A surface treatment method for copper material by micro-arc oxidation

Abstract

The application relates to the technical field of metal material surface treatment, and particularly discloses a copper material micro-arc oxidation surface treatment method. The method comprises the following steps: applying initial pulse power excitation to the copper material in an electrolyte and synchronously performing high-speed optical imaging; performing morphology identification and feature quantification on a discharge arc spot by using a convolutional neural network; judging the discharge state in real time by a deep reinforcement learning model, and dynamically adjusting and controlling pulse power parameters to suppress destructive large arcs; forming a closed-loop feedback control until a compact oxidation film with a predetermined thickness is generated. By the technical scheme, microcosmic dynamic tracking and accurate intervention of the micro-arc discharge process are realized, the uniformity and compactness of the film layer are improved, the waste rate is reduced, and the method has good automation level and industrial application prospect.

Description

Technical Field

[0001] This invention belongs to the field of metal material surface treatment technology, specifically relating to a micro-arc oxidation surface treatment method for copper materials. Background Technology

[0002] With the continuous advancement of materials science and surface engineering technology, micro-arc oxidation, a key technology for in-situ growth of ceramic films on metal surfaces through plasma discharge, has demonstrated its application value in enhancing the performance of copper and its alloys. Copper holds an irreplaceable position in the fields of electronics and precision manufacturing due to its excellent electrical and thermal conductivity. Micro-arc oxidation technology, which constructs a protective coating on copper surfaces with high insulation, corrosion resistance, and wear resistance, has become a core process for improving the overall service life of copper components. This technology primarily involves applying instantaneous high voltage in an electrolyte environment to excite plasma micro-discharge on the workpiece surface, driving a chemical reaction and generating a dense oxide film.

[0003] Real-time control of the discharge state during the micro-arc oxidation process of copper is a key technology determining the final coating quality. This process aims to maintain the uniform distribution and stable evolution of the discharge arc spot on the copper surface through precise intervention of power output parameters. Due to the extremely high thermal conductivity of copper and the drastic impedance fluctuations during the surface film formation process, the system needs to analyze the morphological characteristics and energy distribution of the plasma discharge region in real time to ensure that the discharge energy can be accurately applied to the target area, realizing the transformation from uncontrollable discharge to controllable surface processing.

[0004] Existing technologies largely rely on fixed voltage or current parameters preset by human experience, making it difficult to adapt to the instantaneous and complex changes in the microscopic physicochemical environment of copper surfaces. Traditional monitoring methods cannot capture the arc spot morphology distribution and brightness density during the micro-arc discharge process, resulting in a lack of real-time sensing capability for the health of the discharge status. Conventional linear control logic cannot cope with nonlinear fluctuations that occur during the discharge process, making it difficult to capture and suppress sudden local hot spots. The lag in this control mode makes local heat accumulation highly susceptible to triggering destructive large arcs, breaking down the already formed microscopic film structure, ultimately leading to ablation and uneven coating thickness. Therefore, a copper surface treatment solution based on micro-arc oxidation is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for micro-arc oxidation surface treatment of copper materials, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for micro-arc oxidation surface treatment of copper materials, comprising the following specific steps: Step 1: Apply an initial pulse power supply excitation to the copper workpiece in an electrolyte environment to induce plasma micro-discharge on its surface, and simultaneously start a high-speed image acquisition device to perform continuous frame rate optical observation of the discharge area; Step 2: Based on the continuous image sequence acquired by the high-speed image acquisition device, use a convolutional neural network model to perform morphological recognition, spatial distribution density analysis, and brightness intensity quantification of the discharge arc spots in each frame image to generate a multi-dimensional feature vector characterizing the current discharge state; Step 3: Input the multi-dimensional feature vector into a deep reinforcement learning decision model, which determines in real time whether the current discharge is in a uniform micro-arc state or has evolved into a destructive large arc state according to a preset discharge health evaluation criterion; Step 4: If it is determined to be a destructive large arc state, the deep reinforcement learning decision model outputs the corresponding power supply control command to dynamically adjust the pulse width, pulse frequency, and duty cycle parameters of the pulse power supply to suppress local overheating and restore uniform discharge; Step 5: Repeat steps 1 to 4 to form a closed-loop feedback control process until a dense oxide film layer of a predetermined thickness is generated on the copper surface.

[0007] Preferably, the high-speed image acquisition device in step 1 is an industrial-grade optical imaging device with a frame rate of not less than 10,000 frames per second. Its photosensitive element has a high dynamic range response capability and can clearly capture arc spot details in micro-arc discharge scenarios where strong light and weak light coexist.

[0008] Preferably, the convolutional neural network model in step 2 is trained with a large number of labeled samples, which can accurately distinguish between normal micro-arc spots and abnormal arc spots, and extract key visual features including spot area, edge sharpness, center brightness gradient and neighborhood distribution entropy.

[0009] Preferably, the deep reinforcement learning decision model in step 3 adopts a policy optimization architecture, whose state space is composed of the multi-dimensional feature vectors, whose action space corresponds to the discretized combination of the adjustable parameters of the pulse power supply, and whose reward function is set comprehensively based on the discharge uniformity index and the film growth rate.

[0010] Preferably, the power control command in step 4 is sent to the pulse power module in real time through a digital signal processor. This module can reconstruct the output waveform in microseconds to ensure that the control response speed matches the transient characteristics of the discharge process.

[0011] Preferably, the electrolyte environment includes an alkaline main solvent and various metal salt additives. Its formulation has been experimentally verified to induce stable and controllable micro-arc discharge behavior on the copper surface and promote the orderly arrangement of oxide crystal structures.

[0012] Preferably, the parameter settings of the initial pulse power supply excitation are not fixed, but are adaptively initialized according to the geometry, surface roughness and historical processing records of the copper workpiece, so as to shorten the time required for the system to enter the steady-state control stage.

[0013] Preferably, the closed-loop feedback control process includes an abnormal interruption mechanism. When the state of uniform discharge cannot be restored by parameter adjustment after multiple consecutive detections, the processing is automatically terminated and an alarm signal is triggered to prevent the workpiece from being scrapped due to continuous burning.

[0014] Preferably, the determination of the end point of the formation of the dense oxide film layer not only depends on the preset processing time, but also combines the trend of film layer impedance change obtained by real-time monitoring. When the impedance growth rate is lower than a certain preset threshold, it is considered that the film layer growth is approaching saturation.

[0015] Preferably, the method can also be extended to the micro-arc oxidation treatment of other high thermal conductivity metal materials. Process transfer can be achieved simply by retraining the convolutional neural network model and the deep reinforcement learning decision model according to the physicochemical properties of different materials.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing a closed-loop control mechanism that combines high-speed image acquisition with intelligent algorithms, microscopic dynamic tracking and precise intervention of discharge arc spots during the micro-arc oxidation of copper materials are achieved, solving the problems of ablation and uneven coating caused by local heat accumulation in the traditional fixed parameter mode. 2. This method transforms the originally uncontrollable plasma discharge process into a highly controllable surface processing behavior, which improves the uniformity and density of the film while reducing the scrap rate. 3. Since the entire control process is fully automated and requires no human experience, it has good repeatability and potential for large-scale application, making it suitable for high-end electronic devices, precision connectors, and other industrial fields that require high-performance copper surfaces. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of the data flow between the functional entities according to the present invention; Figure 3 This is a flowchart illustrating the process of performing discharge spot morphology recognition, spatial distribution analysis, and multidimensional feature vector generation on a continuous image sequence acquired by a high-speed image acquisition device based on a convolutional neural network, according to the present invention. Figure 4This is a schematic diagram illustrating the principle framework of the present invention, which uses a deep reinforcement learning decision model to evaluate the discharge health in real time and output pulse power supply control commands to suppress destructive large electric arcs. Figure 5 This is a flowchart illustrating the dynamic reconstruction of pulse power module waveforms and the determination of dense oxide film growth saturation based on a closed-loop feedback mechanism according to the present invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] In this embodiment, the implementation process of a copper micro-arc oxidation surface treatment method is entirely based on a closed-loop dynamic feedback mechanism. Through the deep coupling of high-frequency visual perception and artificial intelligence decision-making, microscopic intervention on plasma discharge behavior is achieved.

[0020] In step 1, an initial pulse power supply is applied to the copper workpiece in an electrolyte environment to excite plasma micro-discharge on its surface, and a high-speed image acquisition device is simultaneously started to perform continuous frame rate optical observation of the discharge area.

[0021] In performing step 1, an electrolyte environment adapted to the physicochemical properties of copper is constructed. This electrolyte environment employs an alkaline aqueous solution system, specifically comprising: potassium hydroxide at a concentration ranging from 5 g / L to 15 g / L as the main electrolyte, providing necessary conductive ions and adjusting the pH of the solution to between 10 and 13; sodium silicate at a concentration ranging from 10 g / L to 30 g / L, providing the silicon source required for film formation during discharge and promoting the formation of ceramic phase oxides; and sodium dihydrogen phosphate at a concentration ranging from 2 g / L to 8 g / L, acting as a complexing agent and buffer to help stabilize the discharge reaction rate. Sodium aluminate at a concentration ranging from 1 g / L to 3 g / L and an additive at a concentration ranging from 0.5 g / L to 1.5 g / L are also added to the electrolyte. This additive is a mixture of sodium fluoride and glycerol in a 1:3 mass ratio, designed to reduce the arc-starting voltage on the copper surface and improve the density of the oxide film.

[0022] During the initial pulse power supply excitation process, the parameters of the pulse power supply are adaptively initialized based on the geometric characteristics of the copper workpiece. For workpieces with complex shapes and high surface area to volume ratios, the initial peak voltage is set between 350 volts and 450 volts, the initial pulse frequency is set between 400 Hz and 800 Hz, and the initial duty cycle is set between 5% and 15%. If the workpiece surface roughness is large, the initial voltage is increased by 10 to 20 volts to ensure that the electrolyte film at the surface protrusions can be broken down. The power supply excitation waveform adopts a bidirectional pulse mode; the positive pulse is used to maintain plasma discharge, and the negative pulse is used to eliminate charge accumulation on the workpiece surface and prevent excessively intense partial discharge.

[0023] A high-speed image acquisition device, activated synchronously, is positioned directly above or at the side observation window of the micro-arc oxidation reactor. This device employs an industrial-grade high-speed camera with a frame rate of 10,000 to 20,000 frames per second. Its image sensor is based on complementary metal-oxide-semiconductor (CMOS) technology and possesses a high dynamic range response of at least 120 dB. This characteristic ensures that, even in scenarios where the strong scintillation spot generated by micro-arc discharge coexists with a relatively dark background liquid environment, the camera can retain details in dark areas and suppress overexposure in bright areas. A narrow-band filter is positioned in front of the camera lens, with its center wavelength set at 450 nm or 650 nm. This filter aims to filter out light scattered by impurities in the electrolyte, retaining only the characteristic spectral images representing the excited-state radiation of specific metal elements.

[0024] In step 2, based on the continuous image sequence acquired by the high-speed image acquisition device, a convolutional neural network model is used to perform morphological recognition, spatial distribution density analysis, and brightness intensity quantification on the discharge arc spots in each frame of the image, thereby generating a multi-dimensional feature vector characterizing the current discharge state.

[0025] In step 2, the convolutional neural network model acts as a visual feature extractor. Each frame of the acquired raw image undergoes a preprocessing module, including grayscale conversion, median filtering for noise reduction, and histogram equalization to enhance the contrast between the arc spots and the background. The processed image is then input into a pre-trained deep convolutional neural network. This network architecture consists of multiple alternating convolutional layers, pooling layers, and non-linear activation layers. In the convolutional layers, multiple 3×3 or 5×5 convolutional kernels are used to perform sliding convolution operations on the image, extracting the edges, textures, and local geometric structures of the arc spots.

[0026] The model can accurately distinguish between normal micro-arc spots and abnormal arc spots. Normal micro-arc spots appear as clusters of pixels that are small in size, uniform in brightness, and highly dispersed in the image; while abnormal arc spots (i.e., destructive large arcs) are characterized by rapidly expanding area, brightness far exceeding a preset threshold, jagged radiating edges, and spatial stagnation between consecutive frames. Key visual features extracted by the model include: Spot area characteristics: Calculate the total number of pixels in all identified discharge regions in a single frame image, and calculate the maximum and average areas of a single independent spot.

[0027] B. Edge sharpness characteristics: The gradient change rate of the spot boundary is calculated using the Sobel operator or the Laplace operator to characterize the intensity of the discharge process.

[0028] C. Center brightness gradient: Analyze the brightness distribution curve that decreases from the center of the spot to the edge to determine the concentration of energy release.

[0029] D. Neighborhood distribution entropy: By calculating the entropy value of the probability density function of the spatial distribution, the spatial uniformity of the arc spot distribution on the copper surface is quantified. The higher the entropy value, the more uniform the distribution.

[0030] All the above features are mapped and concatenated into a multidimensional feature vector, which serves as a digital mapping of the discharge state within that time slice. This vector not only contains the spatial static features at the current moment, but also, through the introduction of a temporal difference operator, includes the rate of change of arc spot morphology between adjacent frames.

[0031] In step 3, the multidimensional feature vector is input into the deep reinforcement learning decision model, which then determines in real time whether the current discharge is in a uniform micro-arc state or has evolved into a destructive large arc state based on the preset discharge health evaluation criteria.

[0032] In performing step 3, the deep reinforcement learning decision model employs a proximal policy optimization (PPO) architecture as its core algorithm. This model treats the micro-arc oxidation control process as a Markov decision process. Its state space is composed of the multi-dimensional feature vectors generated in step 2, which can comprehensively reflect the instantaneous physical state of the discharge system. The model internally includes a policy network and an evaluation network.

[0033] The discharge health evaluation criterion is a comprehensive quantitative indicator. The health score is set between 0 and 1; the closer the score is to 1, the closer the discharge is to an ideal micron-level uniform distribution; the lower the score, the more the system deviates from a steady state, posing a risk of evolving into a large-scale arc. The evaluation logic is as follows: when the number of arc spots per unit area is greater than a preset density threshold, and the average brightness variation coefficient of a single spot is less than a preset fluctuation threshold, the system is considered to be in a healthy state. Conversely, if the brightness integral value within a specific coordinate region continuously increases, and the geometric centroid of that region shifts very little within 10 consecutive frames, the system automatically determines that heat accumulation has occurred at that location, and the health score decreases.

[0034] The reward function design of the deep reinforcement learning decision-making model comprehensively considers both discharge uniformity and film growth rate. A positive reward is given if the decision action maintains high health without reducing the system's average current density; a negative penalty is given if a large arc causes the health to drop below the alarm limit. Through this self-playing training method, the model learns how to predict the arc evolution trend under different copper surface conditions.

[0035] In step 4, if the state is determined to be a destructive large electric arc, the deep reinforcement learning decision model outputs the corresponding power supply control command to dynamically adjust the pulse width, pulse frequency and duty cycle parameters of the pulse power supply in order to suppress local overheating and restore uniform discharge.

[0036] When performing step 4, once the decision model identifies an abnormal decline in health, its action space immediately outputs a set of optimized power supply parameter combinations. The action space corresponds to a discretized combination of adjustable pulse power supply parameters. For example, the pulse width adjustment step is set to 1 microsecond, and the frequency adjustment step is set to 50 Hz.

[0037] The specific power control logic includes: a. Pulse Width Suppression: If local overheating is determined, the decision model prioritizes reducing the positive pulse width, compressing it from the current 100 microseconds to 40 microseconds or even lower. Reducing the pulse width limits the total energy injected into the workpiece surface in a single pulse, allowing sufficient time for heat in the micro-region to diffuse to the substrate and electrolyte.

[0038] b. Frequency Tuning: While reducing the pulse width, appropriately increase the pulse frequency to maintain a constant overall average current. Increasing the frequency increases the frequency of energy input per unit time, which helps to induce more tiny discharge centers on the surface and prevents energy from accumulating at a single fault point.

[0039] c. Duty cycle optimization: By adjusting the ratio of pulse on-time to off-time, the duration of the cooling phase is increased, thereby suppressing the continued existence of large electric arcs at the microscale.

[0040] The power regulation command is sent to the pulse power module via a digital signal processor (DSP) within microseconds. This module integrates a high-speed insulated-gate bipolar transistor (IGBT) array, capable of reconstructing the output waveform within 10 microseconds of receiving the command. This rapid response ensures that the system can extinguish or reshape the destructive arc into a micro-arc through active energy redistribution before it causes substantial ablation damage to the oxide film.

[0041] Step 5 involves repeating steps 1 to 4 to form a closed-loop feedback control process until a dense oxide film of a predetermined thickness is formed on the surface of the copper material.

[0042] During step 5, the system maintains high-frequency cyclic monitoring. To determine the film formation endpoint, this embodiment introduces a real-time film impedance monitoring mechanism. During the negative phase of the pulsed power supply, the instantaneous values ​​of induced voltage and current are obtained through a sampling circuit, and the complex impedance of the workpiece surface is calculated. Due to the extremely high resistivity of the oxide film, the total impedance of the system exhibits a monotonically increasing trend as the film thickness and density increase.

[0043] The logic for determining the endpoint of dense oxide film formation is as follows: The system records the impedance change curve with processing time in real time and calculates its first derivative (i.e., impedance growth rate). When the impedance value reaches the preset target threshold and its growth rate is lower than a preset small threshold (e.g., less than 2% per minute) for 120 consecutive seconds, it is determined that the film growth has reached saturation, and the system automatically enters the slow voltage reduction and shutdown stage.

[0044] The closed-loop feedback control process includes an abnormal interruption mechanism. If, during execution, the discharge health fails to recover to above the safe range after 30 consecutive parameter adjustments by the deep reinforcement learning model, or if an uncontrolled, sudden surge in current is detected, the system will immediately trigger hard protection, forcibly disconnecting the power module output and issuing an audible and visual alarm signal. This mechanism prevents the scrapping of copper workpieces due to algorithm failure or extreme operating conditions, ensuring the safety of the production process.

[0045] Throughout the process, due to the high thermal conductivity of copper, the electrolyte temperature is controlled within a range of 25 degrees Celsius ± 2 degrees Celsius through a heat exchange circulation system. The flow rate of the circulation pump is adjusted according to the surface area of ​​the workpiece to ensure that the electrolyte is circulated at least 3 times per minute. The flow washes away the heat and microbubbles generated in the discharge area, and together with the aforementioned intelligent control algorithm, ensures the uniformity of oxide film formation.

[0046] Example 2: Based on Example 1, Example 2 further refines the technical details of multi-scale feature fusion of convolutional neural network model and deep reinforcement learning model, and demonstrates the adaptive implementation process of the method on copper workpieces with different geometric structures.

[0047] For copper parts with complex microporous structures or narrow, deep groove structures, the visual characteristics of discharge arc spots exhibit high spatial limitations. In performing step 2, the convolutional neural network model used in this embodiment introduces a multi-scale dilated convolution structure. Dilated convolution expands the receptive field without increasing computational complexity by inserting holes between the convolution kernel parameters, enabling the model to capture both the microscopic edge features of a single arc spot and the long-range spatial relationships between arc spots in different regions.

[0048] To address the generation of multidimensional feature vectors, this embodiment adds a "partial discharge clustering feature" based on convolutional feature map statistics. This feature identifies regions with excessively high or low discharge activity by calculating the discharge probability distribution of different grid regions in the image. Specifically, the quantization method is as follows: the original 512×512 pixel image is divided into 16×16 sub-regions, and the proportion of pixels in each sub-region whose brightness exceeds 5 times the background average is counted. If the difference in the proportion between adjacent sub-regions exceeds a preset difference threshold, the corresponding uniformity component in the feature vector will decrease, triggering the decision model in step 3 to intervene.

[0049] In the deep reinforcement learning decision-making process of step 3, this embodiment employs a recurrent reinforcement learning architecture including a Long Short-Term Memory (LSTM) network layer. Since micro-arc oxidation is a time-dependent physical process, the current discharge state is often influenced by the heat accumulation of the previous few pulse cycles. By introducing an LSTM layer, the model can memorize the state trajectory within the past 500 milliseconds. The state space contains not only the current visual feature vector but also the action history of the past three time steps. The reward function of the decision model is further refined, and its mathematical expression is translated into the following textual description: The reward value equals the sum of the discharge uniformity component and the film growth efficiency component, minus the energy consumption penalty term. The discharge uniformity component is proportional to the spatial distribution entropy of the spots in the image; the film growth efficiency component is proportional to the logarithm of the average current density; and the energy consumption penalty term is proportional to the square of the peak voltage. Through this complex reward mapping, the model can automatically optimize the control path with the lowest energy consumption while ensuring film formation quality.

[0050] Specifically, in step 4, power control, when the system is applied to process large-size planar copper plates, the evolution of the electric arc often exhibits lateral drift characteristics. The control commands not only involve adjusting pulse parameters but also the coordinated control of multiple parallel pulse channels. The pulse power module is divided into multiple independently controlled power sub-units. When the high-speed camera detects signs of arc accumulation in the left region of the copper plate, the DSP-issued command specifically reduces the pulse frequency of the corresponding left electrode group while keeping the parameters of the right region unchanged. This regionalized differential control capability improves surface consistency when processing large workpieces.

[0051] In step 5 of the film formation endpoint determination, this embodiment, in addition to relying on the impedance change trend, also introduces feedback determination based on visual features. As the oxide film gradually thickens and densifies, the plasma discharge on its surface exhibits a transformation from "large-particle, high-brightness arc spots" to "fine, dense, flickering sparks." The convolutional neural network calculates the average diameter of the sparks in the image in real time. When the average diameter of the identified spots decreases and the spot density per unit area reaches saturation (i.e., the overlap of spot positions within adjacent pulse cycles is extremely high), combined with the determination condition that the impedance growth rate is below a threshold, the system determines that the film formation process is complete.

[0052] The method in this embodiment also possesses excellent process transfer capabilities. When the processed object is changed from pure copper to tellurium copper alloy or chromium zirconium copper alloy, only the image data and corresponding impedance data generated during the first 10 minutes of processing of the new material need to be fed back to the model. Utilizing transfer learning technology, the neural network only needs to be trained with a small number of samples to adapt to the discharge characteristics of the new material. For example, tellurium copper alloy produces a lower characteristic spectral brightness during discharge; the model will automatically adjust the camera's exposure compensation coefficient and redefine the brightness gradient threshold in the health evaluation criteria to ensure that the control accuracy is not affected by changes in material composition.

[0053] Example 3: This Example 3 describes a large-scale automated production line implementation scheme based on the method of the present invention, focusing on multi-station collaborative control and ensuring the long-term operational stability of the system.

[0054] In step 1, to adapt to industrial-scale mass production, a high-speed image acquisition device is integrated onto a gantry-type support with autofocus and multi-dimensional translation capabilities. Whenever a new copper workpiece enters the reaction station, the support system automatically adjusts the camera focal length via a laser rangefinder to ensure the discharge area remains within the optimal imaging plane. The initial pulse power supply excitation parameters are set not only based on the workpiece geometry but also incorporate real-time monitoring of the electrolyte conductivity and temperature. The electrolyte management system automatically detects the charge carrier concentration every 60 seconds and corrects the initial voltage setting in real time based on minute drifts in conductivity, with fluctuations controlled within ±5 volts.

[0055] In the algorithm processing stages of steps 2 and 3, this embodiment employs a strategy combining edge computing and cloud-based model updates. Due to the massive amount of image data (10,000 frames per second), a high-performance graphics processing unit (GPU) cluster is deployed on-site for real-time inference. The feature extraction time for each frame of the convolutional neural network is compressed to less than 0.5 milliseconds, ensuring that the total latency of the feedback loop does not exceed 2 milliseconds. This response speed is sufficient to capture and intervene in the microsecond-level arc transition behavior during the discharge process. All feature vectors and corresponding control decisions are encrypted and uploaded to the central server. Offline incremental training of the deep reinforcement learning model is performed through big data analysis. After processing 1,000 products, the system automatically pushes updated model parameters to the edge to continuously optimize the system's decision-making intelligence.

[0056] At the hardware execution level of step 4, the pulse power module employs a high-frequency series resonant inverter topology. Compared to traditional chopper power supplies, this structure can generate square wave pulses with steeper edges and a rise time of less than 1 microsecond. This fast-response hardware foundation enables the precise execution of complex pulse sequences issued by deep reinforcement learning models. For example, the model can generate a special "pulse cluster" pattern, where 10 high-frequency short pulses are followed by long pulses. The impact effect of the short pulses induces a large number of nucleation sites on a highly thermally conductive copper substrate, and the long pulses promote the growth of oxide crystals, resulting in a bilayer oxide film with both hardness and toughness.

[0057] At the end of the closed-loop process in step 5, this embodiment adds a sub-step for online surface quality assessment based on machine vision. After the power is completely turned off, the high-speed camera switches to high-resolution static imaging mode and scans the generated oxide film surface using multi-angle structured light. The system calculates the uniformity of surface reflectivity and the consistency of color texture, combined with the average health record during processing, to generate a unique digital quality profile for each workpiece. If the average health of a product falls below the safety threshold multiple times during processing, the system marks it as "key re-inspection" in the digital label, realizing full lifecycle quality traceability of the production process.

[0058] This embodiment also features in-depth optimization for photoelectric interference in industrial environments. A hardware-level synchronization trigger circuit is added between steps 1 and 2. This circuit ensures precise synchronization between the camera's shutter opening moment and the forward conduction moment of the pulse power supply, with an offset of less than 10 nanoseconds. This captures the moment of strongest plasma luminescence, improving the signal-to-noise ratio. For the problem of chemical bubble obstruction caused by electrolyte, a dynamic masking operator is integrated into the algorithm. This operator can identify and eliminate false arc spot features generated by floating bubbles by calculating the motion vectors of pixels, ensuring the authenticity and reliability of the data input to the reinforcement learning decision model.

[0059] Throughout the closed-loop feedback control process, the system not only focuses on the current film quality but also possesses predictive maintenance capabilities. By analyzing the long-term evolution of the power module output current ripple, the decision model can identify the degree of electrochemical loss of the electrode plates. When the ripple amplitude exceeds the preset normal range, the system will prompt the operator to check the electrode spacing or replace the filter element of the circulation system, ensuring the process reproducibility of the surface treatment method during thousands of hours of continuous operation.

[0060] Example 4: This example details the application process of the method of the present invention on copper alloy components with extremely high thermal conductivity (such as copper liners of aerospace engine thrust chambers). This process has extremely high requirements for the uniformity and adhesion of the film layer.

[0061] For this type of workpiece, in step 1, the electrolyte environment is further optimized to a low-temperature micro-arc oxidation system. Nanoparticle additives, such as nano-alumina or nano-zirconia particles, are introduced into the electrolyte at a concentration of 2 grams per liter. These particles are suspended in the liquid phase and, under the combined action of electrophoretic force and pressure waves generated by plasma discharge, are mechanically embedded into the pores of the growing oxide film. To facilitate this physical process, the initial pulse power supply is set to an asymmetric AC mode, with the absolute value of the negative voltage set to 30% of the positive voltage, aiming to remove the loose layer on the film surface through periodic cathodic ablation.

[0062] In step 2, the convolutional neural network model adds the ability to analyze the chromaticity of the discharge spectrum. Since different intensities of plasma discharge correspond to different electron temperatures, the proportions of their emitted spectral components will vary. The model estimates the effective electron temperature gradient of the micro-region discharge by analyzing the pixel value ratios of the red, green, and blue channels in the image. This temperature gradient feature, together with the spatial distribution density and morphological features from step 2, forms an enhanced multidimensional feature vector, providing deeper physical information for subsequent deep reinforcement learning decisions.

[0063] In performing step 3, the reward function of the deep reinforcement learning decision model incorporates indirect control over the internal stress of the film. By adding sensor feedback of the workpiece substrate temperature to the real-time cyclic feedback, the reward function includes not only a uniformity index but also a component negatively correlated with the substrate temperature rise rate. If the model's action causes the substrate temperature to rise too quickly, the reward value will decrease.

[0064] Specifically, in step 4, power regulation, the output of the pulse power module is configured as a three-level or five-level topology for this extreme operating condition. The deep reinforcement learning decision model can output more nuanced level transition instructions, such as adding a small step level to the falling edge of the main pulse to mitigate the sudden change in plasma pressure at discharge termination. This microsecond-level waveform detailing is crucial for eliminating micro-cracks on the copper surface.

[0065] In the closed-loop process of step 5, this embodiment also performs real-time feedback control on the surface roughness after film formation. By using a convolutional neural network to statistically analyze the uniformity of the distribution of fine sparks in the image online, the system can predict the surface roughness of the film. If the predicted value exceeds the set upper limit, the decision model will automatically increase the duty cycle of the negative pulse, utilizing the weak electrolytic polishing effect of the negative current to perform real-time micro-smoothing on the surface of the generated oxide film. When the film impedance increases and the impedance change tends to be extremely small, the system automatically stops processing and enters the electrolyte circulation cooling and cleaning program to ensure that the electrolyte remaining on the workpiece surface is quickly removed, preventing secondary corrosion.

[0066] The method in this embodiment is not limited to processing copper materials, but has also been successfully applied to the surface of aluminum-based composite materials with similar thermal conductivity. Experiments show that by retraining the convolutional neural network model and the deep reinforcement learning decision model according to the physicochemical properties of different materials, the system can achieve the same level of film uniformity control on this type of new material as it does on copper after only about 50 batches of adaptive learning. This demonstrates that the closed-loop control mechanism integrating computer vision and intelligent algorithms proposed in this invention has versatility and technical robustness, and can adapt to the surface treatment needs of various complex high thermal conductivity metal materials.

[0067] 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. A method for micro-arc oxidation surface treatment of copper materials, characterized in that, Includes the following steps: An initial pulse power supply is applied to the copper workpiece in an electrolyte environment to induce plasma micro-discharge on the surface of the copper workpiece, and a high-speed image acquisition device is simultaneously activated to perform continuous frame rate optical observation of the discharge area. Based on the continuous image sequence acquired by the high-speed image acquisition device, a convolutional neural network model is used to perform morphological recognition, spatial distribution density analysis and brightness intensity quantification on the discharge arc spots in each frame of the image, generating a multi-dimensional feature vector characterizing the current discharge state. The multidimensional feature vector is input into the deep reinforcement learning decision model, which then determines whether the current discharge is in a uniform micro-arc state or has evolved into a destructive large arc state based on a preset discharge health evaluation criterion. If the state is determined to be a destructive large electric arc, the deep reinforcement learning decision model outputs the corresponding power supply regulation command to dynamically adjust the parameters of the pulse power supply in order to suppress local overheating and restore uniform discharge. The above steps are repeated to form a closed-loop feedback control process until a dense oxide film layer of a predetermined thickness is generated on the surface of the copper workpiece.

2. The method for micro-arc oxidation surface treatment of copper materials according to claim 1, characterized in that: The electrolyte environment consists of an alkaline aqueous solution system, which includes a main electrolyte, a film-forming silicon source, a complexing buffer, a film-forming aid, and mixed additives. The main electrolyte is potassium hydroxide, which is used to provide conductive ions and adjust the pH of the electrolyte environment to a preset alkaline range. The silicon source for film formation is sodium silicate, which is used to provide a silicon source for the generation of ceramic phase oxides; The complexing buffer is sodium dihydrogen phosphate, used to stabilize the discharge reaction rate; the film-forming aid is sodium aluminate; the mixed additive is a mixture of sodium fluoride and glycerol in a preset mass ratio, used to reduce the arc-starting voltage on the surface of the copper workpiece. The parameters of the initial pulse power supply excitation are adaptively initialized based on the geometry, surface roughness, and historical processing records of the copper workpiece. When the copper workpiece has a preset geometric feature of high surface area to volume ratio, a higher initial peak voltage, initial pulse frequency, and initial duty cycle are set. When the surface roughness of the copper workpiece exceeds the preset reference value, the liquid film at the surface protrusion is broken down by increasing the initial peak voltage. The initial pulse power supply excitation adopts a bidirectional pulse mode, using a positive pulse to maintain the plasma micro-discharge and a negative pulse to eliminate the charge accumulation on the surface of the copper workpiece.

3. The method for micro-arc oxidation surface treatment of copper materials according to claim 2, characterized in that: The high-speed image acquisition device is placed directly above or at the side observation window of the micro-arc oxidation reaction tank, and adopts a high-speed camera based on complementary metal-oxide-semiconductor technology. Its photosensitive element has a preset high dynamic range response capability, which is used to preserve dark details and suppress overexposure in bright areas in the scenario where the strong scintillation spot generated by the plasma micro-discharge coexists with the dark background liquid phase environment. The high-speed camera has a narrow-band filter in front of its lens, with its center wavelength set in the nanometer band. This filter is used to filter the light scattered by impurities in the electrolyte and retain the characteristic spectral image that characterizes the excited state radiation of a specific metal element. The high-speed image acquisition device is integrated on a gantry bracket with automatic focusing and multi-dimensional translation capabilities. The camera focal length is automatically adjusted by a laser rangefinder to ensure that the discharge area is within the imaging plane. A hardware-level synchronization trigger circuit is provided between the high-speed image acquisition device and the pulse power supply to ensure that the camera shutter opening time is synchronized with the forward conduction time of the pulse power supply.

4. The method for micro-arc oxidation surface treatment of copper materials according to claim 3, characterized in that: The convolutional neural network model serves as a visual feature extractor, performing preprocessing operations on each frame of the image. These preprocessing operations include grayscale conversion, median filtering for noise reduction, and histogram equalization. The convolutional neural network model includes multiple alternating convolutional layers, pooling layers, and nonlinear activation layers; in the convolutional layers, multiple convolutional kernels of preset sizes are used to perform sliding convolution operations on the image to extract the edges, textures, and local geometric structures of the arc spots; For copper workpieces with complex microporous structures or narrow deep groove structures, the convolutional neural network model introduces a multi-scale dilated convolutional structure to expand the receptive field by inserting holes between convolutional kernel parameters, which is used to capture the micro-edge features of individual arc spots and the long-range spatial correlation between arc spots in different regions. The multidimensional feature vector also includes partial discharge clustering features based on convolutional feature map statistics. By dividing the original image into multiple grid-like sub-regions, the proportion of pixels in each sub-region whose brightness exceeds the background average by a preset multiple is counted.

5. The method for micro-arc oxidation surface treatment of copper materials according to claim 4, characterized in that: The multidimensional feature vector includes the following visual features: First, the spot area feature is obtained by calculating the sum of all pixels in the discharge regions identified in a single frame image and by statistically analyzing the maximum and average areas of a single independent spot. Second, edge sharpness characteristics: the gradient change rate of the spot boundary is calculated using the gradient operator to characterize the intensity of the discharge process. Third, the central brightness gradient is used to determine the concentration of energy release by analyzing the brightness distribution curve that decreases from the center of the spot to the edge. Fourth, neighborhood distribution entropy: by calculating the probability density function entropy value of the spatial distribution, the spatial distribution uniformity of the arc spot on the surface of the copper workpiece is quantified. The probability density function entropy value is positively correlated with the distribution uniformity. The multidimensional feature vector incorporates the rate of change of arc spot morphology between adjacent frames by introducing a temporal difference operator; The convolutional neural network model estimates the effective electron temperature gradient of micro-region discharge by analyzing the proportion of chromaticity components of arc spots in the image in real time, and uses the effective electron temperature gradient as a component of the multidimensional feature vector.

6. The method for micro-arc oxidation surface treatment of copper materials according to claim 5, characterized in that: The deep reinforcement learning decision model adopts a proximal policy optimization architecture and defines the micro-arc oxidation regulation process as a Markov decision process. The state space of the deep reinforcement learning decision model is composed of the multidimensional feature vectors, the action space corresponds to the discretized combination of the adjustable parameters of the pulse power supply, and the reward function is set based on the discharge uniformity index and the film growth rate. The deep reinforcement learning decision model adopts a recurrent reinforcement learning architecture that includes a long short-term memory network layer to memorize the state trajectory within a preset time. Its input state space includes the current visual feature vector and the action history within a preset step size. The reward function is defined as follows: the reward value is equal to the sum of the discharge uniformity component and the film growth efficiency component, minus the energy consumption penalty term; The discharge uniformity component is proportional to the spatial distribution entropy of the spots in the image, the film growth efficiency component is proportional to the logarithm of the average current density, and the energy consumption penalty term is proportional to the square of the peak voltage.

7. The method for micro-arc oxidation surface treatment of copper materials according to claim 6, characterized in that: The discharge health evaluation criterion is a quantitative indicator. When the number of arc spots per unit area is greater than a preset density threshold and the average brightness variation coefficient of a single spot is less than a preset fluctuation threshold, the system is judged to be in a healthy state. If the brightness integral value in a specific coordinate region continues to rise, and the displacement of the geometric centroid of that specific coordinate region within a consecutive preset number of frames is less than a preset displacement threshold, then it is determined that heat accumulation has occurred, and the score corresponding to the discharge health evaluation criterion decreases. The deep reinforcement learning decision model analyzes the multidimensional feature vectors to predict the evolution trend of the electric arc, and outputs an optimized combination of power parameters when it detects an abnormal decline in health. The adjustment logic of the power supply parameter combination includes: limiting the total energy injected into the workpiece surface by reducing the positive pulse width, increasing the pulse frequency to maintain a constant average current, and adjusting the ratio of pulse on-time to off-time to increase the duration of the cooling phase.

8. The method for micro-arc oxidation surface treatment of copper materials according to claim 7, characterized in that: The power control command is sent to the pulse power module through a digital signal processor. The pulse power module integrates a high-speed insulated gate bipolar transistor array to reconstruct the output waveform within a preset microsecond time. The pulse power module adopts a high-frequency series resonant inverter topology to generate square wave pulses with a rise time shorter than a preset duration. When processing large-size planar copper plate workpieces, the pulse power module is divided into multiple independent and controlled power sub-units. The deep reinforcement learning decision model outputs regionalized differential control commands to reduce the pulse frequency for electrode groups that show signs of arc accumulation. The deep reinforcement learning decision model is also used to generate pulse group patterns, that is, a long pulse is followed by a preset number of high-frequency short pulses. The impact effect of the short pulses is used to induce nucleation points on the copper workpiece, and the long pulses are used to promote the growth of oxide crystals.

9. The method for micro-arc oxidation surface treatment of copper materials according to claim 8, characterized in that: The closed-loop feedback control process introduces a real-time monitoring mechanism for film impedance. During the negative phase of the pulse power supply, the instantaneous values ​​of induced voltage and current are obtained through the sampling circuit, and the complex impedance of the copper workpiece surface is calculated. The logic for determining the endpoint of the formation of the dense oxide film is as follows: the system records the curve of the change of the complex impedance with the processing time in real time, and calculates the order derivative of the complex impedance, i.e. the impedance growth rate. When the value of the complex impedance reaches a preset target threshold, and the impedance growth rate is continuously lower than a preset small threshold for a preset duration, the film growth is determined to be saturated. The determination of the film formation endpoint also incorporates visual feature feedback. The average diameter of the sparks in the image is calculated in real time through the convolutional neural network. When the average diameter of the identified spots shrinks to below the preset ratio of the initial stage and the spot density per unit area reaches saturation, the film formation process is determined to be complete in conjunction with the determination condition of the complex impedance.

10. The method for micro-arc oxidation surface treatment of copper materials according to claim 9, characterized in that: The closed-loop feedback control process is equipped with an abnormal interruption mechanism. When the deep reinforcement learning decision model adjusts the parameters for a preset number of times and the discharge health still fails to recover to the safe range, or when an uncontrolled sudden surge in current is detected, the power module output is forcibly disconnected and an alarm signal is issued. The copper micro-arc oxidation surface treatment method also includes an online surface quality assessment step based on machine vision. After the power is turned off, the high-speed camera is switched to a high-resolution static imaging mode, and the generated oxide film surface is scanned by multi-angle structured light to calculate the uniformity of surface reflectivity and the consistency of color texture, and to generate a quality digital file for each workpiece. The copper micro-arc oxidation surface treatment method analyzes the evolution of the output current ripple of the pulse power module to identify the degree of electrochemical loss of the electrode plate, and triggers a maintenance prompt when the ripple amplitude exceeds the normal range. The electrolyte environment is temperature-controlled within a preset fluctuation range through a heat exchange circulation system. The flow rate of the circulation pump is adjusted according to the surface area of ​​the workpiece to remove heat and bubbles generated in the discharge area.

Images