Flexible powder metallurgy press production process optimization control system

Abstract

The invention belongs to the technical field of powder metallurgy processes, and particularly relates to a flexible powder metallurgy press production process optimization control system, which comprises a central processing unit which establishes a command and execution cooperative relationship with modules through a system bus and is responsible for distributing tasks, transmitting parameters and receiving feedback for various modules, the overall orderly and efficient operation of the system is realized; the data sensing module comprises a powder feeding unit, a pressing unit, a pressure maintaining unit, a demolding unit and a whole-process online detection unit, multiple types of sensors are arranged, whole-process real-time data sensing is achieved, collected data are preprocessed in real time through edge computing nodes, and data time sequence consistency is ensured; according to the execution driving module, a plurality of piezoelectric ceramic actuators are evenly arranged at the connecting end of the punch and the press main shaft in the circumferential direction, the actuators are controlled to stretch out and draw back through a high-voltage driving power source, and reverse compensation force is generated.

Description

Technical Field

[0001] This invention belongs to the field of powder metallurgy process technology, specifically a flexible powder metallurgy press production process optimization control system. Background Technology

[0002] As a key process in high-end manufacturing, the production process control level of powder metallurgy press, the core equipment of powder metallurgy technology, directly determines the quality stability and production efficiency of high-precision parts such as automotive gearbox gears and aerospace structural components.

[0003] As downstream industries continue to upgrade their demands for component dimensional tolerances, density uniformity, and production efficiency, the shortcomings of existing powder metallurgy press control systems in process coordination capabilities are becoming increasingly prominent, and have become a key bottleneck restricting the high-quality development of the industry.

[0004] Existing systems typically deploy sensors for a single process, failing to cover the entire process from powder feeding, pressing, pressure holding, demolding, and inspection. Low data acquisition frequency and asynchronous timing prevent the linkage and analysis of parameters across processes. When data fluctuates, the pressing unit cannot predict and adjust pressure parameters in advance, directly causing excessive product density deviation. Furthermore, high-speed pressing results in significant punch vibration, leading to fluctuations in green compact density and increased scrap rates. Traditional passive vibration reduction systems suffer from response delays and cannot meet dynamic compensation requirements. Therefore, this invention provides a flexible powder metallurgy press production process optimization control system. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a flexible powder metallurgy press production process optimization control system, comprising: The central processing unit (CPU) establishes a command and execution coordination relationship with modules through the system bus. It is responsible for assigning tasks to various modules, transmitting parameters, and receiving feedback to achieve the overall orderly and efficient operation of the system. The power module supplies power to the central processing unit; The data sensing module includes a powder feeding unit, a pressing unit, a pressure holding unit, a demolding unit, and a full-process online detection unit. It is equipped with multiple types of sensors to achieve real-time data sensing throughout the entire process. The collected data is preprocessed in real time through edge computing nodes to ensure data time sequence consistency. The collaborative hub module builds a data linkage model based on industrial Ethernet, and realizes collaborative linkage of data by using process correlation analysis, real-time bottleneck location, and cross-unit data sharing. The DRL model building module aims to maximize the overall pass rate and minimize energy consumption to construct DQN agents. The RL decision-making process module utilizes offline pre-training, online real-time decision-making, and dynamic exploration mechanisms to achieve collaborative control. The execution layer response module receives parameter instructions from the RL agent and then adjusts the parameters at the millisecond level through the servo control system to ensure action synchronization. The feedback optimization closed-loop module includes: real-time feedback of quality data from the online detection unit, parameter optimization linkage between the collaborative hub and the RL agent, and storage and reuse of the optimal parameter combination. The adjusted product quality data is fed back to the collaborative hub in real time through the online detection unit. The collaborative hub then transmits the adjustment action and quality feedback data back to the RL agent to update the reward value and optimize the model parameters. The vibration sensing module deploys multiple acceleration sensors on the top of the punch and the side wall of the mold to collect vibration acceleration signal data of the punch in the vertical and horizontal directions in real time. The data is then uploaded to the collaborative hub after being filtered by a low-pass filter. The LMS adaptive control module is integrated into the collaborative central algorithm layer and adjusts the compensation parameters in real time through the LMS adaptive filtering algorithm. The drive module is equipped with multiple piezoelectric ceramic actuators evenly arranged circumferentially at the connection end between the punch and the press spindle. The actuators are controlled to extend and retract by a high-voltage drive power supply to generate a reverse compensation force.

[0007] Preferably, the powder feeding unit includes functions for raw material confirmation, parameter matching, powder feeding start-up, and data acquisition, used to realize the conveying of powder raw materials; it also includes a material identification module for identifying, but not limited to, powder material and loose density attributes, located in the raw material confirmation stage of the powder feeding unit; The pressing unit includes functions such as powder feeding confirmation, dynamic adaptation of pressing parameters, pressing start, and connection of the pressure holding unit. It is a system module used to press the powder raw material in the mold cavity into shape. When the pressing force reaches the target value and the upper mold stroke is in place, the pressing unit sends a compliance signal to the coordination central function module, and the coordination central function module instructs the pressure holding unit to start the pressure holding process.

[0008] The pressure holding unit is a system module that performs pressure holding operations on the pressed product and transmits relevant data after the pressure holding is completed. The pressure holding unit sends the pressure holding data to the collaborative central function module, which forwards it to the demolding unit and triggers the RL decision layer module to predict the demolding resistance range.

[0009] The demolding unit is used to demold the product, transfer it to the next process, and upload data.

[0010] The demolding unit receives demolding parameters from the coordination center and fine-tunes the parameters based on the demolding resistance predicted by the RL decision layer. The full-process online detection unit is a system module that detects, classifies, and feeds back data on the finished products; the RL agent is an intelligent decision-making entity that updates the reward function based on the full-process data, optimizes the parameters for the next batch, and updates the parameter combination of qualified batches to the optimal parameter library.

[0011] The preferred approach includes RFID tag preprocessing for raw material packaging, inbound identification triggering, core identification, compliance judgment, and closed-loop connection, specifically implemented as follows: The raw material packaging RFID tag preprocessing involves completing the preparatory work before material identification through tag writing and tag attachment operations. The inbound identification trigger activates the reader through preset trigger conditions, causing the powder delivery unit to respond and enter the material receiving state; The core identification is based on RFID tag data reading and parsing, and sequentially performs the steps of tag wake-up and data reception, data parsing and format conversion, and data integrity verification. The compliance judgment is achieved by comparing the read material attributes with a preset rule base to determine the verification result, thereby realizing material attribute verification and matching. The closed-loop connection executes the qualified data upload operation based on the compliance judgment result, performs abnormal handling and linkage operation for abnormal situations, and performs the retry operation after the abnormality is resolved, ultimately achieving accurate identification of materials.

[0012] Preferably, the information written in the tag includes basic information and traceability information; the core verification dimensions for comparing the attributes with the rule base are compliance of loose packing density, validity of shelf life, and matching degree of material with production requirements; the anomaly handling and linkage operation specifically involves a dual operation of local alarm and abnormal information upload.

[0013] Preferably, the collaborative central functional module includes three sub-modules: process correlation analysis, real-time bottleneck location, and cross-unit data sharing. The implementation process of each sub-module is as follows: The process correlation analysis submodule sequentially executes the steps of preliminary data modeling and rule definition, real-time data input and preprocessing, dynamic correlation calculation and risk marking, correlation result output and downstream linkage. The bottleneck real-time location submodule achieves bottleneck location through steps of defining and monitoring abnormal triggering conditions, tracing and investigating multi-dimensional data, and confirming and outputting bottleneck results. The cross-unit data sharing submodule completes data sharing in the following order: definition of the scope and standards of shared data, sharing triggering mechanism and data matching, data sharing execution and feedback optimization.

[0014] Preferably, the DRL model construction module further includes: a state space. Action space Reward function ; The reward function The final basic formula for the reward function is determined as follows: ; in, Incentives will be given for improving the pass rate. As a reward for energy consumption reduction rate, Penalty for parameter adjustment range For parameter adjustment range, For energy consumption reduction rate, To improve the pass rate.

[0015] Preferably, it also includes an offline pre-training unit, an online real-time decision-making unit, and a dynamic exploration mechanism unit; the offline pre-training unit is used to build the decision foundation to realize the model's initial decision-making ability, and its execution process includes multi-dimensional dataset construction, DQN model architecture construction, phased training, and effect verification; the online real-time decision-making unit is used for dynamic and precise control to realize rapid parameter adjustment, and its execution process includes state acquisition synchronization, intelligent decision-making and verification, instruction issuance monitoring, and decision visualization; the dynamic exploration mechanism unit is used to cope with scene changes to balance exploration and stability, and its execution process includes trigger condition judgment, ε-greedy strategy execution, risk control, and result solidification.

[0016] Preferably, the execution layer response module is mainly divided into the following two steps: unit execution details and synchronization guarantee measures. At the same time, the unit execution details are mainly completed by the powder feeding unit, pressing unit and demolding unit.

[0017] Preferably, the online inspection unit is used to simultaneously collect three types of core quality data—dimensional accuracy data, internal quality data, and surface defect data—after completing product inspection, and ensures rapid response of the collaborative hub through data transmission and triggering mechanisms; after receiving the core quality data, the collaborative hub forms action-quality corresponding data pairs and processes them according to the reward function. The formula calculates the real-time reward value and sends the action-quality-reward value data back to the RL agent for storage in the experience replay pool. The parameter management unit is used to determine the corresponding parameters as the optimal parameter combination when the preset conditions are met for three consecutive generation cycles, and the collaboration center automatically generates an optimal parameter report.

[0018] Preferably, the central processing unit is electrically connected to the power module; the power module can supply power to the central processing unit so that it can operate normally.

[0019] The beneficial effects of this invention are as follows: 1. The present invention discloses a flexible powder metallurgy press production process optimization control system. This system breaks through the limitations of the traditional rigid production mode. Through full-process data collaboration and dynamic calling of the optimal parameter library, it realizes rapid adaptation for multi-variety, small-batch production. For product changeovers with different materials or structures, there is no need to rely on repeated manual disassembly and debugging. Based on the raw material attributes of the material identification module and the rapid exploration capability of the DRL model, parameter matching and verification can be completed within 30-60 minutes, shortening the changeover cycle by more than 80% compared to the traditional changeover cycle, effectively meeting the production needs of modern manufacturing industry for rapid iteration and multiple batches.

[0020] 2. The flexible powder metallurgy press production process optimization control system described in this invention uses real-time data linkage and DRL intelligent decision-making across the entire process of powder feeding, pressing, holding pressure, demolding, and detection to construct a dynamic mapping relationship between process parameters and product quality. It can locate bottlenecks such as uneven powder feeding and abnormal demolding resistance within two production cycles and automatically optimize parameters. In practical applications, the product defect rate can be reduced from 5%-12% to ≤2%, and the dimensional tolerance compliance rate can be stabilized at over 99.5%. Simultaneously, through multi-objective optimization, unit product energy consumption is reduced by 10%-15%, significantly reducing raw material waste and energy consumption, and lowering production costs.

[0021] 3. The flexible powder metallurgy press production process optimization control system described in this invention achieves a vibration amplitude of ≤2μm through the nanometer-level displacement resolution of the piezoelectric ceramic actuator and the real-time iteration of the LMS algorithm, which is more than 75% lower than the uncompensated state (3-8μm). This significantly improves the product dimensional accuracy and density uniformity. It works seamlessly with the original control system without modifying the main structure of the press. Only the sensor and actuator need to be added to the punch end. The total delay from vibration signal acquisition to compensation force output is ≤50μs, which can offset high-frequency sudden vibrations during the pressing process in real time, such as instantaneous impact vibrations caused by powder agglomeration. Attached Figure Description

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

[0023] Figure 1 This is a flowchart of the production process optimization control system in this invention; Figure 2 This is a flowchart illustrating the implementation steps of each unit in this invention. Detailed Implementation

[0024] 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.

[0025] like Figure 1As shown, an embodiment of the present invention provides a production process optimization control system for a flexible powder metallurgy press, comprising: The central processing unit (CPU) establishes a command and execution coordination relationship with modules through the system bus. It is responsible for assigning tasks to various modules, transmitting parameters, and receiving feedback to achieve the overall orderly and efficient operation of the system. The power module supplies power to the central processing unit, enabling it to operate normally. The data sensing module includes a powder feeding unit, a pressing unit, a pressure holding unit, a demolding unit, and a full-process online detection unit. It deploys multiple types of sensors. The powder feeding unit has a built-in laser flow sensor to collect real-time powder feeding rate and uniformity data; a weighing sensor to detect the remaining powder in the hopper and the amount of powder fed per batch; a pressing unit has a built-in pressure sensor to control the pressing force in real-time; a displacement sensor to detect the stroke accuracy of the upper and lower molds in real-time; a temperature sensor to detect the mold temperature; a pressure holding unit has a built-in pressure stability sensor; a timer to detect the pressure holding time deviation; and a demolding unit has a built-in displacement sensor to detect the demolding speed in real-time. The system includes a tension sensor to detect demolding resistance, an online inspection unit to record product dimensional tolerances using a laser profilometer, an ultrasonic density meter to detect internal density distribution, and a vision camera to detect surface defects in real time. The collected data is preprocessed in real time via edge computing nodes and uploaded to the collaborative hub at a frequency of 50ms / time to ensure data consistency and avoid collaborative deviations caused by data delays in each process. This enables real-time data perception across the entire process. Finally, the collected data is preprocessed in real time via edge computing nodes to ensure data consistency and avoid collaborative deviations caused by data delays in each process. The collaborative hub module, based on an industrial Ethernet-based data linkage model, utilizes process correlation analysis, real-time bottleneck location, and cross-unit data sharing to achieve collaborative data linkage. Process correlation analysis: It establishes a causal chain in real time between powder feeding, pressing pressure, holding time, demolding speed, and product quality. When the powder feeding fluctuates by ±5%, it automatically marks subsequent pressing units as high-risk processes. Real-time bottleneck location: It automatically traces upstream processes through time-series data comparison and triggers bottleneck warnings. Abnormal processes and their impact range are marked through a visual interface. Cross-unit data sharing: It breaks down the data barriers of a single press. When multiple presses produce the same series of products, the collaborative hub shares the optimal process parameters, avoiding repeated trial and error. The DRL model building module aims to maximize the overall pass rate and minimize energy consumption, constructing a DQN intelligent agent. Its core design is divided into: a state space (S) containing 12 core parameters from the real-time perception layer plus 30 minutes of historical process data features; an action space (A) defining the set of actions with adjustable parameters for each process and setting equipment safety constraints; and a reward function (R) employing a weighted multi-objective reward formula: R = α × pass rate improvement rate + β × energy consumption reduction rate - γ × parameter adjustment range, where α = 0.6, β = 0.3, and γ = 0.1, dynamically balancing quality, efficiency, and equipment stability. Simultaneously, it requires requirement analysis and screening of the state dimension, with requirement decomposition based on powder... The characteristics of the entire metallurgical process are defined, and the state space must cover four dimensions: real-time process parameters, equipment operating status, product quality feedback, and historical trend characteristics. This ensures that the correlation between parameters, quality, and energy consumption can be accurately reflected. Initial dimension screening: 18 raw data items from the sensing layer, such as powder feeding rate, pressing force, and mold temperature, are sorted out. Pearson correlation analysis (calculating the correlation coefficient between each parameter and the product qualification rate) is used to screen core dimensions. Parameters with an absolute value of correlation coefficient ≥ 0.3 are retained, such as the correlation coefficient between powder feeding uniformity and qualification rate of 0.62 and the correlation coefficient between pressing force stability of 0.58. Weakly correlated parameters, such as the humidity of the powder silo environment with a correlation coefficient of 0.08, are eliminated. 12 real-time core parameters are initially determined. The RL decision-making process module utilizes offline pre-training, online real-time decision-making, and dynamic exploration mechanisms to achieve collaborative control. Offline pre-training involves training the RL model based on historical production data, enabling the agent to initially grasp the parameter-quality mapping relationship. First, a multi-dimensional dataset is constructed, containing over 150,000 data sets from three presses over 12 months, including iron-based, copper-based, and cemented carbide products, and 15 different products. This dataset is then processed using an isolation forest to remove 3.2% outliers, SMOTE to supplement data, and to extract 12 derived features. Next, a DQN model is designed: an input layer of 15-dimensional state vectors → a convolutional layer with two kernels to extract temporal features → a fully connected layer with two hidden layers to prevent overfitting → an output layer of 15-dimensional action vectors. Training is conducted in three stages, with a fixed range of 10,000 to 500,000 steps. With a learning rate of 0.001, the decision accuracy is ≥75%. From 510,000 to 1,500,000 steps, the learning rate is reduced to 0.0001+PER, and the reward weights α=0.55 and β=0.35 are adjusted. When the pass rate is ≥95%, the energy consumption is reduced by 8%. From 1,510,000 to 2,000,000 steps, 50,000+ steps of parameter transfer samples are trained in 20 scenarios. Finally, offline testing (20,000 sets of data verifying a pass rate of ≥92% and energy consumption ≤1.1kWh / pcs) and simulation (success rate of ≥85% in handling 100 abnormal scenarios) confirm its usability. Online real-time decision-making: the collaborative center inputs the current state (S) into the RL agent every 100ms, and the agent outputs the optimal action (A) and simultaneously outputs the decision basis. The collaborative central system collects data every 50ms, generating a 15-dimensional state vector. Timestamp alignment and caching are used, resulting in a latency processing rate ≤0.5%. The system inputs the maximum action based on the Q value of the RL input at 100ms. After dual verification (pass rate ≥99.8%) for equipment safety (pressing force ≤90% of rated) and process boundaries (powder feeding ≥ minimum material value), the system issues data in steps according to Profinet, prioritizing powder feeding followed by pressing. The response time is ≤20ms, with adjustments made for deviations exceeding ±2%. A decision report is simultaneously output for visualization and manual intervention. A dynamic exploration mechanism is in place: when raw material batches change or product specifications switch, the RL agent initiates ε-greedy exploration, with ε gradually decreasing from 0.3 to 0.05. Under the premise of ensuring stable production, we will quickly explore the optimal parameter combination in new scenarios. Dynamic exploration will be initiated when the loose density of raw materials changes by ±0.1g / cm³, product type changes without parameters, or the pass rate drops by 5%. We will adopt dynamic ε-greedy (initial ε=0.3, mid-term ε=0.1, late-term ε=0.05) and stratified exploration (key parameter exploration range 1.2 times, minor parameter exploration range 1.1 times, and no exploration safety threshold). We will combine circuit breaking (pass rate <90% or energy consumption exceeding 1.5kWh / pcs) and small-batch trial production (50 pieces) to control risks. After repeated / cross-validation / compliance verification, effective parameters (pass rate ≥99.2%, etc.) will be stored in the optimal parameter library for reuse in related scenarios. The execution layer response module receives parameter instructions from the RL agent and then adjusts the parameters at the millisecond level through the servo control system to ensure action synchronization. The feedback optimization closed-loop module includes: real-time feedback of quality data from the online detection unit, parameter optimization linkage between the collaborative hub and the RL agent, and optimal parameter combination storage and reuse. The online detection unit feeds back adjusted product quality data to the collaborative hub in real time. The collaborative hub then transmits the adjustment actions and quality feedback data back to the RL agent, updates the reward value, and optimizes the model parameters. The real-time feedback of quality data from the online detection unit includes data acquisition dimensions and accuracy, data transmission and triggering mechanisms; the parameter optimization linkage between the collaborative hub and the RL agent includes data association and reward value calculation, and RL model parameter update logic; and the optimal parameter combination storage and reuse includes storage judgment criteria and processes, parameter reuse, and support for model transformation. The vibration sensing module deploys three high-precision accelerometers on the top of the punch and the side wall of the mold, with a range of ±500g and a sampling frequency of 1kHz. These sensors collect vibration acceleration signals in the vertical (pressing) and horizontal (swaying) directions of the punch in real time. After the data is processed by a 50Hz low-pass filter, it is uploaded to the collaborative hub at a frequency of 100μs / time. The vibration sensing module collects and analyzes the vibration data. After the pressing process starts, the accelerometers synchronously collect the punch vibration data. When the punch moves downward to press the powder, high-frequency vibrations of 20-200Hz are easily generated due to factors such as uneven powder particles and mold gaps. The amplitude is usually 3-8μm. After receiving the vibration acceleration signal, the collaborative hub converts it into vibration displacement data through integration calculation, such as acceleration 0.5g → displacement 4.9μm. The data is then compared with a preset threshold (2μm). If the threshold is exceeded, an active compensation process is triggered. The LMS adaptive control module, integrated into the collaborative central algorithm layer, uses vibration acceleration signal → vibration displacement calculation → compensation force output as its core logic. It adjusts compensation parameters in real time through the LMS adaptive filtering algorithm to achieve active vibration cancellation. At the same time, it embeds the RL decision layer for collaboration. It uses the vibration signal collected by the acceleration sensor as the error signal and the punch pressing speed and real-time pressing force (from the original pressure sensor) as reference signals, and inputs them into the LMS algorithm. The LMS algorithm updates the filtering coefficients in real time through the minimum mean square error criterion, with an iteration step size of 0.001. It calculates the magnitude and direction of the compensation force required to cancel the current vibration. For example, when the vertical vibration is 4.9μm, an upward compensation force of 3.2kN needs to be generated. The compensation command is synchronized to the RL decision layer to ensure that the compensation action and the pressing parameters (such as pressing force adjustment) are coordinated without conflict. The actuator drive module features four piezoelectric ceramic actuators evenly distributed circumferentially at the connection end between the punch and the press spindle. These actuators have a displacement resolution of 0.1 nm and a maximum output force of 5 kN. They are evenly distributed at 90° angles along the circumference of the punch-spindle connection end, corresponding to the punch's X, Y, and Z axes and the yaw direction. Two actuators (Z1 and Z2) are arranged vertically (in the pressing direction) to counteract the punch's vertical vibrations, such as the high-frequency vertical jumping caused by the powder reaction force during pressing. Two actuators (X and Y) are arranged horizontally to counteract the punch's left / right / forward / backward yaw vibrations, such as the punch offset vibration caused by uneven die clearance. Each actuator is flexibly connected to the punch and spindle, with an elastic coefficient of 5 N / μm, ensuring high actuator extension and contraction force. The effect is transmitted to the punch, while avoiding additional vibration caused by rigid connection. The actuator extension and retraction are controlled by a high-voltage drive power supply with an output voltage of 0-150V and a response time ≤10μs, generating a reverse compensation force. After receiving the compensation command from the LMS algorithm, the high-voltage drive power supply adjusts the output voltage within 10μs: for example, for a vertical vibration of 4.9μm, a 120V voltage is output to the corresponding piezoelectric ceramic actuator, driving it to extend by 0.8μm, generating a reverse compensation force. During the compensation process, the accelerometer provides real-time feedback on the vibration displacement change, and the LMS algorithm corrects the compensation parameters every 100μs: if the vibration displacement drops to 2.5μm, the voltage is adjusted to 90V, and the actuator extension and retraction is changed to 0.6μm, until the vibration amplitude stabilizes at ≤2μm, and the compensation is complete. Accuracy ±0.2μm; Vibration direction recognition and actuator selection: The coordinating hub determines the dominant vibration direction by the signal difference between three acceleration sensors (e.g., 0.5g from the top sensor and 0.3g from the sidewall sensor). If the vertical direction is the dominant vibration, actuators Z1 and Z2 are activated first. Precise voltage-displacement-force conversion: Based on the inverse piezoelectric effect of piezoelectric ceramics, voltage and displacement have a linear relationship. The actuator's voltage-displacement coefficient is 5.3nm / V. The LMS algorithm calculates the required compensation displacement and then reverses it to deduce the driving voltage. Each actuator has a built-in displacement sensor with a resolution of 0.01nm, which monitors the actual expansion and contraction in real time and compares it with the commanded displacement. If the deviation exceeds 5% (e.g., commanded 0.8μm, actual 0.76μm), immediate adjustment is made. The drive voltage is increased from 150V to 155V to ensure compensation accuracy. When the actuator output force exceeds the rated value by 5kN, if the voltage is mistakenly set to 160V, the high-voltage drive power supply will automatically cut off, and the central alarm will be triggered simultaneously to prevent actuator damage. The characteristics of piezoelectric ceramics are affected by temperature. For every 10°C increase in temperature, the displacement coefficient decreases by 0.5%. Temperature sensors are deployed around the actuator with a sampling frequency of 100Hz. When the temperature exceeds 40°C, the algorithm automatically corrects the voltage. For example, at 45°C, the voltage increases from 150V to 151.9V to maintain stable compensation displacement. Finally, a feedback optimization closed loop is embedded. After pressing, the online detection unit (laser profilometer) detects the dimensional accuracy of the product, such as gear pitch deviation. If the vibration compensation is in place, the pitch deviation will be reduced from ±0.The vibration data is reduced from 0.015mm to ±0.008mm, and the compensated mass data is sent back to the collaborative center. The collaborative center stores the vibration data, compensation parameters, and mass data in association as samples for LMS algorithm parameter optimization. For example, for the iron-based powder pressing scenario, the LMS iteration step size is optimized to 0.0008, and the state space of the RL model is updated to include the punch vibration amplitude in the state dimension, thereby improving the overall decision-making accuracy. By leveraging the nanometer-level displacement resolution of piezoelectric ceramic actuators and the real-time iteration of the LMS algorithm, vibration amplitude is reduced to ≤2μm, a reduction of over 75% compared to the uncompensated state (3-8μm). This significantly improves product dimensional accuracy and density uniformity. It seamlessly integrates with the existing control system, requiring no modification to the press's main structure. Only the sensor and actuator need to be added to the punch end. The total delay from vibration signal acquisition to compensation force output is ≤50μs, which can offset high-frequency sudden vibrations during the pressing process in real time, such as instantaneous impact vibrations caused by powder agglomeration.

[0026] like Figure 1 As shown, the powder feeding unit includes raw material confirmation, parameter matching, powder feeding start-up, and data acquisition functions to realize the conveying of powder raw materials. It also includes a material identification module for identifying, but not limited to, powder material and loose density attributes. Located in the raw material confirmation stage of the powder feeding unit, the module synchronously uploads the information to the collaborative hub. After parameter matching, the collaborative hub calls the optimal parameter library and matches the initial powder feeding parameters, including single powder feeding quantity and powder feeding rate, according to the current product specifications. This information is then sent to the powder feeding unit controller, initiating powder feeding. The powder feeding valve opens according to the controller's command, and the powder is conveyed to the mold cavity through the powder feeding pipeline. During this process, the laser flow sensor collects the powder feeding rate every 10ms to monitor whether there is any deviation. Within a preset range, if the rate fluctuation exceeds ±10%, an abnormal powder feeding rate signal is immediately generated. A weighing sensor is installed at the bottom of the powder hopper, recording the powder hopper weight change every 50ms. The actual powder feeding amount is calculated based on the weight difference, compared with the preset value, and synchronously uploaded to the coordination center. Finally, data acquisition is used for anomaly handling and parameter fine-tuning. If the coordination center receives data showing insufficient powder feeding, it immediately issues an adjustment command: increases the powder feeding valve opening by 5% and extends the powder feeding time by 0.2s. If the laser flow sensor detects uneven powder feeding, it triggers the powder feeding pipeline vibrator to clear possible powder blockage in the pipeline. At the same time, the abnormal information is synchronized to the pressing unit, prompting subsequent pressing processes to prepare for parameter adaptation. The pressing unit, encompassing functions such as powder feeding confirmation, dynamic adaptation of pressing parameters, pressing start, and connection to the pressure holding unit, is a system module used to press powder raw materials in the mold cavity into shape. When the pressing force reaches the target value and the upper mold stroke is in place, the pressing unit sends a compliance signal to the coordination central function module. The coordination central function module instructs the pressure holding unit to start the pressure holding process. The pressing unit controller adjusts the initial parameters based on the powder feeding data. When the powder feeding amount is 1.3% lower than the preset value, the pressing force is increased from the initial 120kN to 122kN, and the pressure holding time is extended from the initial 2s to 2.1s to ensure that the powder is fully compacted. After the pressing starts, the upper mold cylinder descends according to the adjusted parameters. During this process, the pressing... The force sensor collects the pressing force every 20ms and plots the pressing force and displacement curves in real time. If the pressing force suddenly increases at a certain moment, it immediately feeds back to the controller and stops the downward movement of the hydraulic cylinder. The displacement sensor is installed on the top of the upper mold and records the upper mold stroke every 10ms. When the stroke approaches the preset value, the hydraulic cylinder is controlled to decelerate to avoid overpressure. The temperature sensor is embedded in the mold side wall and monitors the mold temperature every 50ms. If the temperature rises to 85℃, the mold cooling system is started. Finally, the pressure holding unit is connected. When the pressing force reaches the target value and the upper mold stroke is in place, the pressing unit sends a pressing target signal to the coordination center. The coordination center synchronously sends instructions to the pressure holding unit to start the pressure holding process and uploads the pressing process data for archiving. Pressure Holding Unit: This system module performs pressure holding operations on the pressed product and transmits relevant data after pressure holding is completed. First, pressure holding is executed. The pressure holding valve maintains the current pressing force as instructed, and the timer starts timing. During this process, a pressure stability sensor collects pressure fluctuation values ​​every 10ms. If the fluctuation exceeds ±2%, it is immediately fed back to the controller. Pressure loss is compensated by adjusting the air pressure of the pressure holding valve. The timer compares the actual pressure holding time with the target value in real time. If time delay is caused by pressure compensation, the pressure holding time is automatically extended by 0.1s to ensure the pressure holding effect. When pressure holding ends, the pressure holding time meets the target, and the pressure fluctuation stabilizes within ±1%, the pressure holding unit sends a pressure holding completion signal to the coordination center and synchronously uploads the pressure holding data. The coordination center forwards the pressure holding data to the demolding unit, prompting the demolding process to prepare to start. Simultaneously, it triggers the RL decision layer: based on powder feeding, pressing, and pressure holding data, it predicts the possible resistance range during the demolding process, thus achieving data transmission. The demolding unit includes parameter reception and adjustment, demolding initiation, and product transfer. First, parameter reception and adjustment: the demolding unit receives demolding parameters from the collaborative hub and, combined with the demolding resistance predicted by the RL decision layer, fine-tunes the parameters. Demolding initiation: the demolding motor drives the lower mold to move upward. During this process, the displacement sensor collects the lower mold stroke every 10ms to monitor whether it moves upward synchronously and avoids skewing. If the stroke deviation on one side exceeds ±0.05mm, the driving force on both sides of the motor is adjusted to correct the skewing. The tension sensor is installed at the connection end between the lower mold and the motor and collects the demolding resistance every 20ms. If the resistance exceeds the predicted value by 15%, the demolding speed is immediately reduced to 0.5mm / s, and a demolding resistance abnormality signal is sent to the collaborative hub. Finally, product transfer: when the lower mold stroke reaches the target value and the product is completely removed from the mold cavity, the demolding unit sends a demolding completion signal and synchronously controls the conveyor belt to start, transferring the product to the online detection unit. The demolding unit uploads the final data to the collaborative hub, completing the data loop of the molding process. The online inspection unit includes positioning and scanning, sub-item inspection execution, result judgment and classification, data feedback and model optimization. First, the product enters the inspection area via a conveyor belt. A vision positioning camera captures the product's position, guiding a robotic arm to fix the product on the inspection platform. Simultaneously, a barcode scanner reads the product's unique identification code, linking it to data from previous processes. Sub-item inspection is then performed, including dimensional inspection, density inspection, and surface inspection, achieving comprehensive product inspection. Dimensional inspection: A laser profilometer scans the product, detecting key dimensions and generating a dimensional deviation report. Density inspection: An ultrasonic density meter emits ultrasonic waves that penetrate the product, calculating the internal density distribution based on the sound wave propagation speed. If the density of a certain area is below 92%, it is marked as a density defect. Surface inspection: A high-resolution vision camera captures the product surface from six different angles, covering the entire surface. The results are then analyzed using images... The algorithm identifies scratches and dents, generates a surface defect distribution map, and finally, the inspection system determines the product grade according to preset standards. The products are mainly categorized as qualified: dimensional deviation +0.012mm, density uniformity 96.2%, and no defects. A qualified signal is sent to the coordination center, which controls the conveyor belt to send the product to the qualified product warehouse. Unqualified products have 0.8mm scratches, are marked as surface scratches, and their defect images are uploaded to the coordination center. Simultaneously, the products are diverted to the unqualified product area for manual re-inspection. Afterwards, combining data feedback and model optimization, the coordination center sends the inspection results back to the RL decision layer. The RL agent updates the reward function based on data from the entire process of powder feeding, pressing, holding pressure, demolding, and inspection, optimizes the parameters for the next batch, and updates the parameter combinations of the qualified batches to the optimal parameter library, completing the closed-loop process.

[0027] like Figure 1As shown, the material identification module mentioned above mainly refers to the preprocessing of RFID tags for raw material packaging. This involves tag writing and tag attachment. Tag writing: After the powder raw materials are packaged at the manufacturer, the core attributes of the material are written into the RFID tag using a dedicated coding device. The data format follows a preset protocol. Tag attachment: The RFID tag with the written data is fixed in a prominent position on the outside of the raw material packaging. The tag surface is covered with a waterproof and wear-resistant film to prevent data loss due to wear or moisture during transportation. Afterwards, the raw material is identified upon entering the warehouse. Triggering conditions and reader activation enable the powder feeding unit to accept the material. Triggering conditions: When the manual or automatic feeding system delivers the raw material packaging to the powder hopper inlet, the infrared sensor installed at the inlet detects the approaching packaging and immediately sends an identification start signal to the RFID reader. Reader activation: After receiving the trigger signal, the RFID reader switches from standby to operation. In this state, the antenna is activated to emit ultra-high frequency electromagnetic waves, covering a range of 0.5-1m around the feed inlet, ensuring that the tag can be quickly captured after entering the identification area. RFID tag data reading and parsing is the core identification process, divided into tag wake-up and data reception, data parsing and format conversion, and data integrity verification. Tag wake-up and data reception: After the raw material packaging enters the identification area, the passive RFID tag obtains energy through the electromagnetic waves received by the antenna, activates the internal chip, and feeds back the stored data to the reader. The reader receives the tag signal through the antenna, and after filtering and amplification, transmits the raw data to the local data processing module. Data parsing and format conversion: The module's built-in parsing algorithm converts the binary data into a readable text format according to a preset protocol. Data integrity verification: After parsing, the system automatically checks whether the data fields are complete. If a key field is missing, a data incompleteness warning is immediately triggered, and the reader reads the data three times, with an interval of 0 seconds between each reading.If, after 5 seconds, three reads still result in missing fields, a tag malfunction is identified, and the process proceeds to the exception handling stage. Material attribute verification and matching constitute the compliance judgment step. This is achieved through two steps: comparing attributes with the rule base and determining the verification result. Attribute comparison with the rule base: The system calls the material attribute verification rule base, comparing the parsed material attributes with the corresponding standard attributes of the material in the rule base. Verification result determination: If all dimensions pass verification (compliant, within shelf life, and material matching), a qualified identification result is generated. If any dimension fails, an abnormal identification result is generated, and the abnormality type is recorded. Finally, data upload and subsequent linkage form a closed-loop connection, mainly divided into qualified data upload, exception handling and linkage, and exception resolution and retry. Qualified data upload: If the identification is deemed qualified, the data... The transmission module uploads complete material attributes to the collaborative hub via industrial Ethernet at a frequency of once per second until the collaborative hub returns a data reception confirmation signal. Upon receiving the data, the collaborative hub associates it with the current production task, providing raw material data support for subsequent parameter matching. Anomaly handling and linkage: If an identification anomaly is detected, the system performs two operations: local alarm and anomaly information upload. Anomaly resolution and retry: After the operator arrives, if the reading anomaly is caused by label misalignment, the raw material packaging position is adjusted, and the retry identification button is clicked. The system repeats steps 3-5. If the anomaly is due to the raw material itself, a qualified raw material is replaced, and the identification process is re-executed until a successful identification is achieved. At this point, the blocking cylinder retracts, allowing the raw material to enter the powder hopper, thus achieving material identification.

[0028] like Figure 1As shown, the label writing proposed in claim 3 mainly consists of basic information and traceability information. Basic information refers to the material code, material type, standard value of loose density, and shelf life. Traceability information refers to the manufacturer code and quality inspection report number. The core verification dimensions for comparing attributes with the rule base include loose density compliance, shelf life validity, and material matching with production requirements. Loose density compliance: The parsed loose density is 2.9 g / cm³, and the allowable range for loose density of iron-based powder in the rule base is 2.8-3.2 g / cm³, thus it is considered compliant. If the parsed value is 2.7 g / cm³, the density is considered abnormal. Shelf life validity: The system reads the current date, compares it with the 6-month shelf life and production date on the label, calculates the remaining shelf life, and determines whether it is within the acceptable range. If the current date has exceeded the shelf life, it is considered expired. Material matching with production needs: The collaborative hub issues the material type required for the current production task in advance. The system compares the material type field in the label. If they match, the match is successful. If it is copper-based powder, the material does not match. Anomaly handling and linkage mainly perform two operations: local alarm and anomaly information upload. Local alarm: The reader's built-in sound and light alarm is activated, with a flashing red light and a buzzer prompting at a frequency of 1 time / second. At the same time, the blocking cylinder at the powder hopper inlet extends to prevent abnormal raw materials from entering the powder hopper. Anomaly information upload: Information such as the anomaly type, original label data, and verification failure dimension is uploaded to the collaborative hub. The collaborative hub displays the anomaly in a pop-up window on the monitoring interface and pushes a reminder to the mobile phone of the on-site operator.

[0029] like Figure 1As shown, the process correlation analysis mentioned in the collaborative hub function is mainly divided into preliminary data modeling and rule definition, real-time data input and preprocessing, dynamic correlation calculation and risk labeling, correlation result output and downstream linkage. Preliminary data modeling and rule definition includes: basic correlation model construction: based on the powder metallurgy process mechanism, a process parameter correlation matrix is ​​built into the collaborative hub. The matrix's row dimension represents upstream process parameters, and the column dimension represents downstream process parameters / product quality indicators, with basic correlation weights labeled. Dynamic threshold setting: for each type of correlation, a fluctuation warning threshold is set. When the powder feeding volume fluctuates by ±5%, it corresponds to a high-risk powder feeding uniformity fluctuation in the pressing unit; when it fluctuates by ±8%, it corresponds to a product density deviation risk. The threshold can be dynamically updated based on historical production data. Once a month, real-time data input and preprocessing are performed, including data access: the collaborative hub receives real-time data uploaded by the sensing layer via industrial Ethernet. This data includes data from the powder feeding unit, pressing unit, pressure holding unit, and online detection unit. Data alignment and cleaning are performed: synchronization is achieved via timestamps, using the collaborative hub clock as a reference to correct time deviations in sensor data, ensuring an error ≤10ms. Discrete data from different processes are integrated into a time-series data chain. Simultaneously, the 3σ criterion is used to remove outliers to avoid interfering with correlation analysis. Dynamic correlation calculation and risk labeling are performed, including real-time correlation degree calculation: every 100ms, the collaborative hub calls a weighted correlation algorithm to calculate the risk correlation degree of downstream processes based on the current fluctuation values ​​of parameters in each process and the basic correlation weights. The formula is: Downstream process risk correlation = Σ upstream parameter fluctuation range × corresponding correlation weight For example, if the powder delivery volume fluctuates by +6% with a correlation weight of 0.7, and the powder delivery uniformity fluctuates by -3% with a correlation weight of 0.5, then the risk correlation of the pressing unit = 6% × 0.7 + 3% × 0.5 = 0.042 + 0.015 = 0.057. Risk level determination and marking: Set a risk correlation level standard. If the calculated result is 5.7%, the pressing unit will be automatically marked as medium risk, and the process will be marked in yellow on the collaborative central monitoring interface. Simultaneously, a risk correlation report will be generated, clearly indicating the risk source, correlation logic, correlation results output, and downstream linkage. The process is divided into parameter pre-adjustment suggestion generation: For downstream processes marked as medium / high risk, the collaborative hub automatically generates parameter adjustment suggestions based on the correlation model. For example, when the pressing unit is medium risk, it is recommended to increase the pressing force by 3kN, and it is marked that the powder feeding volume exceeds the threshold fluctuation by +6%, and the insufficient powder volume needs to be compensated by increasing the pressing force to avoid low density. Result push: The risk mark, correlation report and parameter adjustment suggestions are pushed to two terminals simultaneously: ① On-site operation terminal, touch screen display, response time ≤50ms, for operators' reference; ② RL decision-making level. Real-time bottleneck location mainly consists of defining and monitoring abnormal trigger conditions, tracing and investigating multi-dimensional data, and confirming and issuing early warnings for bottleneck results. The definition and monitoring of abnormal trigger conditions is further divided into abnormal type classification: a powder metallurgy production abnormality database is built into the collaborative hub, listing common abnormal scenarios, including: ① Quality abnormalities, referring to low product density, out-of-tolerance dimensions, and surface cracking; ② Process parameter abnormalities, referring to pressure meeting standards but large fluctuations in holding pressure, and sudden increases in demolding resistance; ③ Equipment abnormalities, each type of abnormality has a clear judgment standard, such as a product density <7.2g / cm³ being judged as a low density abnormality, while the standard value for iron-based products is 7.3-7.5g / cm³. Monitoring: The collaborative hub scans all process data every 200ms, comparing the current data with the standard threshold range. If the anomaly judgment criteria are met, the bottleneck location process is immediately triggered, and the anomaly trigger time and initial anomaly data are recorded. Multi-dimensional data tracing and investigation are conducted, including reverse tracing of quality anomalies: If the anomaly is low product density, the collaborative hub first retrieves historical data from the online detection unit to confirm whether the anomaly is a persistent problem. Then, it reverse-correlates the parameter data of the demolding-holding-pressing-powder feeding process, focusing on the following: Holding unit: whether there is a holding time < 2s and a holding pressure fluctuation > ±3%; Pressing unit: whether the pressing force does not reach the target value and the mold temperature is < 50℃; Powder feeding. Unit: Does the powder feeding amount <15g and powder feeding uniformity <92% exist? Cross-validation: If the powder feeding amount is found to be 14.2g for 5 consecutive times, further cross-validate the equipment status data of the powder feeding unit to confirm whether the insufficient powder feeding amount is caused by insufficient opening of the powder feeding valve, rather than a problem with the raw material itself. Algorithm-assisted localization: Call the random forest tracing algorithm and input the full process data of 10 production cycles before and after the anomaly into the algorithm. The algorithm outputs a bottleneck probability ranking based on historical anomaly cases, such as insufficient powder feeding amount > insufficient holding time > insufficient pressing force. Bottleneck result confirmation and warning output include a manual review interface reserved: For high-probability bottlenecks determined by the algorithm, the collaborative central system automatically... The parameter curves for the corresponding time period are captured and displayed on the operating terminal for operators to review. Early warning information is generated and pushed out: After confirming the bottleneck, standardized early warning information is generated, including the bottleneck process (also referring to the powder feeding unit), abnormal parameters (e.g., powder feeding amount 14.2g, lower than the standard 15g), the scope of impact (the product density may remain low for the next 5 production cycles), and recommended measures (adjusting the powder feeding valve opening to 85%). Early warning information is pushed out in three ways: ① A pop-up window on the monitoring interface flashes red until operator confirmation; ② The workshop's audible and visual alarm activates the alarm in the corresponding powder feeding unit area, beeping twice per second; ③ Push notifications to management personnel's mobile app, including screenshots of the anomaly and handling suggestions. Cross-unit data sharing is mainly divided into the scope and standard definition of shared data, the sharing trigger mechanism and data matching, and the execution and feedback optimization of data sharing. The scope and standard definition of shared data includes shared data classification: a cross-unit shared database is established in the collaborative hub to store shareable data in categories, including: ① basic parameters, product specifications, material type; ② optimal process parameters, powder feeding rate, pressing force, holding time, etc., which must meet a pass rate of ≥99.2% for three consecutive production cycles; ③ anomaly handling plans; ④ equipment status data. Data standardization format: the format and units of shared data are unified, for example, powder feeding rate is uniformly expressed in grams (g), pressing force in kN, and pass rate in percentage (%), to avoid parameter misuse due to inconsistent formats. The shared triggering mechanism and data matching are divided into trigger scenario definition: setting trigger conditions for cross-unit data sharing, including: ① new press going online; ② product model change; ③ local parameter optimization. The intelligent data matching algorithm: when a sharing requirement is triggered, the collaborative hub calls the parameter matching algorithm to filter data in the shared database based on current production needs. The matching logic is: prioritize matching parameters with the same equipment model + completely matching product specifications + the same material; if no completely matching data is found, match parameters with similar equipment models + similar product specifications + the same material; the matching result must meet the parameter applicability ≥ 85%, where applicability = equipment matching degree × 0.4 + product matching degree × 0.4 + material matching degree × 0.2, and equipment / product / material completely matching. The result is 100%; otherwise, a message will be displayed indicating no available shared parameters, requiring the initiation of RL exploration mode. Data sharing execution and feedback optimization are divided into: Parameter distribution and synchronization: The collaborative hub distributes the matched shared parameters to the target press via industrial Ethernet. A step-by-step verification mechanism is used during the distribution process: First, the parameters are distributed to the target press's parameter buffer. The target press verifies whether the parameters meet its own equipment constraints. If they do, it returns a successful reception. The collaborative hub then triggers a parameter activation command to make the parameters officially effective. The entire process takes ≤100ms. Sharing effect feedback and iteration: After the target press uses the shared parameters for production, the collaborative hub collects its product quality data, pass rate, and dimensional deviation every production cycle (approximately 30 seconds), along with the source of the shared parameters. Comparison of press quality data: If the pass rate of the target press is ≥ 98% of the pass rate of the shared source press, the sharing is deemed valid, and the parameter is added to the local optimal parameter library of the target press. If the pass rate of the target press is < 95% of the pass rate of the shared source press, the shared parameter needs to be adapted. The collaborative hub automatically calls the RL decision layer to fine-tune the parameters based on the real-time data of the target press, such as the mold temperature being 5℃ higher than that of the source press. The adapted parameters are then updated to the cross-unit shared database for subsequent sharing by similar equipment. Access control and data security: Three levels of data sharing permissions are set: administrators can modify sharing rules, technicians can review shared parameters, and operators can only view them. At the same time, the shared data is encrypted during transmission to prevent data from being tampered with or leaked.

[0030] like Figure 1 As shown, the construction of the DRL model also includes: state space (S), action space (A), and reward function (R). The state space (S) in the construction of the DRL model includes the following steps: real-time parameter refinement, data standardization, historical data window selection, historical feature extraction, and dimension fusion; real-time parameter refinement: for the 12 core parameters after selection, the dimension definition, unit, and value range are clearly defined to ensure the data format is consistent, as follows: Dimension Category Parameter name unit Range of values Data source Powder delivery unit Powder delivery rate g / s 6-10 Laser flow sensor Powder delivery unit Powder delivery uniformity % 85-100 Laser flow sensor (fluctuation calculation) Suppression unit Real-time suppression force kN 80-150 pressure sensor Suppression unit mold temperature ℃ 40-90 Temperature sensor Pressure holding unit Pressure holding time deviation s -0.5-+0.5 Timer (difference from target value) Pressure holding unit Pressure stability % 95-100 Pressure stability sensor Demolding unit Demolding speed mm / s 0.5-1.2 Displacement sensor (rate calculation) Demolding unit Demolding resistance kN 5-12 Tension sensor Online detection unit Product dimensional tolerances mm -0.03-+0.03 Laser profilometer Online detection unit Product density g / cm³ 7.0-7.6 Ultrasonic density meter Energy consumption monitoring Unit product energy consumption kWh / pcs 0.8-1.5 Electricity metering module Equipment status Press load rate % 60-100 Motor current sensor (conversion) Data standardization: Min-Max standardization is used to map each parameter to the [0,1] interval. The formula is: Standardized value = (Actual value - Minimum value) / (Maximum value - Minimum value), avoiding model training bias due to differences in parameter units; Historical data window selection: Based on the characteristics of the production cycle, a single cycle is about 30 seconds, and 30 minutes (60 production cycles) are selected as the historical data window to ensure that parameter trend changes can be captured; Historical feature extraction: Three types of historical features are calculated using the sliding window method: Trend features: such as the linear fitting slope of the pass rate within 30 minutes, a positive slope indicates an increase in the pass rate, and a negative slope indicates a decrease; Fluctuation features: such as the standard deviation of the pressing force within 30 minutes, reflecting the stability of the parameter, the smaller the standard deviation, the better the stability; Extreme value features: such as the maximum value of the demolding resistance within 30 minutes, to determine whether there is a potential risk of equipment overload; Dimension fusion: The 12-dimensional real-time parameters are concatenated with the 3-dimensional historical features to form a 15-dimensional state vector (S∈R). 15 ), which serves as the input state for the RL agent; Motion Space (A): Basic adjustment range setting and dynamic constraint optimization; Basic adjustment range setting: Refer to process standards and equipment rated parameters to set the basic adjustment range for each motion dimension, as follows: Action dimension Adjust parameters Basic adjustment range Equipment safety constraints Process precision constraints Powder delivery unit Powder delivery volume ±2% Each adjustment shall not exceed 5% of the powder hopper capacity. Avoid density deviation exceeding ±0.05 g / cm³ Suppression unit Suppressive force ±3kN Do not exceed 90% (135kN) of the press's rated load. Avoid overpressure to prevent mold damage. Pressure holding unit Holding time ±0.5s Not less than 1 second of minimum pressure holding time Avoid looseness inside the product Demolding unit Demolding speed ±0.1mm / s Not less than the minimum demolding speed of 0.5 mm / s To prevent product cracking Dynamic constraint optimization: The constraint range is dynamically adjusted based on the equipment operating status. For example, when the press load rate is ≥95%, the upper limit of the pressing force adjustment is automatically reduced to ±2kN (from ±3kN) to avoid equipment overload. When the loose density of powder decreases by 0.2g / cm³, the powder feeding adjustment range is expanded to ±3% (from ±2%) to ensure sufficient powder quantity. Reward Function (R): Weight determination, pass rate improvement rate reward (R1), energy consumption reduction rate reward (R2), parameter adjustment range penalty (R3); Weight determination: Five process experts were invited to score the three objectives using the Analytic Hierarchy Process (AHP) to construct a judgment matrix and calculate the weights: Pass rate improvement rate (α): Experts unanimously agreed that it had the highest priority, with a weight calculation result of 0.6; Energy consumption reduction rate (β): Second highest, with a weight calculation result of 0.3; Parameter adjustment range (γ): As a constraint term, the weight calculation result is 0.1, a negative weight, the larger the adjustment range, the less the reward; Finally, the basic formula of the reward function was determined: R = α×R1 + β×R2 - γ×R3; R1 is the reward for the qualified rate improvement rate, R2 is the reward for the energy consumption reduction rate, and R3 is the penalty for the parameter adjustment range; Reward for the qualified rate improvement rate (R1): Data collection: Obtain the current batch qualified rate (Q_current) from the online detection unit every 1 production cycle (30 seconds) and compare it with the previous batch qualified rate (Q_last); Calculation logic: Adopt a segmented reward mechanism to prevent the model from overly pursuing the qualified rate and ignoring other goals: If Q_current > Q_last + 2%: R1 = 1.0, full score reward, encouraging a significant increase; If Q_last - 1% ≤ Q_current ≤ Q_last + 2%: R1 = 0.5, basic reward, maintaining stability; If Q_current < Q_last - 1%: R1 = -0.8, penalty, to avoid a decrease in the qualified rate; Special case: If Q_current ≥ 99.5%, the top level in the industry, an additional 0.2 reward points will be added to guide the model to strive for a high qualified rate; Reward for the energy consumption reduction rate (R2): Data collection: Obtain the current energy consumption per unit product (E_current) from the electric energy metering module and compare it with the historical optimal energy consumption, E_optimal, based on the data statistics of the past 3 months; Calculation logic: If E_current < E_optimal - 5%: R2 = 0.8, high - amount reward, encouraging significant energy conservation; If E_optimal - 5% ≤ E_current ≤ E_optimal: R2 = 0.4, basic reward, maintaining low energy consumption; If E_current > E_optimal + 5%: R2 = -0.5, penalty, to avoid excessive energy consumption; Constraint condition: If the reduction in energy consumption causes the qualified rate to decrease by more than 1%, then R2 is forced to be set to 0 to avoid sacrificing quality for energy consumption; Penalty for the parameter adjustment range (R3): Data collection: Calculate the sum of the absolute values of the parameter adjustment ranges between the current action and the previous action; Standardization processing: Standardize each parameter adjustment range according to the proportion of the adjustment range, for example, the pressing force adjustment of 3 kN accounts for 100% of its ±3 kN adjustment range, and the standardized value is 1.0; The powder feeding amount adjustment of 2% accounts for 100% of its ±2% adjustment range, and the standardized value is 1.0; Calculation logic: R3 = the standardized total adjustment range, with a value range of 0 - 4.0. Because there are 4 action dimensions, the maximum value of each dimension after standardization is 1.0. For example, if the total adjustment range after standardization is 0.8, then R3 = 0.8.

[0031] Such as Figure 1As shown, offline pre-training in the RL decision-making process refers to building the decision foundation, achieving the model's initial decision-making capability through four steps: multi-dimensional dataset construction, DQN model architecture, phased training, and performance verification. Multi-dimensional dataset construction involves collecting 150,000+ sets of data from 3 different models of presses over 12 months, including iron-based, copper-based, and cemented carbide products, representing 15 different products. This data is then processed using Isolation Forest to remove outliers (removing 3.2%), SMOTE to supplement data, and 12 derived features extracted for later use. The DQN model architecture is a four-layer structure: Input layer: 15-dimensional state vector → Convolutional layer: 2 convolutional kernels to extract temporal features → Fully connected layer: 2 hidden layers to prevent overfitting → Output layer: 15-dimensional action. Vector; Phased training: Basic learning, 10,000-500,000 steps, fixed learning rate 0.001, mastering basic patterns such as fan delivery volume and density, decision accuracy ≥75%; Multi-objective optimization, 510,000-1,500,000 steps, dynamically reducing the learning rate to 0.0001, adjusting reward weights (α=0.55, β=0.35), achieving an 8% energy consumption reduction when the pass rate is ≥95%; Scenario adaptation, 1,510,000-2,000,000 steps, trained in 20 scenarios, adding parameters to transfer samples to improve adaptability; Effect verification: offline testing with 20,000 sets of data, pass rate ≥92%, energy consumption ≤1.1kWh / pcs, simulating 100 abnormal scenarios, with a success rate of ≥85%; Online real-time decision-making refers to dynamic and precise control, achieving rapid parameter adjustment in four steps: status acquisition synchronization, intelligent decision-making and verification, command issuance monitoring, and decision visualization. Status acquisition synchronization: The collaborative hub collects data every 50ms, preprocesses it to generate a 15-dimensional vector, aligns it with timestamps and uses caching to ensure timing consistency, with a latency processing rate ≤0.5%. Intelligent decision-making and verification: 100ms input to the RL agent selects the action with the highest Q value, followed by dual verification of equipment safety (pressing force ≤90% of rated) and process boundaries (powder feeding ≥ minimum material value) (pass rate ≥99.8%). Command issuance monitoring: Profinet protocol is used for step-by-step issuance, from powder feeding to pressing, with real-time comparison of execution values; if the deviation exceeds ±2%, a supplementary command is issued. Decision visualization: Output reports include influencing factors, expected effects, and historical cases, displayed on large screens / terminals, supporting manual intervention. The dynamic exploration mechanism refers to a four-step balanced exploration and stabilization approach to cope with scenario changes: triggering conditions, ε-greedy strategy, risk control, and result solidification. Triggering conditions: activation occurs when the loose density of raw materials changes by ±0.1g / cm³, the product is changed without historical parameters, or the pass rate drops by 5% consecutively. ε-greedy strategy: initial stage (cycles 1-10): ε=0.3 (30% exploration), mid-stage (cycles 11-30): ε drops to 0.1, late stage (after cycle 31): ε=0.05, exploration stops when the optimal parameter is found (pass rate ≥99.2%). Layered exploration: the exploration range for key parameters (powder feeding / pressing force) is 1.2 times the conventional range, and for secondary parameters (demolding speed) it is 1.1 times the conventional range, with a safety threshold for prohibiting exploration. Risk control: when the pass rate is <90% or energy consumption exceeds 1.5kWh / pcs, a circuit breaker is triggered and the process is reversed. For major scenarios, a small-batch trial production of 50 pieces is conducted first. Result solidification: after the effective parameters are verified by repetition (3 batches) + cross-validation (same model press) + compliance, they are stored in the optimal parameter library.

[0032] like Figure 1 As shown, the execution layer response module is mainly divided into the following two steps: unit execution details and synchronization guarantee measures. The unit execution details are mainly completed using the powder feeding unit, pressing unit, and demolding unit. Powder feeding unit: The powder feeding valve (electric proportional valve) adjusts its opening according to instructions. For example, a +1.5% increase in powder feeding corresponds to an increase in valve opening from 60% to 62%, with a response time ≤20ms, ensuring real-time powder quantity adaptation. Pressing unit: The pressing cylinder is driven by a servo hydraulic system. A +2kN pressing force instruction triggers a fine-tuning of the cylinder pressure valve, with a pressure control accuracy of ±0.5kN, avoiding overpressure or underpressure. Demolding unit: The demolding motor (servo motor) receives a demolding speed instruction of ±0.1mm / s and uses an encoder to provide real-time feedback of the rotation speed, adjusting the response time to ≤15ms to prevent product cracking. Synchronization assurance measures: Prioritization of instructions: Instructions from the powder feeding unit are executed first and issued 50ms in advance to ensure that the pressing process starts only after the powder quantity is in place, thus avoiding process disconnection; Real-time status feedback: Each execution unit transmits actual parameters to the coordination center every 10ms, such as the current powder feeding valve opening and cylinder pressure. If the deviation exceeds ±2%, it is immediately adjusted to ensure synchronized action.

[0033] like Figure 1As shown, the real-time feedback of quality data from the online inspection unit in the feedback optimization closed-loop module mainly refers to the simultaneous collection of three types of core quality data after the online inspection unit completes product inspection: dimensional accuracy data, internal quality data, and surface defect data. Through data transmission and triggering mechanisms, a rapid response from the collaborative hub is ensured. Data collection dimensions and accuracy: After completing product inspection, the online inspection unit needs to simultaneously collect three types of core quality data: Dimensional accuracy data: The laser profilometer outputs the actual values ​​and tolerance deviations (accuracy ±0.005mm) of key product dimensions (such as gear inner diameter and pitch). For example, an inner diameter of 20.01mm has a tolerance of +0.01mm and a standard of ±0.02mm. Internal quality data: The ultrasonic density meter generates a density distribution heatmap of the product, marking the areas with the lowest density, such as a density of 7 in the central area. 1g / cm³, lower than the standard of 7.3g / cm³, and overall density uniformity, such as 94%, while the standard is ≥95%; Surface defect data: The vision camera uses image recognition algorithms to count the type (scratches, dents, flash), quantity, and size of defects, such as a 0.6mm long scratch, and generates the coordinates of the defect location; Data transmission and triggering mechanism: Within 100ms after the inspection is completed, the inspection unit uploads the quality data to the collaborative hub via industrial Ethernet (Profinet protocol). The data packet includes the product's unique identifier (SN code), inspection timestamp, quality parameters, and pass / fail judgment result; If it is judged as unqualified, such as out-of-tolerance dimensions or insufficient density, the system automatically triggers the abnormal feedback priority, and the data transmission delay is compressed to within 50ms to ensure rapid response of the collaborative hub; The parameter optimization linkage between the collaborative hub and the RL agent involves the collaborative hub receiving quality data, forming action-quality data pairs, calculating real-time reward values ​​according to the reward function formula, and finally transmitting the action-quality-reward data pairs back to the RL agent. The agent stores these data pairs in an experience replay pool with a capacity of 100,000 records. Data association and reward value calculation: After receiving quality data, the collaborative hub associates it with previously issued adjustment actions by the RL agent using the product SN code and timestamp, such as increasing powder delivery volume by 1.5% or increasing pressing force by 2kN, forming action-quality data pairs. The real-time reward value is calculated using the reward function formula R = α × pass rate improvement rate + β × energy consumption reduction rate - γ × parameter adjustment amplitude. For example, after adjustment, the pass rate increases from 95% to 97%, an improvement rate of 2%, and energy consumption decreases from 1.2kWh to 1.15kWh. Wh, with a reduction rate of 4.2% and a standardized parameter adjustment range of 0.6, the reward value R = 0.6 × 2% + 0.3 × 4.2% - 0.1 × 0.6 = 0.012 + 0.0126 - 0.06 = -0.0354, a negative reward, indicating that the parameters need further optimization; RL model parameter update logic: The collaborative hub transmits action-quality-reward value data pairs back to the RL agent, which stores them in the experience replay pool with a capacity of 100,000 records. Every 50 sets of data are accumulated, the agent initiates a model parameter update, adjusting the weight parameters of the DQN network through the gradient descent algorithm with a learning rate of 0.0001, optimizing the state-action mapping relationship. For example, for scenarios where the density is still low after increasing the amount of powder delivered, the weights of the network output layer are adjusted so that actions with synchronously increased suppression force are prioritized in subsequent similar states; The optimal parameter combination is determined for inclusion and reuse after meeting certain conditions for three consecutive production cycles. The collaborative hub then automatically generates an optimal parameter report. The inclusion criteria and process are as follows: An optimal parameter combination is considered complete when the following conditions are met for three consecutive production cycles, with 50-100 products produced per cycle: ① Pass rate ≥ 99.2%; ② Dimensional tolerance compliance rate 100%; ③ Density uniformity ≥ 96%; ④ Unit product energy consumption ≤ 1.1 kWh. The collaborative hub automatically generates an optimal parameter report, including parameter details (powder feed rate, pressing force, etc.) and applicable scenarios (material type, product specifications, raw material batch). Quality data statistics are completed within 24 hours after online review by process personnel and officially stored in the optimal parameter library. Parameter reuse and changeover support: When production scenarios match, such as the same material and product specifications, the collaborative hub automatically calls the parameters in the optimal parameter library as the initial process parameters before changeover, without the need for re-adjustment. If the scenarios are similar, such as similar product specifications and the same material, the system fine-tunes the optimal parameters based on the parameter mapping algorithm, such as converting the powder feeding amount according to the size ratio. For example, the original product with an inner diameter of 20mm corresponds to a powder feeding amount of 15g, while the new product with an inner diameter of 19mm is converted to 14.25g, further shortening the changeover cycle.

[0034] like Figure 1 As shown, the central processing unit (CPU) and the power module are electrically connected; the power module supplies power to the CPU, enabling its normal operation. The power module provides stable power support for the entire feedback optimization closed loop, ensuring data acquisition, transmission, and decision execution. Key unit power supply guarantees: It provides compatible voltages for the online detection unit (laser profilometer and ultrasonic density meter require 220V±5% regulated power, vision camera requires 12V DC power), the collaborative hub (industrial Ethernet switch requires 48V PoE power), and the RL intelligent agent computing unit (server requires 380V three-phase power), ensuring detection accuracy (e.g., laser profilometer). When the power supply fluctuation is ≤±2%, the dimensional detection error is ≤0.005mm, and the operation stability is maintained. Power supply stability measures: UPS uninterruptible power supply is used to prevent sudden power outages from causing loss of quality data. The power outage recovery time is ≥30 minutes. Voltage filtering modules are configured to suppress fluctuations in the workshop power grid, such as voltage spikes generated by machine tool start-up and shutdown. The power supply ripple coefficient is ≤1%. Abnormal power supply response: When the power supply voltage is detected to exceed the threshold, such as below 198V or above 242V, an alarm is immediately triggered and the backup power supply is switched. At the same time, the RL intelligent agent parameter update and optimal parameter entry operation are suspended to avoid abnormal power supply affecting data accuracy.

[0035] 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 flexible powder metallurgical press production flow optimization control system, characterized by, Comprise: A central processing unit, which establishes a cooperative relationship of command and execution between the system bus and the modules, is responsible for allocating tasks, passing parameters and receiving feedback for various modules, and realizes orderly and efficient operation of the whole system; A power module, which supplies power for the central processing unit; A data sensing module, which includes a powder feeding unit, a pressing unit, a pressure maintaining unit, a demolding unit, and an online detection unit throughout the whole process, and deploys multiple types of sensors to realize real-time data sensing throughout the whole process, and through edge computing nodes, the collected data is preprocessed in real time to ensure data time sequence consistency; A collaborative hub function module, which builds a data linkage model based on industrial Ethernet, and realizes data collaborative linkage through process correlation analysis, real-time bottleneck positioning, and cross-unit data sharing; A DRL model construction module, which constructs a DQN agent with the goal of maximizing the overall process qualification rate and minimizing energy consumption; An RL decision-making process module, which realizes collaborative control through offline pre-training, online real-time decision-making, and dynamic exploration mechanism; An execution layer response module, which realizes millisecond-level parameter adjustment through servo control system after receiving parameter instructions from RL agent to ensure action synchronization; A feedback optimization closed loop module, which includes real-time feedback of online detection unit quality data, parameter optimization linkage between collaborative hub and RL agent, optimal parameter combination storage and reuse, and real-time feedback of adjusted product quality data to the collaborative hub through the online detection unit, and the collaborative hub returns the adjustment action and quality feedback data to the RL agent to update the reward value and optimize the model parameters; A vibration sensing module, which deploys multiple acceleration sensors on the top of the punch and the side wall of the die to real-time collect vibration acceleration signal data in the vertical and horizontal directions of the punch, and the data is uploaded to the collaborative hub after low-pass filter processing; An LMS adaptive control module, which is integrated into the algorithm layer of the collaborative hub, and adjusts the compensation parameters in real time through the LMS adaptive filter algorithm; An execution driving module, which uniformly arranges multiple piezoelectric ceramic actuators around the connection end of the punch and the main shaft of the press, and controls the extension and contraction of the actuators through a high-voltage driving power supply to generate a reverse compensation force.

2. A flexible powder metallurgical press production process optimization control system according to claim 1, characterized in that: The powder feeding unit includes raw material confirmation, parameter matching, powder feeding start, and data acquisition functions, and is used to realize the transportation of powder raw materials; it also includes a material identification module for identifying attributes such as powder material and bulk density, which is located in the raw material confirmation link of the powder feeding unit; The pressing unit covers powder feeding confirmation, dynamic adaptation of pressing parameters, pressing start, and connection function of the pressure maintaining unit, and is used for the system module of pressing the powder raw materials in the die cavity; when the pressing force reaches the target value and the upper die stroke is in place, the pressing unit sends a signal to the collaborative hub function module, and the collaborative hub function module instructs the pressure maintaining unit to start the pressure maintaining process; The pressure maintaining unit performs pressure maintaining operation on the pressed product and transmits related data after the pressure maintaining is completed; the pressure maintaining unit sends the pressure maintaining data to the collaborative hub function module, which is forwarded to the demolding unit and triggers the RL decision layer module to predict the demolding resistance range through the collaborative hub function module; The demolding unit is used for completing product demolding, transferring to the next process, and uploading data. The demolding unit receives demolding parameters issued by the collaborative hub, and combines the process of fine-tuning demolding resistance parameters predicted by the RL decision layer. The full-process online detection unit is a system module for detecting, classifying, and feeding back data of the molded product. The RL agent updates the reward function based on full-process data, optimizes the parameters of the next batch, and updates the qualified batch parameter combination to the optimal parameter library as an intelligent decision-making subject.

3. A flexible powder metallurgical press production process optimization control system according to claim 2, characterized in that: The raw material packaging RFID tag preprocessing, warehouse identification triggering, core identification, compliance judgment, and closed-loop connection are implemented as follows: The raw material packaging RFID tag preprocessing completes the preparation work before material identification through tag writing operation and tag attachment operation; The warehouse identification triggering activates the reader through the preset triggering condition, so that the powder feeding unit responds and enters the material receiving state; The core identification takes RFID tag data reading and analysis as the core, and sequentially performs the steps of tag awakening and data receiving, data analysis and format conversion, and data integrity verification; The compliance judgment compares the read material attributes with the preset rule library to complete the verification result determination, realizes material attribute verification and matching; The closed-loop connection performs qualified data upload operation according to the compliance judgment result, performs abnormal handling and linkage operation for abnormal conditions, performs retry operation after the abnormality is removed, and finally realizes accurate identification of the material.

4. A flexible powder metallurgical press production process optimization control system according to claim 3, characterized in that: The information written in the tag includes basic information and traceability information; the core verification dimension of the attribute and rule library comparison is the compliance of bulk density, the validity of shelf life, and the matching degree of material and production demand; the abnormal handling and linkage operation specifically includes local alarm and abnormal information upload double operation.

5. A flexible powder metallurgical press production process optimization control system according to claim 4, characterized in that: The collaborative hub function module includes three sub-modules of process correlation analysis, real-time bottleneck positioning, and cross-unit data sharing, and the implementation process of each sub-module is as follows: The process correlation analysis submodule sequentially performs the steps of early data modeling and rule definition, real-time data input and preprocessing, dynamic correlation calculation and risk marking, correlation result output and downstream linkage; The bottleneck real-time positioning submodule realizes bottleneck positioning through the steps of abnormal trigger condition definition and monitoring, multi-dimensional data tracing and investigation, bottleneck result confirmation and early warning output; The cross-unit data sharing submodule completes data sharing in the order of shared data range and standard definition, shared trigger mechanism and data matching, data sharing execution and feedback optimization.

6. The flexible powder metallurgical press production process optimization control system of claim 1, wherein: The DRL model construction module further includes: a state space , an action space , and a reward function ; The reward function The final reward function base formula is determined as: ； wherein, is a pass rate improvement rate reward, is an energy consumption reduction rate reward, is a parameter adjustment amplitude penalty, is a parameter adjustment amplitude, is an energy consumption reduction rate, is a pass rate improvement rate.

7. The flexible powder metallurgical press production process optimization control system of claim 1, wherein: It also includes an offline pre-training unit, an online real-time decision unit and a dynamic exploration mechanism unit; the offline pre-training unit is used to build a decision basis to realize the preliminary decision-making ability of the model, and its execution process includes multidimensional data set construction, DQN model architecture construction, phased training and effect verification; the online real-time decision unit is used for dynamic and accurate regulation to realize rapid parameter adjustment, and its execution process includes state acquisition synchronization, intelligent decision and verification, instruction issuing monitoring and decision visualization; the dynamic exploration mechanism unit is used to balance exploration and stability in response to changes in the scene, and its execution process includes trigger condition judgment, ε-greedy strategy execution, risk control and result solidification.

8. A flexible powder metallurgical press production process optimization control system according to claim 7, characterized in that: The execution layer response module mainly includes the following two steps: unit execution details and synchronization guarantee measures. At the same time, the unit execution details mainly use the powder feeding unit, the pressing unit and the demolding unit to complete the execution work.

9. A flexible powder metallurgical press production process optimization control system according to claim 8, characterized in that: The online detection unit is used to synchronously collect three types of core quality data, i.e., size precision data, internal quality data and surface defect data after product detection, and to guarantee the fast response of the coordination center through data transmission and triggering mechanism; the coordination center forms action-quality corresponding data pairs after receiving the core quality data, and calculates real-time reward values according to a reward function and returns the action-quality-reward value data pairs to the RL agent for storage in an experience replay pool. The parameter management unit is used to determine the corresponding parameters as the optimal parameter combination when the preset condition is met for three consecutive generation cycles, and to automatically generate an optimal parameter report by the collaborative hub.

10. A flexible powder metallurgical press production process optimization control system according to claim 9, characterized in that: The central processing unit and the power module are electrically connected; It can power the central processing unit to enable it to operate normally.

Images