Cold stamping quality control method based on multi-source data fusion and real-time optimization

Abstract

The invention relates to the technical field of aluminum alloy cold stamping quality control, in particular to a multi-source data fusion and real-time optimization-based cold stamping quality control method, which comprises the following steps of: inputting a fusion feature space into a pre-trained LSTM (Long Short Term Memory) model, and associating a real-time stamping speed with a dwell time parameter through an attention mechanism; combined prediction of springback value deviation and defect probability is realized for the first time, a process parameter correction value is dynamically generated based on a prediction result, and a thermal expansion compensation algorithm and micron-sized mold compensation are combined, so that the quality regulation response speed reaches a millisecond level, the compensation precision is improved to + / -2 [mu] m, and the error is reduced by more than 50% compared with that of traditional PID (Proportion Integration Differentiation) control.

Description

Technical Field

[0001] The present invention relates to the technical field of aluminum alloy cold stamping quality control, and particularly relates to a cold stamping quality control method based on multi-source data fusion and real-time optimization. Background Art

[0002] At present, the aluminum alloy cold stamping forming technology is one of the core processes for lightweight manufacturing of new energy vehicles. Through high-precision and high-efficiency stamping processes, aluminum alloy sheets are processed into body structural parts (such as car doors, hoods, battery pack housings, etc.), significantly reducing the vehicle's overall mass and improving the endurance and energy utilization efficiency.

[0003] The following technical problems exist in the current prior art: 1. The existing technology mainly relies on a single type of sensor (such as a pressure or displacement sensor), with limited data acquisition dimensions, resulting in difficulty in synchronously obtaining multi-modal data such as die pressure, displacement, temperature, and surface defects, affecting the comprehensive monitoring of the stamping process; moreover, the traditional feature fusion method uses fixed weight allocation (such as equal-weight averaging), lacks a dynamic weight adjustment mechanism based on the material's elastic modulus, and does not effectively perform dimensionality reduction processing, resulting in insufficient extraction efficiency of key information such as stamping energy characteristics and deformation models, and a high error rate.

[0004] 2. The prediction of the springback amount of workpieces (referring to the dimensional deviation value caused by the elastic recovery of the material after the aluminum alloy sheet is formed in the traditional cold stamping process, that is, the maximum offset between the actual shape of the stamped part after demolding and the theoretical shape designed by the die) in the existing technology is based on empirical formulas or finite element simulations, and cannot dynamically correlate real-time stamping speed and holding pressure time parameters, with low prediction accuracy and a defect missed detection rate > 20%.

[0005] 3. The existing MES system has not effectively integrated material batches, process parameters, and defect patterns. Quality problem tracing relies on manual cross-system data retrieval, lacks multi-dimensional correlation analysis capabilities, and the existing MES system lacks the ability to correlate multi-source data. Process parameters, defect patterns, and material batches are stored separately, and quality problem tracing relies on manual investigation, which is time-consuming.

[0006] Therefore, a cold stamping quality control method based on multi-source data fusion and real-time optimization that can solve the above problems is needed. Summary of the Invention

[0007] The present invention provides a cold stamping quality control method based on multi-source data fusion and real-time optimization. The present invention inputs the fused feature space into a pre-trained LSTM model, associates the real-time stamping speed and holding time parameters through the attention mechanism, and realizes the joint prediction of the rebound deviation and the defect probability for the first time. The process parameter correction value is dynamically generated based on the prediction result, and the quality control response speed reaches the millisecond level and the compensation accuracy is improved to ±2μm, which is more than 50% lower than the traditional PID control error.

[0008] The technical solution adopted by the present invention to solve the above technical problems is: a cold stamping quality control method based on multi-source data fusion and real-time optimization, which realizes the full process closed-loop control of the production process through the MES system, including the following steps: Step 1: The die pressure, hydraulic displacement, temperature and workpiece surface image data of the stamping equipment are collected through a distributed sensor network and transmitted to the MES system after processing; Step 2: Extract stamping energy characteristics through time-frequency analysis of pressure data, build a dynamic deformation model based on hydraulic displacement, calculate thermal compensation in combination with temperature, and extract defect characteristics through multi-scale morphological processing of images; Step 3: Normalize the energy characteristics, deformation model, thermal compensation and defect characteristics, generate weight coefficients based on the elastic modulus, and construct a three-dimensional fusion feature space through principal component analysis; Step 4: Input the feature space into the pre-trained LSTM model, associate the stamping speed with the holding time, and output the quality prediction matrix containing the springback deviation and defect probability; Step 5: When the springback deviation exceeds ±0.15mm and the defect probability is greater than 85%, the bus is sent to the PLC to adjust the stamping speed and holding time parameters; Step 6: Calculate the mold compensation amount based on the actual springback amount, link the lubrication system to grade and adjust the injection parameters, and perform micron-level compensation; Step 7: After completing 50 stampings, update the LSTM parameters with laser measurement data, and start incremental learning when the prediction error is greater than 0.1 mm for three consecutive times; Step 8: Build a traceable knowledge base in the MES system and store batch data to form a closed-loop archive.

[0009] Furthermore, the distributed sensor network in step 1 includes: A 16-point array piezoelectric pressure sensor group is arranged on the top of the mold cavity, with a sampling frequency of ≥5kHz, and transmits data via the CAN bus protocol; A sub-micron grating displacement sensor installed at the piston rod of the hydraulic cylinder; The working surface of the stamping die is embedded with a 9-point infrared temperature measurement array, and the temperature data is calibrated using the flight time algorithm; On both sides of the conveyor line, dual-light-source polarized industrial cameras are installed to synchronously trigger and capture defect images on the workpiece surface at the 0.05 mm level.

[0010] Furthermore, the construction of the dynamic deformation model based on hydraulic displacement in step 2 includes: Step 2-1: Perform sliding window filtering on the hydraulic displacement data, and adaptively adjust the window length; Step 2-2: Establish a material springback prediction model based on the J-C constitutive equation, and calculate the theoretical deformation amount in combination with the elastic modulus data; Step 2-3: Extract the energy spectrum characteristics of the pressure signal through wavelet packet decomposition, and divide the low-frequency stable region and the high-frequency impact region; Step 2-4: Perform gray projection analysis on the workpiece surface image, and optimize the defect boundary continuity in combination with morphological closing operation.

[0011] Furthermore, the generation of the weight coefficient based on the elastic modulus in step 3 includes: Step 3-1: Query the elastic modulus value of the current batch of aluminum alloy from the material database of the MES system; Step 3-2: Calculate the objective weight coefficient of the stamping energy characteristics based on the entropy weight method; Step 3-3: Generate a material property correction factor according to the elastic modulus; Step 3-4: Reduce the four-dimensional features to three-dimensional space through principal component analysis, and retain more than 95% of the variance information.

[0012] Furthermore, the pre-training of the LSTM model in step 4 includes: Step 4-1: Collect historical production data to construct a training set, and divide the input sequence according to the sliding time window; Step 4-2: Introduce an attention mechanism into the network structure to dynamically allocate the time weights of different features; Step 4-3: Use the Bayesian optimization algorithm to adjust the number of nodes in the LSTM hidden layer; Step 4-4: Perform weighted fusion of the real-time stamping speed parameter and the LSTM output gate.

[0013] Furthermore, the adjustment of the stamping speed and holding pressure time parameters in step 5 includes the following sub-steps: Step 5-1: Calculate the stamping speed correction value according to the thermal expansion coefficient of the aluminum alloy; Step 5-2: Construct a holding pressure time-temperature correlation curve based on historical process data, and dynamically adjust the holding pressure time parameter using fuzzy control rules; Step 5-3: Send the correction parameters to the PLC controller through the PROFINET industrial bus protocol to ensure that the response delay < 50 ms; Step 5-4: Generate parameter adjustment records in the MES system interface and associate the equipment operation log with the current process version number.

[0014] Furthermore, in step 6, the mold compensation amount is calculated based on the actual springback amount, and the injection parameters are adjusted in stages by linking the lubrication system. The micron-level compensation is performed, which includes the following sub-steps: Step 6-1: Analyze the mold deformation trend based on strain energy density and calculate the residual stress distribution; Step 6-2: The mold fine-tuning mechanism is controlled by a servo motor, with a compensation accuracy of ±2μm; Step 6-3: Lubrication levels are divided according to defect probability: when the defect probability is greater than 90%, the lubricant injection amount is double injection at level I; when the defect probability is between 70% and 90%, the lubricant injection amount is standard injection at level II; when the defect probability is less than 70%, the lubricant injection amount is reduced frequency injection at level III; Step 6-4: Use PID control algorithm to dynamically adjust the lubricant injection pressure to 0.2-0.8MPa.

[0015] Furthermore, the incremental learning optimization step in step 7 includes: Step 7-1: Construct a sliding time window to filter key historical data samples, including mold pressure curve, hydraulic displacement time series, temperature monitoring data and defect image samples; Step 7-2: Use Hampel filter to remove abnormal measurement data; Step 7-3: Freeze the weights of the first three layers of the LSTM network and only update the output layer parameters; Step 7-4: Reuse the feature extraction capability of the historical model through transfer learning technology.

[0016] Furthermore, the step 8 of constructing a traceable knowledge base in the MES system includes the following sub-steps: Step 8-1: Run the original data stream using the time series database storage device; Step 8-2: Build a knowledge graph to associate material batch number, mold number and quality defect pattern; Step 8-3: Deploy an OPC UA-based data interface to interconnect with the ERP system; Step 8-4: Establish a SQL traceability query engine to support multi-dimensional root cause analysis.

[0017] The advantages of the present invention are: 1. The present invention collects multi-modal data in real time through a distributed sensor network (pressure, displacement, temperature, image), combines time-frequency analysis, dynamic deformation modeling, and morphological processing to break through the limitation of a single data dimension; further generates dynamic weight coefficients based on the elastic modulus of the material, and constructs a three-dimensional fusion feature space with high information density using principal component analysis, solving the problems of feature redundancy and weight solidification in traditional methods, increasing the extraction efficiency of key quality features by more than 40%, and providing accurate input for subsequent prediction models.

[0018] 2. The present invention inputs the fusion feature space into a pre-trained LSTM model, associates real-time stamping speed and holding time parameters through an attention mechanism, and for the first time realizes the joint prediction of springback deviation (accuracy of ±0.15 mm) and defect probability (threshold of 85%), and dynamically generates process parameter correction values based on the prediction results. Combining the thermal expansion compensation algorithm and micron-level die compensation, the quality control response speed reaches the millisecond level, the compensation accuracy is improved to ±2 μm, and the error is reduced by more than 50% compared with traditional PID control. 3. The present invention realizes full-process closed-loop control through the MES system, updates the LSTM model using laser measurement data every 50 stampings, triggers incremental learning when the continuous prediction error > 0.1 mm, and quickly adapts to material parameter changes by combining transfer learning technology; at the same time, constructs a knowledge graph to associate process parameters with defect patterns, supports multi-dimensional root cause tracing, improves the quality defect location efficiency by 70%, and the system self-adaptability reaches the industrial-level continuous optimization standard. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a flowchart of a cold stamping quality control method based on multi-source data fusion and real-time optimization provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] Example 1: Figure 1The flowchart of a cold stamping quality control method based on multi-source data fusion and real-time optimization provided by the present invention is as follows Figure 1 A cold stamping quality control method based on multi-source data fusion and real-time optimization as shown in the figure realizes the full-process closed-loop control of the production process through the MES system, including the following steps: First, the die pressure, hydraulic displacement, working temperature and workpiece surface image data of the aluminum alloy stamping equipment are collected in real time through a distributed sensor network. After the data format is unified by an industrial protocol gateway, it is directly transmitted to the MES system; the time-frequency analysis is implemented on the pressure data to extract the stamping energy characteristics, a dynamic deformation model is constructed based on the hydraulic displacement data, the thermal compensation amount is calculated in combination with the working temperature data, and the multi-scale morphological processing is performed on the workpiece surface image to extract the defect characteristics; the stamping energy characteristics, deformation model, thermal compensation amount and defect characteristics are normalized, a dynamic weight coefficient is generated based on the material elastic modulus, and a three-dimensional fusion feature space is constructed through principal component analysis; the fusion feature space is input into the pre-trained LSTM model, the real-time stamping speed and the holding time parameters are correlated, and a quality prediction matrix including the springback deviation and the defect probability is output; when the springback deviation exceeds ±0.15 mm and the defect probability > 85%, a stamping speed correction value is generated according to the thermal expansion compensation amount, the process parameters are adjusted according to the holding time-temperature correlation curve, and are sent to the PLC for execution through the industrial bus; the die compensation amount is calculated based on the actual springback amount after execution, and the lubrication system is linked to adjust the injection parameters according to the defect probability grading, and the micron-level precision compensation is executed; after every 50 stamping operations are completed, the LSTM model weight parameters are updated using the laser three-dimensional measurement data, and incremental learning optimization is started when the continuous three prediction errors > 0.1 mm; a traceable knowledge base is constructed in the MES system, and the material parameters, process adjustment records and quality inspection data are stored according to the production batches to form a closed-loop control file.

[0023] Among them, the MES system realizes the full-process closed-loop control of the production process, which means that through the manufacturing execution system, real-time monitoring and feedback adjustment are carried out on each link of the cold stamping process. Specifically, it can be achieved by using an industrial protocol gateway for unified data format processing, issuing control instructions to the PLC for execution through the industrial bus protocol, and constructing a traceable knowledge base to store the full-process data. This feature realizes the dynamic adjustment of process parameters and the traceability of quality data, and solves the problem of data islands in traditional systems. Among them, the distributed sensor network for real-time collection of die pressure, hydraulic displacement, working temperature, and workpiece surface image data of aluminum alloy stamping equipment refers to a multi-modal perception system composed of a piezoelectric pressure sensor group, a grating displacement sensor, an infrared temperature measurement array, and a polarization industrial camera. Specifically, the CAN bus protocol is used to transmit pressure data, the time-of-flight algorithm is used to calibrate temperature data, and dual-light source synchronous triggering is used to collect surface defect images. This feature breaks through the limitation of the data dimension of a single sensor and improves the comprehensiveness of stamping process monitoring. Among them, time-frequency analysis for extracting stamping energy features refers to the extraction of spectral features and the analysis of energy distribution of pressure signals. Specifically, it is achieved by using sliding window filtering and wavelet packet decomposition to divide the low-frequency stable region and the high-frequency impact region. This feature solves the problem of energy evaluation error caused by traditional time-domain analysis ignoring frequency-domain features. Among them, the construction of a dynamic deformation model refers to establishing a material springback prediction model based on hydraulic displacement data. Specifically, the J-C constitutive equation is used in combination with elastic modulus data to calculate the theoretical deformation amount, and the continuity of the defect boundary is optimized through gray projection analysis and morphological closing operation. This feature improves the accuracy of deformation prediction. Among them, principal component analysis for constructing a three-dimensional fusion feature space refers to dimensionality reduction processing of stamping energy features, deformation models, thermal compensation amounts, and defect features. Specifically, it is achieved by using normalization to eliminate dimension differences, calculating objective weight coefficients based on the entropy weight method, and retaining more than 95% of the variance information. This feature solves the problem of multi-dimensional data redundancy and improves the efficiency of feature extraction. Among them, the pre-trained LSTM model outputs a quality prediction matrix, which means using a long short-term memory neural network for time-series feature modeling. Specifically, it is achieved by using an attention mechanism to allocate time weights, adjusting the number of hidden layer nodes through Bayesian optimization, and associating real-time stamping speed and holding time parameters. This feature realizes the joint prediction of springback amount deviation and defect probability. Among them, adjusting process parameters according to the holding time-temperature correlation curve means generating a stamping speed correction value based on the thermal expansion compensation amount. Specifically, it is achieved by using fuzzy control rules to adjust time parameters and issuing parameters to the PLC for execution through the PROFINET industrial bus protocol. This feature realizes the closed-loop control of process parameters with a millisecond-level response speed. Among them, die compensation amount calculation and micron-level precision compensation refer to the analysis of the die deformation trend based on the actual springback amount. Specifically, the strain energy density is used to analyze the residual stress distribution, and a servo motor is used to control the fine-tuning mechanism to perform ±2μm precision compensation. This feature solves the problem of insufficient accuracy in traditional compensation mechanisms.Among them, the incremental learning optimization of the LSTM model refers to updating the model parameters using the laser three-dimensional measurement data, which is specifically achieved by screening key historical data samples using a sliding time window, freezing the weights of the first three layers and only updating the parameters of the output layer, and reusing the feature extraction ability through transfer learning. This feature improves the adaptability of the model to changes in material parameters. Among them, the construction of the traceable knowledge base refers to storing the full-process data of production batches, which is specifically achieved by storing the original data stream in a time-series database, constructing a knowledge graph to associate defect patterns with process parameters, and deploying an OPC UA interface to achieve system interconnection. This feature supports multi-dimensional root cause tracing and rapid positioning of quality problems.

[0024] The core innovation of the present invention lies in realizing the synchronous acquisition of multi-source data in the stamping process through a distributed multi-modal sensor network, constructing a fusion feature space with high information density by combining dynamic weight allocation and principal component analysis, and jointly predicting the springback amount and defect probability based on the LSTM model. Finally, the process parameters and die compensation amount are closed-loop controlled through the MES system to form a full-process quality optimization system, solving the problems of low feature extraction efficiency, insufficient prediction accuracy, and system response delay in traditional methods.

[0025] The working process and principle of the present invention are as follows: multi-source data of aluminum alloy stamping equipment, including die pressure, hydraulic displacement, working temperature, and workpiece surface image, are collected in real time through a distributed sensor network. These data are transmitted to the MES system after being uniformly processed in format through an industrial protocol gateway. The stamping energy characteristics are extracted by performing time-frequency analysis on the pressure data, a dynamic deformation model is constructed based on the hydraulic displacement data, the thermal compensation amount is calculated by combining the working temperature data, and the defect characteristics are extracted by performing multi-scale morphological processing on the workpiece surface image. These features are normalized, and dynamic weight coefficients are generated based on the material elastic modulus. A three-dimensional fusion feature space is constructed through principal component analysis. The fusion feature space is input into a pre-trained LSTM model, and the real-time stamping speed and holding time parameters are associated to output a quality prediction matrix. When the springback deviation or defect probability exceeds the threshold, the system generates a stamping speed correction value according to the thermal expansion compensation amount, adjusts the process parameters and issues them to the PLC for execution. The die compensation amount is calculated based on the actual springback amount after execution, and the lubrication system is linked to adjust the spraying parameters. After every 50 stampings are completed, the system updates the LSTM model parameters using the laser three-dimensional measurement data and starts incremental learning optimization if necessary. The MES system constructs a traceable knowledge base and stores relevant data to form a closed-loop control file.

[0026] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: First, deploy a distributed sensor network on the cold stamping production line. Arrange a 16-point array piezoelectric pressure sensor group on the top of the mold cavity, set the sampling frequency to 5 kHz, and transmit data through the CAN bus protocol. Install a sub-micron grating displacement sensor at the piston rod of the hydraulic cylinder. Embed a 9-point infrared temperature measurement array on the working surface of the stamping die, and calibrate the temperature data using the time-of-flight algorithm. Install dual-light source polarization industrial cameras on both sides of the conveyor line to synchronously trigger the acquisition of workpiece surface images.

[0027] Second, process and extract features from the collected multi-source data. Perform wavelet packet decomposition on the pressure data, extract energy spectrum features, and divide the low-frequency stable region and high-frequency impact region. Perform sliding window filtering on the hydraulic displacement data, and adaptively adjust the window length. Establish a material springback prediction model based on the J-C constitutive equation, and calculate the theoretical deformation amount in combination with elastic modulus data. Perform gray projection analysis on the workpiece surface image, and optimize the defect boundary continuity in combination with morphological closing operation.

[0028] Furthermore, fuse the extracted features. Query the elastic modulus value of the current batch of aluminum alloy from the material database of the MES system. Calculate the objective weight coefficient of the stamping energy feature based on the entropy weight method. Generate a material property correction factor according to the elastic modulus. Reduce the four-dimensional features to three-dimensional space through principal component analysis, retaining more than 95% of the variance information.

[0029] Next, input the fused features into a pre-trained LSTM model for quality prediction. The pre-training process of the LSTM model includes: collecting historical production data to construct a training set, dividing the input sequence according to a sliding time window; introducing an attention mechanism into the network structure to dynamically allocate time weights for different features; using the Bayesian optimization algorithm to adjust the number of nodes in the LSTM hidden layer; and performing weighted fusion of the real-time stamping speed parameter and the LSTM output gate. When the prediction result shows that the springback deviation exceeds ±0.15 mm and the defect probability is greater than 85%, the system calculates the stamping speed correction value according to the coefficient of thermal expansion. Construct a holding time-temperature correlation curve, and adjust the time parameter using fuzzy control rules. Send the parameters to the PLC through the PROFINET industrial bus protocol, and control the response delay within 50 ms. Thus, the system analyzes the deformation trend of the die based on strain energy density and calculates the residual stress distribution. Control the die fine-tuning mechanism through a servo motor, with a compensation accuracy of ±2 μm. Divide the lubrication level according to the defect probability, and dynamically adjust the lubricant injection pressure to 0.2 - 0.8 MPa.

[0030] After every 50 stamping operations are completed, the system updates the LSTM model using laser three-dimensional measurement data. When the prediction error is greater than 0.1 mm for three consecutive times, incremental learning optimization is initiated. A sliding time window is constructed to screen key historical data samples, and the Hampel filter is used to eliminate abnormal measurement data. The weights of the first three layers of the LSTM network are frozen, and only the parameters of the output layer are updated. The feature extraction ability of the historical model is reused through transfer learning technology.

[0031] Finally, a traceable knowledge base is constructed in the MES system. The original data stream of equipment operation is stored using a time-series database. A knowledge graph is constructed to associate material batch numbers, die numbers, and quality defect patterns. A data interface based on OPC UA is deployed to achieve interconnection with the ERP system. An SQL trace query engine is established to support multi-dimensional root cause analysis.

[0032] Through the above solutions, the present invention realizes the real-time collection and fusion processing of multi-source data, improves the efficiency and accuracy of feature extraction. The LSTM model combines real-time process parameters for quality prediction, improving the prediction accuracy. The real-time process parameter adjustment and die compensation mechanism based on the prediction results improve the response speed and accuracy of stamping quality control. Incremental learning and knowledge base construction achieve the continuous optimization and quality traceability of the system. These technical means work together to effectively solve the problems in traditional cold stamping quality control, such as limited data dimensions, low prediction accuracy, lagging parameter adjustment, and difficulty in quality traceability, improving the quality stability and efficiency of cold stamping production.

[0033] The present invention further proposes a distributed sensor network, which includes a 16-point array piezoelectric pressure sensor group arranged at the top of the die cavity, with a sampling frequency ≥5 kHz and data transmitted through the CAN bus protocol; a sub-micron grating displacement sensor installed at the piston rod of the hydraulic cylinder; a 9-point infrared temperature measurement array embedded in the working surface of the stamping die, and the temperature data is calibrated using the time-of-flight algorithm; dual-light source polarization industrial cameras are installed on both sides of the conveyor line to synchronously trigger the acquisition of 0.05 mm-level defect images on the surface of the workpiece.

[0034] Among them, the 16-point array piezoelectric pressure sensor group at the top of the mold cavity covers the mold working surface through equidistant circular arrangement. The interval distance between each sensor is 1 / 8 of the mold diameter, ensuring the integrity of the pressure field distribution detection. The piezoelectric sensor group cooperates with the CAN bus protocol to achieve multi-channel parallel transmission, and the data packet interval time ≤ 0.2 ms to avoid high-frequency signal distortion. The sub-micron grating displacement sensor at the piston rod of the hydraulic cylinder controls the distance between the magnetic grating scale and the reading head in the range of 0.3 - 0.5 mm, with a resolution of 0.1 μm, capable of capturing displacement mutations during the rapid springback stage. The 9-point infrared temperature measurement array on the working surface of the stamping die adopts an isothermal zone division strategy, and the coverage area of each temperature measurement point does not exceed 30 mm². The time-of-flight algorithm calculates the time difference between the emission and reception of infrared pulses, controlling the temperature calibration error within ±1.5°C. The dual-light source polarization industrial camera forms a 30° cross-illumination angle on both sides of the conveyor line. The trigger signal is synchronized with the stroke switch of the stamping machine, and the exposure time is set to 1 / 120 of the stamping cycle to ensure that the error between the image acquisition moment and the workpiece arrival time < 2 ms.

[0035] Specifically, the array piezoelectric pressure sensor group overcomes the locality defect of single-point detection through multi-point layout, and can capture transient pressure fluctuations with a duration ≥ 0.2 ms at a sampling frequency of 5 kHz. The high-resolution design of the grating displacement sensor enables the micron-level displacement change of the hydraulic cylinder to be accurately quantified, providing basic data for the dynamic deformation model. The spatial distribution density of the infrared temperature measurement array cooperates with the time-of-flight algorithm to effectively eliminate the influence of the mold surface oxide layer on the emissivity and real-time output the corrected temperature field distribution of the working surface. The dual-light source polarization camera eliminates the reflection interference on the metal surface through cross-illumination, and the synchronous trigger mechanism ensures obtaining a complete image of the workpiece surface within the stamping cycle. These sensors are transmitted through a mixture of multiple protocols such as CAN bus and PROFINET, achieving timestamp alignment within the MES system and forming a multi-dimensional data stream. For example, during the stamping stroke, when the pressure sensor group captures an instantaneous impact force of 12.5 kN, the grating displacement sensor synchronously records a displacement jump of 0.8 mm, the infrared temperature measurement array detects a temperature rise of 8°C on the mold working surface, and the polarization camera obtains a micro-crack image of 0.06 mm at the edge of the workpiece. The spatio-temporal alignment of multi-source data provides accurate input for subsequent feature fusion, reducing the stamping energy feature extraction error to within 3%.

[0036] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: The distributed sensor network includes a 16-point array piezoelectric pressure sensor group arranged on the top of the die cavity, with a sampling frequency of 5 kHz and data transmitted through the CAN bus protocol. A sub-micron grating displacement sensor is installed at the piston rod of the hydraulic cylinder. A 9-point infrared temperature measurement array is embedded in the working surface of the stamping die, and the temperature data is calibrated using the time-of-flight algorithm. Dual-light-source polarized industrial cameras are installed on both sides of the conveyor line for synchronously triggering and collecting defect images with a size of 0.05 mm on the workpiece surface.

[0037] Specifically, the 16-point array piezoelectric pressure sensor group is arranged in a 4x4 matrix, covering the key area on the top of the die cavity. The measuring range of each pressure sensor is 0 - 1000 MPa, and the linearity is better than 0.1% FS. The resolution of the sub-micron grating displacement sensor is 0.1 μm, and the measuring range is 0 - 500 mm. The 9-point infrared temperature measurement array uses thermopile sensors, with a temperature measurement range of 0 - 300 °C and an accuracy of ±1 °C. The resolution of the dual-light-source polarized industrial camera is 2048x2048 pixels, and the frame rate can reach 200 fps.

[0038] Furthermore, the CAN bus adopts the high-speed CAN 2.0B protocol with a baud rate of 1 Mbps. The time-of-flight algorithm calibrates the temperature data by measuring the propagation time of infrared light in the air, eliminating the influence of environmental factors. The dual-light-source polarized industrial camera uses polarization filters and cross-polarized light sources to enhance the contrast of surface defects.

[0039] Through the above technical solutions, the present invention realizes the high-precision synchronous acquisition of multi-source data. The comprehensive acquisition of pressure, displacement, temperature, and image data lays a foundation for subsequent multi-modal data fusion. The high sampling frequency and accurate measuring range ensure the timeliness and accuracy of the data. The CAN bus protocol and the time-of-flight algorithm improve the reliability of data transmission and the accuracy of temperature measurement. The dual-light-source polarization imaging technology enhances the detection ability of surface defects. These improvements effectively enhance the comprehensive monitoring ability of the cold stamping process and provide comprehensive and accurate data support for quality control.

[0040] The present invention further proposes the construction of a dynamic deformation model, specifically including the following steps: performing sliding window filtering on the hydraulic displacement data with the window length adaptively adjusted; establishing a material springback prediction model based on the J-C constitutive equation and calculating the theoretical deformation amount in combination with the elastic modulus data; extracting the energy spectrum characteristics of the pressure signal through wavelet packet decomposition and dividing the low-frequency stable area and the high-frequency impact area; performing gray projection analysis on the workpiece surface image and optimizing the continuity of the defect boundary in combination with morphological closing operation.

[0041] Among them, the window length of the sliding window filter is automatically adjusted according to the spectral characteristics of the hydraulic displacement signal. When it is detected that the change in the signal frequency component exceeds the set threshold, the window length is dynamically adjusted within the range of 100 - 500 ms in steps of 50 ms; the J-C constitutive equation introduces the strain rate sensitivity coefficient and the temperature softening effect parameter, and combines the material elastic modulus to calculate the theoretical deformation amount. The model inputs include the stamping speed, the holding pressure time, and the measured temperature data; wavelet packet decomposition uses the db4 wavelet basis function for 5-layer decomposition, and divides the pressure signal energy spectrum into a low-frequency stable region of 0 - 400 Hz and a high-frequency impact region of 400 - 2000 Hz; gray-scale projection analysis generates projection histograms along the X / Y axis directions of the workpiece surface image, and the boundary is optimized by morphological closing operation using a 3×3 circular structural element.

[0042] Specifically, after the hydraulic displacement signal is filtered by the adaptive sliding window, high-frequency noise is effectively suppressed and the true deformation characteristics are retained. The dynamic adjustment of the window length is achieved by real-time calculating the signal power spectral density, and the window length change is triggered when the main frequency band fluctuation exceeds ±10%. For the deformation model established based on the J-C constitutive equation, by introducing the strain hardening index n value and the strain rate sensitivity coefficient m value, and combining the measured data of the elastic modulus E, the theoretical deformation amount ΔL = ƒ(σ, ε˙, T) is calculated, where σ is the equivalent stress, ε˙ is the strain rate, and T is the die temperature. After the pressure signal is decomposed by wavelet packet, the energy ratio of the low-frequency stable region is used to evaluate the stability of the stamping process, and when the energy peak value in the high-frequency impact region exceeds the threshold, the equipment status warning is triggered. After the defect area is located by the gray-scale projection analysis of the workpiece surface image, the morphological closing operation fills the boundary fracture pixel points, so that the defect contour closure degree is increased to more than 95%, providing continuous boundary data for subsequent feature extraction. Through the collaborative execution of the above steps, the prediction error of the dynamic deformation model is reduced to within 0.05 mm, the recognition accuracy of the high-frequency impact characteristics is increased to 98%, and the defect boundary positioning accuracy reaches the level of ±2 pixels.

[0043] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: The construction of the dynamic deformation model includes the following steps: performing sliding window filtering on the hydraulic displacement data, and the window length is adaptively adjusted. Specifically, the Kalman filter algorithm is used to preprocess the hydraulic displacement data, and the sliding window length is dynamically adjusted according to the data fluctuation amplitude. When the fluctuation amplitude is large, the window length is shortened, and when the fluctuation amplitude is small, the window length is extended to balance the filtering effect and real-time performance.

[0044] A material springback prediction model is established based on the J-C constitutive equation, and the theoretical deformation amount is calculated by combining the elastic modulus data. Further, the Johnson-Cook constitutive equation in the prior art is used to describe the mechanical behavior of aluminum alloy materials at high strain rates. Combining the real-time obtained elastic modulus data, the theoretical deformation amount is calculated through finite element analysis. The energy spectrum characteristics of the pressure signal are extracted by wavelet packet decomposition, and the low-frequency stable region and the high-frequency impact region are divided. Among them, the db4 wavelet basis is used to perform 5-layer wavelet packet decomposition on the pressure signal, the energy characteristics of each frequency band are extracted, and the signal is divided into the low-frequency stable region and the high-frequency impact region according to the energy distribution. Gray projection analysis is performed on the workpiece surface image, and the continuity of the defect boundary is optimized by combining morphological closing operation. Thus, first, the workpiece surface image is grayscale processed, and then gray projection analysis in the horizontal and vertical directions is performed to identify potential defect regions. Further, a morphological closing operation is adopted, and a 3x3 structural element is used to optimize the defect boundary to improve the boundary continuity.

[0045] Through the above technical solutions, the present invention realizes the adaptive filtering processing of hydraulic displacement data and improves the data quality. The material springback prediction model established based on the J-C constitutive equation can accurately describe the mechanical behavior of aluminum alloy materials during the stamping process. The wavelet packet decomposition method effectively extracts the energy spectrum characteristics of the pressure signal, providing a reliable basis for subsequent analysis. The gray projection analysis combined with the morphological operation optimizes the defect boundary recognition effect. The comprehensive application of these technical means significantly improves the accuracy and reliability of the dynamic deformation model, providing strong support for cold stamping quality control.

[0046] The present invention further proposes that the generation of the dynamic weight coefficient includes the following sub-steps: querying the elastic modulus value of the current batch of aluminum alloy from the material database of the MES system; calculating the objective weight coefficient of the stamping energy characteristics based on the entropy weight method; generating a material characteristic correction factor according to the elastic modulus; reducing the four-dimensional characteristics to three-dimensional space through principal component analysis, retaining more than 95% of the variance information.

[0047] Among them, the elastic modulus query operation is realized through the material batch tracking module built in the MES system. The database stores the measured value range of the elastic modulus of aluminum alloy plates, and the data update frequency is synchronized with the production batch. During the calculation process of the entropy weight method, the pressure time-frequency characteristics, the output value of the displacement dynamic model, the thermal compensation amount, and the defect area parameters are constructed into an evaluation matrix, and the dispersion degree of each characteristic index is calculated through information entropy to generate the initial weight coefficient. The material characteristic correction factor is dynamically adjusted according to the deviation ratio between the measured value and the standard value of the elastic modulus. When the elastic modulus is higher than the standard value, the correction factor increases the weight of the displacement model according to a linear relationship; when the elastic modulus is lower than the standard value, the weight of the thermal compensation amount is increased. During the principal component analysis process, the four-dimensional characteristic matrix is orthogonally transformed into a three-dimensional space vector, and the information retention amount after dimensionality reduction is controlled by the cumulative variance contribution rate threshold.

[0048] Specifically, in the material database of the MES system, the elastic modulus values of the current batch of aluminum alloy are stored in XML format, including the melting furnace number, rolling batch number, and third-party inspection report number. The query operation matches the production work order number with the material warehousing record through SQL statements to obtain the measured elastic modulus value in real time. During the calculation of the entropy weight method, each index of the stamping energy feature matrix is normalized to a value in the range of 0-1. In the process of calculating the information entropy, the natural logarithm is used to calculate the entropy value of each index, and the objective weight coefficient is determined by the proportion of the difference between the entropy value and 1. The material property correction factor is generated through the elastic modulus ratio function. For example, when the measured elastic modulus is 70 GPa and the standard value is 69 GPa, the correction factor increases the weight of the displacement model by 1.4%. When performing principal component analysis, the eigenvalues of the covariance matrix are arranged in descending order, and the first three principal components are selected to make the cumulative variance contribution rate reach the 95% threshold. The coordinates of the three-dimensional feature vector after dimensionality reduction are calculated through the linear combination of the eigenvector and the original data. This process ensures that the dynamic weight allocation of different material batches not only reflects the objective law of the data but also combines the physical properties of the materials, enabling the fusion feature space to eliminate redundant dimensions while retaining key quality information.

[0049] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: The generation of the dynamic weight coefficient includes the following sub-steps: Query the elastic modulus value of the current batch of aluminum alloy from the material database of the MES system. Specifically, the elastic modulus value of the aluminum alloy material used in the current production batch is extracted from the material database of the MES system through an SQL query statement. For example, for 6061-T6 aluminum alloy, its elastic modulus is approximately 68.9 GPa.

[0050] Calculate the objective weight coefficient of the stamping energy feature based on the entropy weight method. Further, the entropy weight method is used to process multi-dimensional data such as pressure and displacement collected during stamping to calculate the information entropy of each feature, and then the objective weight coefficient is obtained. Among them, the greater the information entropy, the higher the uncertainty of the feature, and the smaller its weight coefficient.

[0051] Generate a material property correction factor according to the elastic modulus. Thus, the obtained elastic modulus value is substituted into a preset correction formula to generate a material property correction factor for adjusting the weight. For example, the calculation method of the correction factor = 1 + (E - E0) / E0 can be adopted, where E is the elastic modulus of the current batch of materials and E0 is the reference elastic modulus.

[0052] The four-dimensional features are reduced to three-dimensional space through principal component analysis, retaining more than 95% of the variance information. Specifically, the four-dimensional feature data such as stamping energy characteristics, deformation models, thermal compensation amounts, and defect characteristics are input into the principal component analysis algorithm. By calculating the eigenvalues and eigenvectors, the first three principal components with a cumulative contribution rate of more than 95% are selected to achieve dimensionality reduction processing.

[0053] Through the above technical solutions, the present invention realizes dynamic weight adjustment based on the elastic modulus of the material, improving the accuracy of feature fusion. At the same time, the data dimension is effectively reduced through principal component analysis, key information is retained, and the subsequent processing efficiency is improved. This method overcomes the limitations of the traditional fixed weight allocation method, enabling the fused features to better reflect the material characteristics and processing status, providing a more reliable data basis for cold stamping quality control.

[0054] The present invention further proposes a pre-training method for the LSTM model, including the following steps: collecting historical production data to construct a training set, and dividing the input sequence according to a sliding time window; introducing an attention mechanism into the network structure to dynamically allocate time weights for different features; using the Bayesian optimization algorithm to adjust the number of nodes in the LSTM hidden layer; and performing weighted fusion of the real-time stamping speed parameter and the LSTM output gate.

[0055] Among them, the historical data training set adopts a sliding time window mechanism, and the window length is dynamically adjusted according to the stamping rhythm; the attention mechanism realizes feature weight allocation by calculating the time-step correlation matrix; the Bayesian optimization takes the prediction error as the optimization objective function to search for the optimal number of hidden layer nodes; the output gate weighted fusion uses a gating mechanism to linearly superimpose the real-time process parameters and the model output.

[0056] Specifically, the historical production data is segmented into sliding windows of 50 time steps according to the stamping cycle, and each window contains sequences of pressure, displacement, temperature, and image features. The attention mechanism generates a weight distribution by calculating the cosine similarity of each time step, enabling the hydraulic displacement mutation feature to obtain a higher weight. The Bayesian optimization determines the number of hidden layer nodes to be 128 in 20 iterations, reducing the training error by 12% compared to the fixed structure. In the output gate fusion stage, the real-time stamping speed parameter is multiplied by the LSTM hidden state to form a speed-state joint vector, and the Sigmoid function is used to control the information transfer ratio. This fusion mechanism reduces the standard deviation of the prediction error of the model under the condition of sudden change in stamping speed from 0.08 mm to 0.05 mm.

[0057] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: The pre-training of the LSTM model includes the following steps: First, collect historical production data to construct a training set and divide the input sequence according to a sliding time window. Specifically, extract the cold stamping production data of the past six months from the MES system database, including multi-dimensional time series such as pressure curves, displacement time series, and temperature changes. Use a 30-second sliding time window and divide the training samples at a step of 5 seconds to generate approximately 100,000 training sequences.

[0058] Then, introduce an attention mechanism into the network structure to dynamically allocate time weights for different features. When specifically implemented, add an attention layer after the output of the LSTM unit to calculate the weight coefficients at different time steps. The attention weights are normalized through the softmax function to achieve adaptive weighting of the features at key time points.

[0059] Furthermore, use the Bayesian optimization algorithm to adjust the number of nodes in the LSTM hidden layer. First, define the search space as 32 to 256 nodes and set the Gaussian process regression model as the surrogate model. Through iterative optimization, evaluate the performance of models with different numbers of nodes on the validation set, and finally determine that the optimal number of hidden layer nodes is 128. The present invention can perform weighted fusion of the real-time stamping speed parameter and the LSTM output gate. The specific operation is to perform element-wise multiplication fusion of the stamping speed value at the current moment after linear transformation and the activation value of the LSTM output gate. Thereby, introduce the real-time process parameter information into the prediction process of the network and improve the adaptability of the model to dynamic working conditions.

[0060] Through the above technical solutions, the present invention can make full use of historical production data to construct a high-quality training set, extract key time series features through the attention mechanism, and determine the optimal network structure in combination with Bayesian optimization. At the same time, integrate real-time process parameters into the prediction process, improving the modeling ability and prediction accuracy of the LSTM model for the dynamic changes in the cold stamping process. This method effectively solves problems such as the difficulty of traditional models in capturing long-term dependence relationships and insufficient extraction of features at key time points, providing a reliable data basis for subsequent quality prediction and process parameter optimization.

[0061] The present invention further proposes to adjust the process parameters according to the holding pressure time-temperature correlation curve, including the following sub-steps: calculate the stamping speed correction value according to the coefficient of thermal expansion; construct the holding pressure time-temperature correlation curve and adjust the time parameter using fuzzy control rules; send the parameters to the PLC through the PROFINET industrial bus protocol, with a response delay < 50ms; generate a parameter adjustment record on the MES interface and associate the device operation log with the process version number.

[0062] Among them, the stamping speed correction value is dynamically calculated according to the deviation between the material thermal expansion coefficient and the measured temperature, and the correction range is 0.15 - 0.8 mm / s; the holding time-temperature correlation curve is constructed by the piecewise linear interpolation method, and the fuzzy control rules are combined to process the non-linear mapping relationship between temperature and time; the PROFINET protocol ensures the real-time transmission of parameters through the time-sensitive network technology, and the data frame structure includes a timestamp verification field; the parameter adjustment records are stored in XML format, and the associated fields include the device ID, operator number, and process version hash value.

[0063] Specifically, the speed correction amount is determined by multiplying the thermal expansion coefficient by the measured value of the temperature sensor, so that the stamping speed adapts to the thermal deformation trend of the mold. The fuzzy control rules quantify the temperature deviation and change rate into five fuzzy levels, and the defuzzification outputs the time adjustment amount through the weighted average method to solve the non-linear problem of the temperature-time relationship. The PROFINET protocol adopts the periodic synchronous transmission mode, sending control instructions every 8 ms, and the timestamp verification mechanism eliminates the influence of network jitter. The parameter adjustment records and the device operation logs are bidirectionally bound through the transaction ID, and the MES system automatically generates the version number and associates it with the material batch for storage. Thus, the closed-loop adjustment of process parameters is completed within 0.2 seconds from temperature monitoring, and the parameter adjustment records support version traceability queries in three dimensions: time, device, and operator.

[0064] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: Calculate the stamping speed correction value according to the thermal expansion coefficient. First, obtain the thermal expansion coefficient of the current workpiece material, and then combine the measured mold temperature in real time to calculate the thermal expansion amount of the material. Based on the deviation between the thermal expansion amount and the target size, use the preset speed-size relationship curve to determine the correction value of the stamping speed. Construct the holding time-temperature correlation curve and adjust the time parameter using fuzzy control rules. Through the analysis of historical production data, establish the correlation curve between the holding time and the mold temperature. Take the current mold temperature as the input, and after fuzzy processing, infer the adjustment amount of the holding time according to the preset fuzzy rule base.

[0065] Send the parameters to the PLC through the PROFINET industrial bus protocol, and the response delay is within 50 ms. Utilize the real-time performance of the PROFINET protocol to package the corrected stamping speed and holding time parameters into a data frame and transmit them to the PLC controller through the industrial Ethernet. The PLC immediately executes the update after receiving the parameters, and the response delay of the whole process is controlled within 50 milliseconds.

[0066] Generate parameter adjustment records in the MES interface, associate device operation logs with process version numbers. After each parameter adjustment, the MES system automatically generates an adjustment record, including information such as the adjustment time, parameter values before and after adjustment, and the reasons triggering the adjustment. At the same time, this record is associated and stored with the current device operation logs and process version numbers for subsequent traceability and analysis.

[0067] Through the above technical solutions, the present invention realizes real-time and precise adjustment of stamping process parameters. Based on thermal expansion compensation and fuzzy control, the stamping speed and holding pressure time are dynamically optimized, improving the workpiece forming accuracy. The use of high-speed industrial bus transmission ensures the timeliness of parameter adjustment, minimizing production delays to the greatest extent. The complete parameter adjustment record and version management mechanism provide reliable data support for quality problem analysis, effectively improving the quality control level of the cold stamping process.

[0068] The present invention further proposes a method for calculating the die compensation amount, including the following steps: performing sliding window filtering on the hydraulic displacement data with the window length adaptively adjusted; establishing a material springback prediction model based on the J-C constitutive equation and calculating the theoretical deformation amount in combination with the elastic modulus data; extracting the energy spectrum characteristics of the pressure signal through wavelet packet decomposition and dividing the low-frequency stable region and high-frequency impact region; performing gray projection analysis on the workpiece surface image and optimizing the defect boundary continuity in combination with morphological closing operation. Analyzing the die deformation trend based on the strain energy density and calculating the residual stress distribution; controlling the die fine-tuning mechanism through a servo motor with a compensation accuracy of ±2μm; dividing the lubrication level according to the defect probability: when the defect probability > 90%, the injection amount of the lubricant is double injection of grade I, when the defect is 70% - 90%, the injection amount of the lubricant is the reference injection of grade II, and when the defect probability < 70%, the injection amount of the lubricant is the reduced-frequency injection of grade III; using the PID control algorithm to dynamically adjust the lubricant injection pressure to 0.2 - 0.8 MPa.

[0069] Among them, the die deformation trend is determined based on the strain energy density analysis, and the residual stress distribution data is obtained through finite element simulation calculation; the servo motor drives the die fine-tuning mechanism through a ball screw with a displacement resolution of 0.1μm; the lubrication level division is based on the statistical distribution model of the defect probability, and the probability threshold is determined through training with historical production data; the injection pressure adjustment uses a closed-loop PID control with the control period synchronized with the die compensation action.

[0070] Specifically, for the calculation of the die compensation amount, the deformation trend of the die is first determined through strain energy density analysis. The residual stress distribution data is obtained through finite element simulation and used to guide the compensation direction of the servo motor. The servo motor drives the die fine-tuning device through a ball screw mechanism with a displacement resolution of 0.1 μm, achieving a compensation accuracy of ±2 μm. The lubrication system automatically switches the working mode according to the online calculated defect probability: when the defect probability exceeds 90%, the double injection mode is triggered, and the injection volume is increased to 200% of the reference value; when the defect probability is in the range of 70%-90%, the standard injection volume is maintained; when the probability is less than 70%, the reduced-frequency injection mode is adopted, and the injection frequency is reduced by 50%. The injection pressure is dynamically adjusted within the range of 0.2-0.8 MPa by a PID controller, and the pressure sensor provides real-time feedback to form a closed-loop control. The control period is synchronized with the die compensation action to ensure the timing matching of the lubricant supply and mechanical compensation. Through the synergistic effect of hierarchical lubrication control and dynamic pressure regulation, this solution reduces the lubricant consumption by 30% while ensuring the consistency of the compensated surface quality.

[0071] Specifically, the lubrication levels are classified as follows: Defect suppression: The injection volume of the lubricant directly affects the contact state between the material and the die. High defect probability (>90%): Double injection (Level I) can form a thicker oil film, reducing the risk of surface scratches and material tearing. Low defect probability (<70%): Reduced-frequency injection (Level III) avoids the material sliding offset caused by an overly thick oil film and prevents secondary deviation of the springback amount. Association with the compensation amount: When the defect probability is high (such as material cracking), the stiffness distribution of the workpiece changes, and the die compensation amount needs to be recalibrated. PID pressure regulation: Dynamically matching the compensation amount: The injection pressure (0.2-0.8 MPa) is controlled in real time through the PID algorithm to ensure the synchronization of the lubricant flow rate and the compensation action. High pressure (0.8 MPa): Quickly form a lubricating film to cooperate with the rapid fine-tuning of the die. Low pressure (0.2 MPa): Micro-lubrication avoids interfering with the die-material contact state after high-precision compensation.

[0072] Through the above steps, the closed-loop linkage of mechanical compensation and process compensation can be achieved. Forward control chain: Die compensation (mechanical) → Correct geometric deviation → Reduce springback; Lubrication adjustment (process) → Suppress surface defects → Stabilize material mechanical properties → Improve the accuracy of compensation amount calculation. Reverse feedback chain: If insufficient lubrication leads to defects (such as scratches) → Change in the actual stiffness of the workpiece → Failure of springback prediction → Re-trigger model update is required. It can solve the limitations of single compensation means. Through Class I lubrication + die compensation, the springback control accuracy can be improved from ±0.15 mm to ±0.05 mm. Through the above method, the accurate calculation and execution of die compensation amount are realized. Based on strain energy density analysis, the high-stress areas of the die can be accurately identified, providing an accurate basis for compensation. A fine-tuning mechanism driven by a servo motor is adopted to achieve high-precision compensation of ±2 μm. The lubrication level is dynamically adjusted according to the defect probability, effectively reducing surface defects caused by friction. The PID algorithm controls the injection pressure to ensure uniform distribution of the lubricant. These measures comprehensively improve the stability of the cold stamping process and product quality.

[0073] The present invention further proposes a method for constructing a traceable knowledge base in the MES system, including the following steps: using a time series database to store the original data stream of equipment operation; constructing a knowledge graph to associate material batch numbers, die numbers with quality defect patterns; deploying a data interface based on OPC UA to achieve interconnection with the ERP system; establishing an SQL trace query engine to support multi-dimensional root cause analysis.

[0074] Among them, the time series database adopts a distributed architecture to store tens of thousands of sensor data streams per second, and the data compression rate is not less than 70%; the knowledge graph establishes a triple relationship model through a graph database, and the nodes include material physical properties and process parameters; the OPC UA interface realizes two-way communication of 2000 data per second; the SQL engine supports time series correlation query, and the response time is less than 3 seconds.

[0075] Specifically, the original data stream of equipment operation is accessed into the time series database through an industrial protocol gateway, and a time stamp partition storage strategy is adopted. Among them, the pressure data is stored at a sampling rate of 5 kHz, and the temperature data is archived at an interval of 0.1 second. When constructing the knowledge graph, the batch number of the aluminum alloy material and the die number are used as entity nodes, and relationship edges are established through quality defect patterns. At the same time, the elastic modulus parameter and historical process adjustment records are associated. Based on the OPC UA data interface, a secure communication channel is configured to realize real-time synchronization of production order data between the MES system and the ERP system, and the transmission delay is controlled within 50 ms. The SQL trace query engine is built-in with a time window function, which can execute cross-data table correlation queries. For example, the pressure curve and temperature monitoring values corresponding to a specific period are automatically associated through the defect occurrence time, realizing multi-dimensional root cause location of quality problems, and the defect analysis efficiency is increased by 70%.

[0076] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: When building a traceable knowledge base in the manufacturing execution system, first use a time series database to store the original data stream of equipment operation. The data stream records the die pressure curve, hydraulic displacement time series, and temperature monitoring data of the stamping equipment in a time series format. Further, build a knowledge graph through a graph database, establish a triple relationship between the material batch number, die number, and quality defect mode. Among them, the material batch number is associated with the material elastic modulus parameter, the die number is associated with the historical compensation record, and the quality defect mode classifies and stores the morphological feature data of cracks, wrinkles, and scratches. Subsequently, deploy a data interface module based on the OPC UA protocol to achieve the interconnection and interoperability between the manufacturing execution system and the enterprise resource planning system, and map the equipment operation parameters and production order information through a unified address space model. Finally, establish a structured query language traceability engine to support multi-dimensional combined queries by time range, defect type, and material batch. The traceability query engine optimizes the query response efficiency through pre-compiled statements, and automatically generates stamping process parameter adjustment suggestions in association with the root cause analysis algorithm.

[0077] Through the above technical solution, the present invention effectively solves the problem of data islands in the quality traceability process, and realizes the efficient association and rapid retrieval of cross-system data. The construction of the knowledge graph enables the organic connection of material attributes, die states, and defect modes. The structured query language engine quickly locates abnormal batches through multi-dimensional combined conditions. The OPC UA interface breaks down the barriers between production data and management systems, thereby forming a complete quality traceability chain, significantly improving the accuracy and efficiency of process anomaly analysis.

[0078] The present invention further proposes a method for building a traceable knowledge base in the MES system, including using a time series database to store the original data stream of equipment operation, building a knowledge graph to associate the material batch number, die number, and quality defect mode, deploying a data interface based on OPC UA to achieve interconnection with the ERP system, and establishing an SQL traceability query engine to support multi-dimensional root cause analysis.

[0079] Among them, the time series database adopts a circular buffer structure, compresses and stores the original waveform data of pressure, displacement, and temperature according to millisecond-level timestamps, and the compression ratio is controlled within 80%. The knowledge graph establishes a triple relationship through a graph database, and probabilistically associates the stamping defect mode with the deviation value of the material elastic modulus and the number of die fine-tuning times. The OPC UA interface is configured with a bidirectional communication protocol to achieve automatic mapping of process parameter adjustment records and ERP production work orders. The SQL traceability query engine is built with a multi-condition indexing algorithm to support combined queries by time interval, equipment number, and defect type, and the response time is less than 2 seconds.

[0080] Specifically, the pressure and displacement signals generated during the operation of the equipment are stored in the time-series database in the form of time series. Each record contains a timestamp, equipment ID, and the original value. The storage period is set to continuously roll over and cover 12 months. The knowledge graph is constructed through the Neo4j graph database. The nodes include die numbers, material batch numbers, and defect type codes. The edge attributes record the correlation strength coefficient. When querying a specific defect pattern, the engine automatically traverses the associated nodes to generate a causal chain. The OPC UA interface establishes a secure channel between MES and ERP. The process parameter adjustment instructions are encapsulated as UA node information and are bidirectionally bound to the material code and process version number in the ERP work order. The SQL traceability engine adopts a columnar storage structure, builds a B+ tree index for the quality inspection data, and optimizes the IO load through predicate pushdown technology when performing a joint query, achieving a second-level response for 500,000 record levels. This method shortens the root cause location time of stamping defects from 30 minutes to within 5 minutes, and improves the data correlation accuracy rate to 98%.

[0081] As a preferred embodiment, the solution of the present invention is specifically implemented as follows: Configure a storage architecture based on the time-series database within the MES system, and deploy an SQL query engine with multi-threaded processing capabilities. This engine has a built-in data index module and establishes an association relationship between the material batch number and the process parameter table. Specifically, create a defect pattern dimension table in the data table design, which includes stamping batch numbers, die numbers, and sensor data fingerprint fields, and realizes bidirectional association with the quality inspection result table through foreign key constraints. When the query engine receives a composite condition query request containing a time range, a material elastic modulus interval, and a die pressure peak value, it automatically parses the query conditions and generates an execution plan, and quickly locates the matching records by parallel scanning the index. For example, when it is detected that a certain batch of workpieces has edge crack defects, the operator inputs a timestamp range, a die number, and a temperature fluctuation threshold. The query engine jointly scans the process parameter history table and the defect image feature table to generate an associated result set containing the material batch number, the hold pressure time adjustment record, and the corresponding defect probability.

[0082] Through the above technical solution, the present invention realizes the cross-table association query of the process parameter adjustment record and the quality inspection data, solves the problem of low efficiency of manually retrieving data across systems, enables the root cause analysis of quality problems to quickly locate specific material batches or abnormal equipment operation parameters based on multi-dimensional data association, and effectively shortens the time cycle of defect traceability.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cold stamping quality control method based on multi-source data fusion and real-time optimization, characterized in that, The MES system realizes the closed-loop control of the entire production process, including the following steps: Step 1: The die pressure, hydraulic displacement, temperature and workpiece surface image data of the stamping equipment are collected through a distributed sensor network and transmitted to the MES system after processing; Step 2: Extract stamping energy characteristics through time-frequency analysis of pressure data, build a dynamic deformation model based on hydraulic displacement, calculate thermal compensation in combination with temperature, and extract defect characteristics through multi-scale morphological processing of images; Step 3: Normalize the energy characteristics, deformation model, thermal compensation and defect characteristics, generate weight coefficients based on the elastic modulus, and construct a three-dimensional fusion feature space through principal component analysis; Step 4: Input the feature space into the pre-trained LSTM model, associate the stamping speed with the holding time, and output the quality prediction matrix containing the springback deviation and defect probability; Step 5: When the springback deviation exceeds ±0.15mm and the defect probability is greater than 85%, the bus is sent to the PLC to adjust the stamping speed and holding time parameters; Step 6: Calculate the mold compensation amount based on the actual springback amount, link the lubrication system to grade and adjust the injection parameters, and perform micron-level compensation; Step 7: After completing 50 stampings, update the LSTM parameters with laser measurement data, and start incremental learning when the prediction error is greater than 0.1 mm for three consecutive times; Step 8: Build a traceable knowledge base in the MES system and store batch data to form a closed-loop archive.

2. The cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that The distributed sensor network in step 1 includes: A 16-point array piezoelectric pressure sensor group is arranged on the top of the mold cavity, with a sampling frequency of ≥5kHz, and transmits data via the CAN bus protocol; A sub-micron grating displacement sensor installed at the piston rod of the hydraulic cylinder; The working surface of the stamping die is embedded with a 9-point infrared temperature measurement array, and the temperature data is calibrated using the flight time algorithm; Dual-light source polarization industrial cameras are installed on both sides of the conveyor line to synchronously trigger the collection of 0.05mm level defect images on the workpiece surface.

3. A cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that, The step 2 of constructing a dynamic deformation model based on hydraulic displacement includes: Step 2-1: Sliding window filtering is performed on the hydraulic displacement data, and the window length is adaptively adjusted; Step 2-2: Establish a material springback prediction model based on the JC constitutive equation, and calculate the theoretical deformation variable in combination with the elastic modulus data; Step 2-3: Extract the energy spectrum characteristics of the pressure signal through wavelet packet decomposition and divide the low-frequency stable area and the high-frequency impact area; Step 2-4: Perform grayscale projection analysis on the workpiece surface image and optimize the continuity of the defect boundary by combining morphological closing operations.

4. A cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that Generating a weight coefficient based on the elastic modulus in step 3 includes: Step 3-1: Query the elastic modulus value of the current batch of aluminum alloy from the MES system material database; Step 3-2: Calculate the objective weight coefficient of the punching energy feature based on the entropy weight method; Step 3-3: Generate material property correction factor based on elastic modulus; Step 3-4: Reduce the four-dimensional features to three-dimensional space through principal component analysis, retaining more than 95% of the variance information.

5. A cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that The pre-training of the LSTM model in step 4 includes: Step 4-1: Collect historical production data to build a training set and divide the input sequence into sliding time windows; Step 4-2: Introduce the attention mechanism into the network structure to dynamically assign time weights to different features; Step 4-3: Use the Bayesian optimization algorithm to adjust the number of LSTM hidden layer nodes; Step 4-4: Weighted fusion of the real-time stamping speed parameters and the LSTM output gate.

6. The cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, wherein Adjusting the punching speed and holding time parameters in step 5 includes the following sub-steps: Step 5-1: Calculate the stamping speed correction value according to the thermal expansion coefficient of aluminum alloy; Step 5-2: Construct a holding time-temperature correlation curve based on historical process data, and dynamically adjust the holding time parameters using fuzzy control rules; Step 5-3: Send the correction parameters to the PLC controller via the PROFINET industrial bus protocol to ensure that the response delay is less than 50ms; Step 5-4: Generate parameter adjustment records in the MES system interface and associate the equipment operation log with the current process version number.

7. A cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that In step 6, the mold compensation amount is calculated based on the actual springback amount, and the injection parameters are adjusted in stages by linking the lubrication system. The micron-level compensation is performed, which includes the following sub-steps: Step 6-1: Analyze the mold deformation trend based on strain energy density and calculate the residual stress distribution; Step 6-2: The mold fine-tuning mechanism is controlled by a servo motor, with a compensation accuracy of ±2μm; Step 6-3: Lubrication levels are divided according to defect probability: when the defect probability is greater than 90%, the lubricant injection amount is double injection at level I; when the defect probability is between 70% and 90%, the lubricant injection amount is standard injection at level II; when the defect probability is less than 70%, the lubricant injection amount is reduced frequency injection at level III; Step 6-4: Use PID control algorithm to dynamically adjust the lubricant injection pressure to 0.2-0.8MPa.

8. A cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that The incremental learning optimization step in step 7 includes: Step 7-1: Construct a sliding time window to filter key historical data samples, including mold pressure curve, hydraulic displacement time series, temperature monitoring data and defect image samples; Step 7-2: Use Hampel filter to remove abnormal measurement data; Step 7-3: Freeze the weights of the first three layers of the LSTM network and only update the output layer parameters; Step 7-4: Reuse the feature extraction capability of the historical model through transfer learning technology.

9. A cold stamping quality control method based on multi-source data fusion and real-time optimization according to claim 1, characterized in that, The step 8 of constructing a traceable knowledge base in the MES system includes the following sub-steps: Step 8-1: Run the original data stream using the time series database storage device; Step 8-2: Build a knowledge graph to associate material batch number, mold number and quality defect pattern; Step 8-3: Deploy an OPC UA-based data interface to interconnect with the ERP system; Step 8-4: Establish a SQL traceability query engine to support multi-dimensional root cause analysis.

Images