Intelligent regulation and control system and method for discharge energy of spark machine

By using a closed-loop control system based on multimodal sensing and LSTM network, the discharge energy in electrical discharge machining is adjusted in real time, solving the trade-off between efficiency and quality in the traditional fixed parameter mode and achieving efficient and stable electrical discharge machining.

CN121315359AActive Publication Date: 2026-01-13DONGGUAN DALING MECHANICAL & ELECTRICAL CO LTD

Patent Information

Application Number
CN202511755415.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-26
Publication Date
2026-01-13
Estimated Expiration
2045-11-26

AI Technical Summary

Technical Problem

In existing electrical discharge machining (EDM), discharge energy regulation is mainly achieved through fixed parameter modes, which cannot respond to changes in the discharge gap state in real time. This leads to a static trade-off between efficiency and quality, resulting in defects such as electrode wear and workpiece burns.

Method used

A closed-loop control system is formed by employing a multimodal gap state sensing module, a discharge pulse real-time classification and feature extraction module, a gap state trajectory prediction module based on a long short-term memory network, and a multi-objective model predictive control module to adjust the discharge energy parameters in real time.

Benefits of technology

It achieves precise matching of discharge energy, improves the stability and efficiency of the processing, avoids unstable discharge, and ensures high-quality processing results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121315359A_ABST
    Figure CN121315359A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of machining and intelligent manufacturing, and particularly relates to an intelligent regulation and control system and method for discharge energy of a spark machine. Comprising a multi-mode gap state sensing module, a high-throughput data synchronous acquisition and preprocessing module, a discharge pulse real-time classification and feature extraction module, an LSTM-based gap state trajectory prediction module, a multi-target model prediction control module and a high-frequency pulse power supply execution module. By fusing the electrical parameters and the in-situ optical turbidity signal, multi-dimensional perception of the discharge state is realized, the future discharge trend is predicted by using an LSTM network, and an optimal pulse parameter combination is solved online by combining an MPC controller. By the adoption of the technical scheme, dynamic optimal balance of the material corrosion rate, the surface quality and the stability in the machining process can be achieved, and the overall machining benefit is improved.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical processing and intelligent manufacturing, in particular to a spark machine discharge energy intelligent regulation and control system and method. BACKGROUND

[0002] As a special processing method for removing conductive materials by using the phenomenon of pulse spark discharge and electric corrosion, the electric spark machining technology has become a key technology in high-end fields such as mold manufacturing and aerospace because it can process super-hard materials, complex cavities and precise microstructures. The core of this technology is the accurate control of pulse discharge energy, which directly determines the material removal rate, workpiece surface quality and electrode loss degree, and is the key to ensuring efficient, stable and precise processing.

[0003] In existing electric spark machining, discharge energy regulation and control is mainly achieved by setting three core parameters of pulse width, pulse interval and peak current of the pulse power supply. The traditional strategy adopts an empirical phased fixed parameter mode: before processing, the parameter combination is preset according to material properties, processing allowance, etc., a large pulse width and high peak current are used in the rough machining stage to pursue high removal rate, and a small pulse width and low peak current are switched to in the fine machining stage to ensure surface quality. This mode has laid a foundation for the industrialization of electric spark machining.

[0004] With the increasing requirements of high-end manufacturing on processing precision, surface integrity and efficiency, the traditional strategy shows principle limitations: electric spark machining is a millisecond dynamic process, the real-time changes of dielectric liquid state and erosion product distribution in the discharge gap directly affect the discharge stability; and the fixed parameter mode is an open-loop control, which ignores the instantaneous feedback of the gap state, resulting in a mismatch between energy supply and actual demand - when the gap state is good, the parameters are conservative, wasting efficiency, and when the gap state is deteriorated, the parameters are aggressive, causing electrode loss, workpiece burn and other defects. This static trade-off between efficiency and quality has become a core bottleneck for technology upgrading.

[0005] Therefore, the application provides a spark machine discharge energy intelligent regulation and control system and method. SUMMARY

[0006] In order to make up for the deficiencies of the prior art and solve at least one technical problem proposed in the background art.

[0007] The technical solution adopted by the application to solve its technical problems is: the spark machine discharge energy intelligent regulation and control system comprises a multi-modal gap state sensing module, a high-throughput data synchronous acquisition and preprocessing module, a discharge pulse real-time classification and feature extraction module, a long short-term memory network-based gap state trajectory prediction module, a multi-objective model predictive control module, and a high-frequency pulse power supply execution module.

[0008] The multi-modal gap state sensing module is arranged in a working area of the electric discharge machining machine tool, and is used for synchronously acquiring multi-dimensional physical signals directly or indirectly related to a discharge gap state. Specifically, the multi-modal gap state sensing module comprises a high-bandwidth electrical parameter sensing unit and an in-situ optical turbidity sensing unit. The high-bandwidth electrical parameter sensing unit is composed of a Hall effect current sensor connected in series with a tool electrode and a workpiece and having a response time less than 0.5 microseconds, and a high-speed differential voltage probe connected across the tool electrode and the workpiece and having an input impedance greater than 10 megaohms, and is used for capturing instantaneous current and voltage waveforms during each discharge pulse without distortion. The in-situ optical turbidity sensing unit is integrally encapsulated in a flow cell structure formed by injection molding of polyether ether ketone (PEEK) material and having a standard fluid interface, and is connected in series with a backflow pipeline of a machining working fluid circulation filtering system and is adjacent to a liquid discharge port of a machining tank. The flow cell integrally forms a straight-through measurement channel with a diameter of 10 millimeters. On one side wall of the measurement channel, an infrared light emitting assembly is installed perpendicularly to a fluid direction; on the other side wall opposite to the emitting assembly, a photoreception assembly is installed. The infrared light emitting assembly encapsulates an infrared light emitting diode with a central wavelength of 940 nanometers, a spectral half-width of 40 nanometers, and a radiation intensity of 750 milliwatts per steradian. The photoreception assembly encapsulates a silicon-based PIN photodiode with a photosensitive area of 10 square millimeters, a peak response wavelength of 940 nanometers, and a dark current less than 1.5 nanoamperes. The in-situ optical turbidity sensing unit quantifies turbidity caused by electric erosion products (mainly carbon particles and metal particles) in the working fluid in real time by measuring a degree of attenuation of infrared light of a specific wavelength by the working fluid flowing through the measurement channel, so as to obtain a physical quantity directly representing a discharge gap chip removal state and a pollution degree.

[0009] The high-throughput data synchronous acquisition and preprocessing module is connected with a signal output end of the multi-modal gap state sensing module, and is used for high-precision synchronous digitization and preliminary hardware-level feature calculation on multi-source heterogeneous sensing signals. It can be understood that a core of the module is a hardware processing platform based on a field programmable gate array (FPGA). Analog input channels of the platform are connected with output ends of the Hall effect current sensor and the high-speed differential voltage probe respectively through two independent analog-to-digital converters (ADCs) with a sampling rate of 50 mega-hertz and a vertical resolution of 16 bits. Another analog input channel of the platform is connected with an output end of a preamplifier circuit at a rear end of the photodiode through an analog-to-digital converter with a sampling rate of 100 kilo-hertz and a vertical resolution of 16 bits. The hardware logic embedded inside the FPGA first uses a unified internal high-precision clock to stamp the data streams from the three channels with a synchronization timestamp. Secondly, the high-speed sampled voltage and current waveform data streams are digitally filtered and processed in real time. Using hardware-implemented peak detection algorithm, zero-crossing detection algorithm and numerical integration algorithm, the peak current, peak voltage, breakdown delay time, pulse duration, single pulse energy and voltage waveform integral area are calculated in each pulse cycle. Next, the photoelectric signal is averaged and filtered to generate a quantitative value that characterizes the real-time turbidity of the processing fluid, denoted as the turbidity index (DCI). Finally, the FPGA packages the synchronization timestamp, key feature data of the original waveform, and the corresponding turbidity index of each pulse into a structured data frame, and transmits it at high speed to the discharge pulse real-time classification and feature extraction module through the PCIe bus interface.

[0010] The discharge pulse real-time classification and feature extraction module is a deterministic computing unit running under a real-time operating system (RTOS). It performs in-depth analysis of the data frames received from the FPGA and accurately classifies the physical properties of each discharge pulse. Internally, this module runs a deterministic classification algorithm based on a multi-dimensional feature space. The input to this algorithm is the data frame for each pulse, including peak current, peak voltage, breakdown delay time, pulse duration, single-pulse energy, and turbidity index. The algorithm, based on a set of preset physical thresholds calibrated through numerous experiments, accurately classifies each pulse into one of the following four basic types: effective erosion discharge, short-circuit pulse, open-circuit pulse, or destructive arc discharge. Specifically, when the peak voltage of the pulse is within the normal discharge voltage range (e.g., 20–30 volts) and the breakdown delay time is greater than a minimum effective delay, it is determined to be an effective erosion discharge; when the peak voltage is lower than the short-circuit threshold and the peak current is greater than a minimum current threshold, it is determined to be a short-circuit pulse; when the peak voltage is close to the power supply open-circuit voltage and the peak current is lower than the open-circuit threshold, it is determined to be an open-circuit pulse; when the pulse duration exceeds the preset upper limit of a normal discharge cycle, and the voltage fluctuates continuously at a low plateau, while the turbidity index (DCI) shows an abnormally rapid increase in a short period of time, it is determined to be a destructive arc discharge. After classification, the module assigns a category label to each pulse and integrates it with the original feature data into an enhanced feature vector, which serves as the input for subsequent processing.

[0011] The gap state trajectory prediction module based on a long short-term memory network receives an enhanced feature vector sequence arranged in chronological order output by the real-time classification and feature extraction module of the discharge pulse, and uses it to probabilistically predict the state evolution trend of the discharge gap over a future period. The core of this module is a long short-term memory network (LSTM) model that has been trained offline and fine-tuned online. The LSTM model's network structure includes an input layer, two stacked LSTM layers, and an output layer. Each LSTM layer contains 256 hidden units. The model's input is a sliding window with a time step of 100, containing an enhanced feature vector sequence of 100 consecutive discharge pulses. The model's output layer uses the Softmax activation function, outputting a four-dimensional probability distribution vector. The four components of this vector correspond to the predicted probability that the next discharge pulse will be an effective erosion discharge, a short-circuit pulse, an open-circuit pulse, or a destructive arc discharge. By inputting the enhanced feature vector sequence at the current time step into the LSTM model, the module can predict the risk probability of gap state deterioration under the current processing parameters, thus providing a forward-looking basis for control decisions.

[0012] The multi-objective model predictive control module (MPC) is the decision-making core of the entire intelligent control system. This module receives the future state probability distribution output by the gap state trajectory prediction module and, in conjunction with a preset multi-objective optimization function, calculates the optimal combination of pulse power supply parameters for the next control cycle through online rolling optimization. The multi-objective optimization function is defined as the mathematical expression of the processing benefit J: ; in, The expected value representing the material erosion rate. The expected value representing the tool electrode wear rate. The expected values ​​representing the surface roughness of the workpiece are calculated using a preset physical model associated with the control variables. This is the total probability of future abnormal discharges (short circuit, open circuit, arc) predicted by the LSTM module; , , , These are weighting coefficients, which can be adaptively adjusted according to the current processing stage (roughing, semi-finishing, finishing). , , It is calculated using a pre-established empirical or semi-empirical physical model related to pulse parameters (peak current, pulse width, pulse interval); The MPC controller, within a preset prediction time domain (e.g., the next 20 pulse cycles), uses the peak current, pulse width, and pulse interval of the pulse power supply as control variables, and a physical model and an LSTM prediction model as the system state transition model. Within a constraint space containing the physical limits of each parameter (e.g., current range 1-50 amperes, pulse width range 2-500 microseconds), it solves for a control sequence that maximizes the cumulative optimization function J in the future prediction time domain. This solution process is completed within each control cycle (e.g., every 5 milliseconds) using an efficient Sequential Quadratic Programming (SQP) algorithm. Ultimately, the MPC module outputs only the first combination of control parameters in the calculated control sequence to the high-frequency pulse power supply execution module.

[0013] The high-frequency pulse power supply execution module receives digital instructions from the multi-objective model predictive control module and accurately generates one or more high-voltage pulses conforming to the instruction parameters, which are then applied between the tool electrode and the workpiece. The main power topology of this module employs a full-bridge inverter circuit, with silicon carbide (SiC) MOSFETs as the switching elements to ensure extremely low switching losses and extremely fast switching speeds (rise / fall time less than 50 nanoseconds). A dedicated digital signal processor (DSP) receives the parameter settings from the MPC module and generates high-precision pulse width modulation (PWM) signals in real time based on these values. This precisely controls the on / off state of the SiC MOSFET, thereby generating processing pulses with set peak current, pulse width, and pulse interval during the discharge gap. This module constitutes the final execution end of a complete closed-loop control circuit.

[0014] As another aspect of the present invention, the present invention also provides a method for intelligent regulation of discharge energy of an EDM machine, which is implemented by running the above-mentioned system and specifically includes the following steps: Step S100: Through the multimodal gap state sensing module, the discharge gap voltage waveform, current waveform, and optical turbidity signal of the working fluid flowing through the gap region are synchronously acquired during the processing.

[0015] Step S200: The high-throughput data synchronous acquisition and preprocessing module is used to digitize the acquired multi-source sensor signals at high speed and synchronously, and the hardware logic is used to calculate the waveform characteristic parameters and corresponding turbidity index of each discharge pulse in real time to generate a structured pulse data frame.

[0016] Step S300: The discharge pulse real-time classification and feature extraction module parses each pulse data frame based on a deterministic classification algorithm in a multi-dimensional feature space, accurately classifies the pulse into effective erosion discharge, short-circuit pulse, open-circuit pulse or destructive arc discharge, and generates an enhanced feature vector containing category labels.

[0017] Step S400: Input the continuously generated enhanced feature vector sequence into the gap state trajectory prediction module based on the long short-term memory network, and use the trained LSTM model to predict the probability distribution of various discharge states in a short time domain in the future.

[0018] Step S500: The multi-objective model prediction control module receives the predicted probability distribution and combines it with a multi-objective optimization function aimed at maximizing processing efficiency. Through online rolling optimization calculation, it solves for the next moment's pulse power supply parameter combination that optimizes the future processing process, namely the optimal peak current, pulse width, and pulse interval.

[0019] Step S600: The calculated optimal parameter combination instruction is sent to the high-frequency pulse power supply execution module, which accurately generates and applies processing pulses that conform to the instruction parameters.

[0020] Step S700: Repeat steps S100 to S600 in a loop to form a continuous, adaptive closed-loop control process until the entire processing task is completed.

[0021] The system and method proposed in this invention have the following beneficial effects: By introducing in-situ optical turbidity sensing, this invention incorporates for the first time a non-electrical signal that directly reflects the physical chip removal state within the gap into the control system. This forms a multi-modal information complement to the traditional electrical parameter signal, greatly improving the accuracy and dimensionality of the system's perception of the true state of the gap, and laying a solid data foundation for achieving refined and intelligent control. The predictive model based on LSTM networks transforms the control strategy from traditional reactive regulation based on historical information to feedforward regulation based on future state prediction. This forward-looking control paradigm enables the system to intervene and adjust energy parameters in advance before the gap state shows signs of deterioration, thereby actively avoiding unstable discharges such as short circuits and arcing, and significantly improving the stability of the processing. By employing a Model Predictive Control (MPC) architecture, the complex control problem is transformed into a constrained multi-objective online optimization problem. This architecture dynamically balances multiple conflicting performance indicators such as material removal rate, electrode wear, and surface quality at each control instant, and finds the Pareto optimal solution for the current state. This completely breaks through the inherent limitation of traditional technologies where efficiency and quality cannot be simultaneously achieved, enabling continuous optimization of overall processing efficiency throughout the entire process. The entire system, from sensing, processing, prediction to control, forms a high-frequency, real-time closed loop with a control cycle at the millisecond level. It can respond quickly to microsecond-level discharge phenomena, ensuring a precise match between energy supply and instantaneous gap demand, thereby maximizing processing efficiency while ensuring the highest processing quality. Attached Figure Description

[0022] The invention will now be further described with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic diagram of the system described in this invention; Figure 2 This is a flowchart illustrating the method described in this invention; Figure 3 This is a schematic diagram of the structure of the multimodal gap state sensing module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the in-situ optical turbidity sensing unit; Figure 5 This is a schematic diagram of the data processing flow of the high-throughput data synchronous acquisition and preprocessing module in this invention; Figure 6 This is a schematic diagram of the voltage and current waveforms of different types of discharge pulses in this invention; Figure 7 This is a schematic diagram illustrating the working principle of the gap state trajectory prediction module in this invention; Figure 8 This is a schematic diagram of the control principle of the multi-objective model prediction control module in this invention.

[0024] In the diagram: 100, Multimodal gap state sensing module; 110, High-bandwidth electrical parameter sensing unit; 111, Hall effect current sensor; 112, High-speed differential voltage probe; 120, In-situ optical turbidity sensing unit; 121, Flow cell structure; 122, Measurement channel; 123, Infrared light emitting component; 124, Photoelectric receiving component; 200, High-throughput data synchronous acquisition and preprocessing module; 300, Discharge pulse real-time classification and feature extraction module; 400, Gap state trajectory prediction module based on long short-term memory network; 500, Multi-objective model predictive control module; 600, High-frequency pulse power supply execution module. Detailed Implementation

[0025] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0026] Reference Figure 1 This invention proposes a complete architecture for an intelligent discharge energy control system for EDM machines. This system aims to achieve closed-loop adaptive control of discharge energy injection through real-time, multi-dimensional sensing and forward-looking prediction of the physical state of the discharge gap during EDM. This continuously optimizes multiple performance indicators such as material removal rate, tool electrode wear rate, and workpiece surface integrity in a dynamically changing machining environment. Logically, the entire system consists of several tightly coupled functional modules, including a multi-modal gap state sensing module 100, a high-throughput data synchronous acquisition and preprocessing module 200, a discharge pulse real-time classification and feature extraction module 300, a gap state trajectory prediction module 400 based on a long short-term memory network, a multi-objective model predictive control module 500, and a high-frequency pulse power supply execution module 600. These modules together constitute a complete, high-frequency closed-loop control loop from state perception to decision execution.

[0027] Specifically, the multimodal gap state sensing module 100 is the sensing foundation of the entire system. Its core task is to capture multi-source heterogeneous signals that can comprehensively characterize the complex physicochemical processes within the discharge gap without distortion and synchronously. In one specific embodiment, the module 100 is precisely integrated into the working area of ​​the electrical discharge machining tool, as shown in reference... Figure 3 As shown, it comprises two functionally complementary sensing units: a high-bandwidth electrical parameter sensing unit 110 and an in-situ optical turbidity sensing unit 120. The high-bandwidth electrical parameter sensing unit 110 is responsible for monitoring the electrical characteristics of the discharge process, and it consists of a Hall effect current sensor 111 and a high-speed differential voltage probe 112; The Hall effect current sensor 111, exemplarily, can be a closed-loop Hall sensor with a bandwidth exceeding 1 MHz and a response time of less than 0.5 microseconds. It is connected in series in the main discharge circuit between the pulse power supply output and the tool electrode (or workpiece) to accurately measure the instantaneous current waveform flowing through the discharge gap. The high-speed differential voltage probe 112, with an input impedance designed to be greater than 10 megohms to minimize the load effect on the measurement circuit, has its measurement endpoint directly connected between the clamping end of the tool electrode and the workpiece clamping base to capture the instantaneous voltage waveform between the tool and the workpiece in real time. The collaborative operation of these two electrical parameter sensors provides the system with the raw data foundation for key electrical characteristics of each discharge pulse, such as breakdown delay, pulse duration, peak current, and peak voltage.

[0028] Furthermore, in order to overcome the limitations of traditional control systems that rely solely on electrical parameters, the in-situ optical turbidity sensing unit 120 aims to directly quantify the chip removal state of the discharge gap, which is a physical quantity that has a decisive influence on processing stability but is difficult to reflect directly through electrical signals. In some embodiments, refer to Figure 4 The mechanical body of the sensing unit 120 is a flow cell structure 121 precision injection molded from polyetheretherketone (PEEK) material. PEEK was chosen because it exhibits excellent chemical inertness, high strength, and good electrical insulation properties against various electrical discharge machining (EDM) working fluids (such as kerosene-based or water-based media). The flow cell structure 121 has a standardized fluid interface, allowing for easy connection in series with the return pipeline of the EDM working fluid circulation filtration system. Its installation position is optimized to be downstream of the EDM tank drain outlet to ensure immediate capture of the working fluid carrying the latest EDM products. The flow cell structure 121 has an integrally molded, smooth-walled, precisely 10 mm diameter straight-through measurement channel 122. An infrared light emitting component 123 and a photoelectric receiving component 124 are arranged radially opposite each other on the wall of the measurement channel 122. The infrared light emitting component 123 internally encapsulates a high-efficiency infrared light-emitting diode with a center wavelength of 940 nm and a spectral half-width of 40 nm, whose radiation intensity can reach 750 mW / spherical degree. The 940 nm band was chosen because infrared light of this wavelength has good penetration in common working fluid media, and can be effectively absorbed and scattered by electro-erosion products (mainly micron-sized carbon black particles and metal spherical particles). The corresponding photoelectric receiving component 124 has a silicon-based PIN photodiode with a photosensitive area of ​​10 square millimeters and a peak response wavelength precisely matched to 940 nm. This diode has an extremely low dark current of less than 1.5 nanoamps, ensuring a good signal-to-noise ratio for detecting weak light signals. In some embodiments, when the working fluid carrying the electro-erosion products flows through the measurement channel 122, the suspended particles absorb and scatter the infrared beam, causing the light intensity reaching the photoelectric receiving component 124 to decrease. By measuring this degree of decrease, the system can quantitatively assess the turbidity of the working fluid in real time, thereby obtaining a physical index that directly reflects the concentration of electro-erosion products and chip removal efficiency within the discharge gap.

[0029] All analog signals from the multimodal gap state sensing module 100 are sent to the high-throughput data synchronous acquisition and preprocessing module 200. This module 200 is a bridge connecting the physical world and the digital control core. Its design goal is to complete the synchronous digitization and preliminary feature extraction of massive amounts of data at the hardware level, so as to reduce the burden of subsequent software processing and ensure the real-time determinism of the entire system. In a preferred embodiment, the hardware core of module 200 is a customized processing platform based on a field-programmable gate array (FPGA), exemplarily using Xilinx's Artix-7 series chips. This FPGA platform integrates multiple high-performance analog-to-digital converters (ADCs). Specifically, two independent analog input channels are connected to the outputs of the Hall effect current sensor 111 and the high-speed differential voltage probe 112, respectively. Each channel is equipped with an ADC with a sampling rate of up to 50 MHz and a vertical resolution of 16 bits. Such a high sampling rate ensures that the rapid rise and fall edges of microsecond-level pulses can be reconstructed without distortion. The other analog input channel is connected to the output of the transimpedance amplifier circuit at the back end of the photodetector 124. Since the frequency of the turbidity signal change is much lower than that of the electrical parameters, this channel is equipped with an ADC with a sampling rate of 100 kHz and a vertical resolution of 16 bits. In some embodiments, refer to Figure 5 In the data processing flow, the hardware description language (such as VHDL or Verilog) logic embedded inside the FPGA first performs a critical task: synchronization. Based on a clock signal generated by a unified, highly stable internal crystal oscillator, each data sampling point from the three ADC channels is timestamped with a precision down to the nanosecond level, thus ensuring the timing consistency of multi-source data from the source. Next, the FPGA's parallel processing capabilities were used for real-time digital signal processing of the high-speed sampled voltage and current waveform data streams. Hardware-implemented finite impulse response (FIR) digital filters were first used to filter out high-frequency noise introduced by the sensor; Subsequently, through hardware-implemented peak detection algorithms (exemplarily, finding the maximum value within a sliding window), zero-crossing detection algorithms, and numerical integration algorithms based on the trapezoidal rule, the FPGA efficiently calculates a series of key waveform characteristic parameters for each discharge pulse cycle, including the peak current Ip, peak voltage Vp, breakdown delay time td (the duration from the rising edge of the pulse voltage to voltage collapse), pulse duration ton, single-pulse energy (the integral of the product of voltage and current over time), and the integral area of ​​the voltage waveform during the pulse. In some embodiments, for low-speed data streams from the optical turbidity sensing unit, the FPGA performs a moving average filtering algorithm to smooth the signal and generate a stable, quantized turbidity index (DCI). Finally, for each identified discharge pulse, the FPGA packages the associated synchronization timestamp, all calculated waveform feature parameters, and the corresponding turbidity index (DCI) into a predefined structured data frame, and streams these data frames to the upper-layer discharge pulse real-time classification and feature extraction module 300 with extremely low latency through the onboard PCIe high-speed bus interface.

[0030] The function of the discharge pulse real-time classification and feature extraction module 300 is to perform deep analysis and semantic annotation of the raw feature data from the underlying hardware. This module is typically deployed on an embedded computing unit running a real-time operating system (RTOS) to ensure the determinism and timeliness of its processing. Its core is to run a deterministic classification algorithm based on a multi-dimensional feature space. The input to this algorithm is the structured data frame representing each independent discharge event transmitted from the FPGA module 200. Based on a set of preset physical thresholds carefully calibrated through extensive offline experiments and physical model analysis, the algorithm accurately classifies each pulse into one of four basic types, each type having a clear indicative meaning for the processing state. In some embodiments, refer to Figure 6 The typical waveforms shown represent four types: effective erosion discharge, which is a spark discharge that occurs normally and can effectively remove workpiece material; short-circuit pulse, which is a low-impedance path formed between the tool and the workpiece in contact with the electro-erosion products or themselves; open-circuit pulse, which is an unloaded pulse that fails to break down due to excessive gap or excessive dielectric strength; and destructive arc discharge, which is a malignant discharge form with an abnormally long duration, concentrated energy, and serious damage to the workpiece surface. Based on practical scenarios, the classification logic can be set as follows: When the peak voltage Vp of a pulse is within a normal discharge voltage range (e.g., between 20V and 30V), and its breakdown delay time td is greater than a minimum effective delay (e.g., 1.0 microseconds, to distinguish it from a spurious discharge caused by momentary contact), the pulse is determined to be an effective erosion discharge. When the peak voltage Vp is lower than a preset short-circuit voltage threshold (e.g., 5V), and the peak current Ip is greater than a minimum conduction current threshold (e.g., 1 Ampere), it is determined to be a short-circuit pulse. When the peak voltage Vp is close to the power supply's open-circuit voltage (e.g., 80V), and the peak current Ip is lower than a very low open-circuit current threshold (e.g., 0.5 Ampere), it is determined to be an open-circuit pulse. It should be noted that for the identification of destructive arc discharge, this invention utilizes the advantages of multimodal information fusion: when the duration ton of a pulse significantly exceeds the currently set pulse width (e.g., reaching 150% of the set value), and its voltage waveform maintains a low plateau (e.g., below 15 volts) for a long time after breakdown, while the turbidity index DCI shows an abnormally steep upward trend in the short term (e.g., within the past 100 pulses), the system comprehensively determines that a destructive arc discharge has occurred. This criterion, which integrates the time-domain characteristics of electrical parameters and the dynamic trend of turbidity, greatly improves the accuracy of identifying concealed arcs. After classification, module 300 adds a category label to each pulse data frame (e.g., using integers 0, 1, 2, and 3 to represent the four types respectively), and integrates it with the original waveform feature data and turbidity index DCI to form a higher-dimensional, more information-rich enhanced feature vector. This vector fully describes the physical properties and consequences of a single discharge event and serves as input to the subsequent prediction module.

[0031] The gap state trajectory prediction module 400 based on a Long Short-Term Memory (LSTM) network is the core of this invention for achieving forward-looking control. It receives a sequence of enhanced feature vectors arranged chronologically, output in real time by module 300. Its goal is no longer to analyze events that have already occurred, but to probabilistically predict the state evolution trend of the discharge gap over a short period in the future. The core of this module is a Long Short-Term Memory (LSTM) network model that has been thoroughly trained offline and fine-tuned online. As a special type of recurrent neural network (RNN), the LSTM network's unique gating mechanism (input gate, forget gate, output gate) makes it highly adept at learning and remembering long-term dependencies in time-series data, which has an inherent physical correlation with the evolution of the discharge state during electrical discharge machining. In a specific network structure design, this LSTM model includes an input layer, two vertically stacked LSTM layers, and a fully connected output layer. Each LSTM layer contains 256 hidden units (neurons), a depth and width sufficient to capture complex dynamic patterns in the discharge sequence. In some embodiments, refer to Figure 7 As shown, the input to this model is a sliding window with a time step of 100. That is, at any time t, the input to the model is a sequence of enhanced feature vectors corresponding to 100 consecutive discharge pulses from time t-99 to t. The output layer of the model uses the Softmax activation function, and its output is a four-dimensional probability distribution vector. The four components of this vector correspond to the predicted probabilities that the next (i.e., at time t+1) discharge pulse is an effective erosion discharge, a short-circuit pulse, an open-circuit pulse, or a destructive arc discharge, and the sum of the four probability values ​​is 1. Understandably, by inputting the latest processing state sequence into this pre-trained LSTM model, module 400 can predict in advance the probability of the gap state evolving towards instability (i.e., short circuit, open circuit, or arcing) under the current combination of processing parameters. This quantitative prediction of future states provides crucial decision-making basis for the control system to shift from passive response to active prevention. The offline training process of this LSTM model is based on a massive experimental database, which contains hundreds of millions of pulse feature vector sequences with precise classification labels collected under different materials, electrodes, and processing parameters.

[0032] The Multi-Objective Model Predictive Control (MPC) module 500 is responsible for formulating the optimal control strategy. This module receives the future state probability distribution vector output from the prediction module 400 and, in conjunction with a predefined multi-objective optimization function, calculates the optimal combination of pulse power supply parameters for the next control cycle through an online rolling optimization method. In some embodiments, refer to Figure 8 The control principle of this module lies in its optimization framework. First, a mathematical objective function J is defined to comprehensively evaluate the efficiency of the processing, which can take the form: ; In this function, The expected value representing the material erosion rate. E[Ra] represents the expected value of the tool electrode wear rate, and E[Ra] represents the expected value of the workpiece surface roughness Ra. These three are key indicators for measuring processing efficiency and quality. This is the total probability of future abnormal discharges (short circuit, open circuit, arc) predicted by the LSTM module 400, i.e. This represents the stability risk of the processing procedure. , , , These are the weighting coefficients for each item. These coefficients are not fixed but can be adaptively adjusted according to the current processing stage. In one example, during the roughing stage, the system assigns a value representing efficiency. A larger value; while in the finishing stage, it will increase the value representing surface quality. The weights. For The three physical quantities, E[Ra], are calculated online using pre-established empirical or semi-empirical physical models related to the pulse power supply parameters (mainly peak current Ip, pulse width ton, and pulse interval tooff). These models can be fitted with a large amount of experimental data. The MPC controller is optimized within a preset prediction time domain (e.g., looking ahead N=20 pulse cycles). It uses the peak current Ip, pulse width ton, and pulse interval tooff of the pulse power supply as adjustable control variables, and the aforementioned physical models and LSTM prediction models together as a state transition model describing how the system state evolves with control input. Within a constrained space containing the physical limits of various control variables (exemplarily, current range 1-50 amperes, pulse width range 2-500 microseconds, and pulse interval range 2-500 microseconds), the MPC controller solves a complex dynamic optimization problem: finding a sequence of control variables {Ip(k), ton(k), toff(k)} for k=t to t+N-1 that maximizes the sum of the accumulated optimization function J over the entire prediction time domain. This solution process must be completed within each control cycle (e.g., every 5 milliseconds), thus requiring an efficient numerical optimization algorithm, such as Sequential Quadratic Programming (SQP). After obtaining the optimal control sequence in each control cycle, the MPC module, following its fundamental principles of "rolling time" and "receding horizon," outputs only the first combination of control parameters in the sequence—the optimal {Ip, ton, toff} setpoints for the next time step (k=t)—to the high-frequency pulse power supply execution module 600. In the next control cycle, the entire process will be repeated based on the new system state measurements.

[0033] The high-frequency pulse power supply execution module 600 is the final execution end of the closed-loop control circuit. It is responsible for accurately converting the digital instructions from the MPC module 500 into high-voltage pulses applied between the tool electrode and the workpiece. To meet the requirements of the MPC module for high-frequency, wide-range dynamic parameter adjustment, this power supply module has been optimized in terms of topology and component selection. For example, its main power topology employs a full-bridge inverter circuit, and the power switching element is a new generation of wide-bandgap semiconductor device—silicon carbide (SiC) MOSFETs. Compared to traditional silicon-based IGBTs, SiC MOSFETs have significantly lower on-resistance and switching losses, as well as much faster switching speeds; their pulse rise and fall times can be controlled to within 50 nanoseconds. This characteristic ensures that the power supply can generate very steep, well-formed rectangular wave pulses and operate efficiently over an extremely wide frequency and duty cycle range. A dedicated digital signal processor (DSP), such as Texas Instruments' (TI) C2000 series, serves as the module's local controller. It receives digital setpoints for peak current, pulse width, and pulse interval from the MPC module 500 via a high-speed serial interface. Based on these settings, the firmware algorithm inside the DSP generates high-precision, high-frequency pulse-width modulation (PWM) signals in real time. These PWM signals, through an isolated gate drive circuit, precisely control the turn-on and turn-off timing of four SiC MOSFETs, thereby forming a processing pulse sequence with the precise peak current, pulse width, and pulse interval required by the instruction at the transformer secondary, i.e., across the discharge gap. The module's rapid response and high-precision execution capabilities are the key physical guarantee for ensuring the ultimate realization of the control effect of the entire intelligent control system.

[0034] As another aspect of the present invention, the present invention also provides a corresponding intelligent control method for EDM discharge energy, which is implemented by running the above-mentioned system, and its detailed process is as follows: Figure 2 As shown, the specific steps include: Step S100: After the system starts processing, the multimodal gap status sensing module 100 starts working continuously, synchronously collecting the discharge gap voltage waveform, current waveform, and optical turbidity signal of the working fluid flowing through the gap area during the processing.

[0035] Step S200: After receiving these analog signals, the high-throughput data synchronous acquisition and preprocessing module 200 immediately performs high-speed and synchronous digitization on them, and uses the hardware parallel processing capability of the FPGA to calculate the waveform characteristic parameters of each discharge pulse and the time-aligned turbidity index DCI in real time, and finally generates a structured pulse data frame stream.

[0036] Step S300: The real-time classification and feature extraction module 300 of discharge pulses receives and parses these data frames one by one. Based on its internal multi-dimensional feature space deterministic classification algorithm, it accurately classifies each pulse into effective erosion discharge, short-circuit pulse, open-circuit pulse or destructive arc discharge, and generates an enhanced feature vector containing category labels.

[0037] Step S400: The sequence of enhanced feature vectors continuously generated over a past period (e.g., 100 pulses) is treated as a whole and input into the gap state trajectory prediction module 400 based on a long short-term memory network. This module uses a pre-trained LSTM model for forward propagation calculation to predict the probability distribution of various discharge states in a short time domain in the future.

[0038] Step S500: The multi-objective model prediction and control module 500 receives the probability distribution containing future risk warnings and uses it as one of the key inputs, substituting it into the preset multi-objective optimization function J. Subsequently, within its prediction time domain, the module solves for the next-moment pulse power supply parameter combination that maximizes the overall benefits of the future processing through online rolling optimization calculations (e.g., the SQP algorithm), namely the optimal peak current Ip, pulse width ton, and pulse interval toff.

[0039] Step S600: The MPC module 500 sends the calculated set of optimal parameter instructions to the high-frequency pulse power supply execution module 600 via a high-speed communication interface. The DSP of this module immediately parses the instructions and accurately generates the corresponding PWM control signals to drive the power circuit to apply one or more processing pulses conforming to the instruction parameters to the discharge gap.

[0040] Step S700: After completing one control-execution cycle, the system immediately returns to step S100, continuously collecting new status information and repeatedly executing the entire process from steps S100 to S600. Through this high-frequency, millisecond-level loop, a continuous, adaptive closed-loop control process is formed until the entire processing task is completed.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent control system for discharge energy of an electrical discharge machine, applied to an electrical discharge machining tool, characterized in that, The system includes: A multimodal gap state sensing module (100) is disposed in the working area of ​​the electrical discharge machining tool to synchronously acquire multi-dimensional physical signals related to the state of the discharge gap; The high-throughput data synchronous acquisition and preprocessing module (200) is connected to the signal output terminal of the multimodal gap state sensing module (100) and is used to synchronously digitize the multidimensional physical signal and perform hardware-level feature calculation to generate structured pulse data frames. The real-time classification and feature extraction module (300) for discharge pulses is connected to the high-throughput data synchronous acquisition and preprocessing module (200) for receiving and parsing the pulse data frame, classifying each discharge pulse according to a preset physical threshold, and generating an enhanced feature vector containing category labels. The gap state trajectory prediction module (400) based on the long short-term memory network is connected to the discharge pulse real-time classification and feature extraction module (300) for receiving the enhanced feature vector sequence arranged in time order and using a long short-term memory network model to probabilistically predict the state evolution trend of the discharge gap in the future. The multi-objective model prediction control module (500) is connected to the gap state trajectory prediction module (400) based on the long short-term memory network. It is used to receive the probabilistic prediction results and, in combination with a preset multi-objective optimization function, calculate the optimal combination of pulse power supply parameters in the next control cycle through online rolling optimization. The high-frequency pulse power supply execution module (600) is connected to the multi-objective model prediction control module (500) and is used to receive the optimal pulse power supply parameter combination instruction and accurately generate a machining pulse that conforms to the instruction parameters and apply it between the tool electrode and the workpiece.

2. The intelligent control system for discharge energy of an EDM machine according to claim 1, characterized in that, The multimodal gap state sensing module (100) includes: A high-bandwidth electrical parameter sensing unit (110) is used to capture the instantaneous current and voltage waveforms during each discharge pulse; An in-situ optical turbidity sensing unit (120) is connected in series in the return pipeline of the processing working fluid circulation filtration system. It is used to quantify the turbidity of the working fluid caused by electro-erosion products in real time by measuring the degree of attenuation of light of a specific wavelength by the working fluid.

3. The intelligent control system for discharge energy of a spark dynamometer according to claim 2, characterized in that, The structure of the in-situ optical turbidity sensing unit (120) includes: A flow cell structure (121) is injection molded from polyetheretherketone material, which has a standard fluid interface and is connected in series in the return pipeline. The flow cell structure (121) has a straight-through measurement channel (122) integrally formed inside. An infrared light emitting component (123) is installed on one side wall of the measuring channel (122) and perpendicular to the direction of the working fluid flow; A photoelectric receiving component (124) is mounted on the other side wall of the measurement channel (122) opposite the infrared light emitting component (123) for receiving infrared light attenuated by the working fluid.

4. The intelligent control system for discharge energy of a spark dynamometer according to claim 3, characterized in that, The infrared light emitting component (123) is internally encapsulated with an infrared light-emitting diode; the photoelectric receiving component (124) is internally encapsulated with a silicon-based PIN photodiode; the measurement channel (122) has a diameter of 10 mm.

5. The intelligent control system for discharge energy of an EDM machine according to claim 2, characterized in that, The core of the high-throughput data synchronous acquisition and preprocessing module (200) is a hardware processing platform based on a field-programmable gate array (FPGA). The platform includes: The instantaneous current and voltage waveforms are digitized using two analog-to-digital converters with a sampling rate of 50 MHz and a vertical resolution of 16 bits. The signal output by the in-situ optical turbidity sensing unit (120) is digitized using an analog-to-digital converter with a sampling rate of 100 kHz and a vertical resolution of 16 bits. All data streams are timestamped using a unified internal clock; Using a hardware-implemented algorithm, the peak current, peak voltage, breakdown delay time, pulse duration, and single-pulse energy of each pulse cycle are calculated. The optical turbidity signal is mean filtered to generate a quantized turbidity index; The synchronization timestamp, waveform feature data, and turbidity index corresponding to each pulse are then packaged into the structured pulse data frame.

6. The intelligent control system for discharge energy of a spark dynamometer according to claim 1, characterized in that, The real-time classification and feature extraction module (300) for discharge pulses internally runs a deterministic classification algorithm based on a multi-dimensional feature space. This algorithm accurately classifies each discharge pulse into one of the following four types based on a set of preset physical thresholds: Effective erosion discharge: determined when the peak voltage of the pulse is within the normal discharge voltage range and the breakdown delay time is greater than the minimum effective delay; Short-circuit pulse: Determined when the peak voltage is lower than the short-circuit threshold and the peak current is greater than the minimum current threshold; Open circuit pulse: Determined when the peak voltage is close to the power supply open-circuit voltage and the peak current is lower than the open circuit threshold; Destructive arc discharge: When the pulse duration exceeds the preset upper limit, the voltage continues to fluctuate at a low level, and the turbidity index calculated by the high-throughput data synchronous acquisition and preprocessing module (200) shows an abnormally rapid increase in a short period of time, a comprehensive judgment is made.

7. The intelligent control system for discharge energy of a spark dynamometer according to claim 1, characterized in that, The long short-term memory network model used in the gap state trajectory prediction module (400) based on the long short-term memory network has a network structure including an input layer, two stacked LSTM layers, each containing 256 hidden units, and an output layer. The input of the model is a sliding window with a time step of 100, and the data in the window is the enhanced feature vector sequence of 100 consecutive discharge pulses. The output layer of the model uses the Softmax activation function to output a four-dimensional probability distribution vector. The four components of this vector correspond to the predicted probability that the next discharge pulse is an effective erosion discharge, a short-circuit pulse, an open-circuit pulse, or a destructive arc discharge.

8. The intelligent control system for discharge energy of a spark dynamometer according to claim 1, characterized in that, The multi-objective model prediction control module (500) calculates the optimal combination of pulse power supply parameters within a preset prediction time domain, using the peak current, pulse width, and pulse interval of the pulse power supply as control variables, and within a constraint space containing the physical limits of each parameter, by solving a dynamic optimization problem aimed at maximizing the accumulated multi-objective optimization function value within the prediction time domain.

9. The intelligent control system for discharge energy of a spark dynamometer according to claim 8, characterized in that, The multi-objective optimization function is defined as the mathematical expression for the processing benefit J: ; in, The expected value representing the material erosion rate. The expected value representing the tool electrode wear rate. The expected values ​​representing the surface roughness of the workpiece are calculated using a preset physical model associated with the control variables. It is the total probability of future abnormal discharges predicted by the gap state trajectory prediction module (400) based on the long short-term memory network; , , , These are weighting coefficients that can be adaptively adjusted according to the processing stage.

10. A method for intelligent regulation of discharge energy in an EDM machine, applicable to the intelligent regulation system for discharge energy in an EDM machine as described in any one of claims 1-9, characterized in that, The method includes the following steps: S100: Through the multi-modal gap state sensing module (100), the discharge gap voltage waveform, current waveform and optical turbidity signal of the working fluid flowing through the gap region are simultaneously acquired during the processing. S200: Through the high-throughput data synchronous acquisition and preprocessing module (200), the acquired multi-source sensor signals are digitized at high speed and synchronously, and the waveform characteristic parameters and corresponding turbidity index of each discharge pulse are calculated in real time using hardware logic to generate structured pulse data frames. S300: Through the real-time classification and feature extraction module (300) of discharge pulses, each pulse data frame is analyzed based on a deterministic classification algorithm in a multi-dimensional feature space, and the pulse is accurately classified into effective erosion discharge, short-circuit pulse, open-circuit pulse or destructive arc discharge, and an enhanced feature vector containing category labels is generated. S400: The continuously generated enhanced feature vector sequence is input into the gap state trajectory prediction module (400) based on the long short-term memory network. Using a trained long short-term memory network model, the probability distribution of various discharge states in a short time domain in the future is predicted in real time. S500: The multi-objective model prediction control module (500) receives the predicted probability distribution and combines it with a multi-objective optimization function aimed at maximizing processing efficiency. Through online rolling optimization calculation, it solves the next moment pulse power supply parameter combination that can optimize the future processing process, namely the optimal peak current, pulse width and pulse interval. S600: The calculated optimal parameter combination command is sent to the high-frequency pulse power supply execution module (600), which accurately generates and applies processing pulses that conform to the command parameters; S700: Repeat steps S100 to S600 in a loop to form a continuous, adaptive closed-loop control process until the processing task is completed.

Citation Information

Patent Citations

  • Electric spark multivariable self-adaptive efficient machining control system and method

    CN119387729A

  • Precise automatic multi-axis numerical control electric spark forming machine data processing method and system

    CN120892935A

  • Pulse and gap control for electrical discharge machining equipment

    US20150360310A1

  • Control device for wire electric discharge machine, control program for wire electric discharge machine, and control method for wire electric discharge machine

    WO2024105760A1

Cited By

  • Electric spark machining efficiency analysis system based on historical working conditions

    CN121820804A