fault diagnosis rule bases, to generate time series processing instruction-state sequences for each cutting unit; The causal graph construction module is used to take the processing actions, state events and control instructions in the time-series processing instructions-state sequence as nodes, and analyze the causal relationship of the nodes based on the causal discovery algorithm of time constraints and conditional independence test to construct a causal graph of multi-station processing behavior. The multi-head collaborative labeling module is used to input the time-series snapshot sequence of the causal graph of multi-station processing behavior into a pre-trained graph neural network and time-series network combination model, extract the dynamic structural features of multi-head collaborative processing and classify them, and output multi-head collaborative control dimension labels. The single-station process label module is used to match and compare the timing-processing instruction-state sequence with the standard processing technology library, calculate processing accuracy, efficiency, wire breakage probability, energy consumption and surface quality indicators, and generate single-station process optimization dimension labels. The full-domain closed-loop control module is used to integrate multi-head collaborative control dimension labels with single-station process optimization dimension labels, associate key evidence of the cause-and-effect graph of multi-station processing behavior, generate full-domain control instructions and processing optimization reports, and send them to the execution unit through the Internet of Things to complete closed-loop control. The AI ​​current adaptive adjustment module, based on the trial algorithm and fuzzy comparison prediction algorithm, realizes self-optimization of pulse width, pulse interval, peak current, processing waveform and servo feed speed, and self-evolving correction parameter library; The intelligent coolant mixing module monitors the coolant status in real time, dynamically calculates the optimal mixing ratio, and automatically completes replenishment, purification, and circulation control. The multi-head synchronous control module enables synchronous start-up, synchronous pause, synchronous reversal, synchronous feed, and synchronous parameter adjustment across multiple workstations, ensuring consistent processing.

7. The intelligent multi-head wire EDM machining control system based on the Internet of Things according to claim 6, characterized in that, The IoT data acquisition module includes: The multi-protocol access unit supports OPCUA, Modbus, and MQTT protocols, enabling unified access for devices, sensors, and actuators; the high-frequency data sampling unit collects processing current, voltage, wire feed speed, tension, temperature, density, and coordinate data at a sampling rate of 1MHz. The data preprocessing unit is used for outlier removal, missing value interpolation, noise filtering, spatiotemporal alignment, and normalization transformation; the time-series data management unit is used for frame segmentation by time window, feature extraction, and synchronous time-series dataset storage and management. The multimodal timing analysis module includes: The visual motion recognition unit is used for target tracking, skeleton and pose extraction, temporal motion classification, and to generate visual modality processing motion streams. The audio and discharge detection unit is used for separating, recognizing and extracting abnormal features of human voice and discharge sound, and generating an audio modal event stream; The sensor event parsing unit maps sensor data into operation events based on the rule engine, generating a sensor modal event stream. The spatiotemporal fusion unit completes multimodal data stream association and evidence fusion based on timestamps and spatial locations, and outputs time-series processing instructions - state sequences.

8. The intelligent multi-head wire EDM machining control system based on the Internet of Things according to claim 6, characterized in that, The causal mapping construction module includes: The node and attribute definition unit is used to define processing actions, events, and instructions as nodes, and to define equipment, workstations, time, and parameters as node attributes; The temporal filtering unit is used to filter candidate causal node pairs according to causal time windows; the conditional independence test unit is used to determine the existence and direction of causal edges based on the kernel conditional independence test. The causal strength calculation unit is used to calculate the causal edge weights based on co-occurrence frequency, sequence dependence, and mutual information; the graph storage and visualization unit is used to store, update, display, and query the causal graph of multi-station processing behavior.

9. A smart multi-head wire EDM machining control system based on the Internet of Things according to claim 6, characterized in that, The multi-head collaborative tagging module includes: The temporal slice unit is used to divide the causal graph into a sequence of temporal snapshots; The graph attention learning unit is used to learn node embedding vectors and aggregate local structural information; the graph-level representation unit is used to generate global graph representation vectors. Temporal feature extraction unit, used to capture co-evolutionary features using long short-term memory network; The collaborative pattern classification unit is used to output collaborative pattern labels and confidence scores, and generate multi-head collaborative control dimension labels.

10. The intelligent multi-head wire EDM machining control system based on the Internet of Things according to claim 6, characterized in that, The single-station process label module includes: The standard process library unit stores standard operating procedures (SOPs), standard parameters, standard trajectories, and standard quality indicators. The deviation calculation unit is used to calculate the deviation by comparing each action, each parameter, and each time sequence. The performance evaluation unit is used to calculate processing accuracy, efficiency, wire breakage probability, energy consumption, and surface quality. The tag generation unit is used to output single-station process optimization dimension tags.

Images