Bending springback meta-learning compensation method integrating machine vision and electrical signal features

Abstract

The present application relates to sheet metal forming technical field, specifically to a kind of bending springback meta-learning compensation method integrated with machine vision and electrical signal characteristics.The method comprises: the multi-modal time series electrical signal in the bending process is synchronously collected;After workpiece unloading, trigger binocular vision system to obtain its three-dimensional point cloud and calculate actual springback angle;From electrical signal, the dynamic process feature related to material plastic deformation is extracted, and is associated with actual springback angle to form sample data pair;The data pair is input into the springback prediction model based on meta-learning framework pre-training, and the model is driven to quickly update parameters;The springback angle of subsequent workpiece in the same batch is predicted using the updated model, and the bending process parameters are automatically compensated accordingly.The present application realizes small sample fast self-adaptation by meta-learning, and combines multi-modal perception to construct accurate "process-result" mapping, which greatly reduces the debugging time and material waste.

Description

Technical Field

[0001] This invention relates to the field of sheet metal forming technology, specifically to a bending springback element learning compensation method that integrates machine vision and electrical signal features. Background Technology

[0002] Sheet metal bending is a key process in metal forming, and its precision directly determines product quality. During bending, the sheet metal experiences springback after unloading due to the release of internal elastic stress, causing a deviation between the forming angle and the target angle. Traditionally, operators rely on experience or repeated trial bending to manually compensate for springback. This method is inefficient, wastes a lot of materials, and is highly dependent on operator experience. With the development of CNC technology, springback prediction and compensation methods based on theoretical formulas or finite element simulations have emerged. However, these methods typically rely on accurate material constitutive models and boundary conditions, making them less adaptable to factors commonly encountered in actual production, such as batch fluctuations in material properties, die wear, and changes in lubrication conditions. Prediction accuracy is difficult to guarantee, especially in flexible production models with multiple varieties and small batches. Re-performing complex simulations or experimental calibrations for each new material and thickness is too costly and uneconomical. In recent years, big data-driven machine learning methods have been introduced to build springback prediction models by training on historical data. However, such methods typically require massive amounts of data covering various working conditions to train a model with strong generalization ability, and they also cannot solve the inherent problem of scarce samples for small-batch new tasks. When the performance of incoming material batches changes, the prediction accuracy of the original model will drop significantly, while collecting enough data from the new batch to retrain the model does not meet the cycle time requirements of rapid production in small batches. Therefore, how to quickly and accurately adapt to the springback characteristics of new batches of materials and achieve precise one-time compensation using only a very small amount of field data (such as the first piece bending test data) under conditions of fluctuating material properties has become a core technical bottleneck that urgently needs to be solved in high-precision, high-flexibility sheet metal bending production.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a bending springback element learning compensation method that integrates machine vision and electrical signal features to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, this invention provides a bending springback element learning compensation method integrating machine vision and electrical signal features, comprising: S1. Acquire multimodal timing electrical signals collected in real time when the bending machine performs the bending process; S2. After unloading and springback, trigger the machine vision system to capture an image of the workpiece bending area, and calculate the actual springback angle based on the image; S3. Extract the dynamic process features related to the plastic deformation of the material from the multimodal time-series electrical signal, and associate them with the actual springback angle to form a sample data pair; S4. Input the sample data into a rebound prediction model pre-trained based on a meta-learning framework to drive the model to quickly update the model parameters. S5. Using the updated springback prediction model, predict and compensate the bending process parameters for subsequent workpieces made of the same batch of materials.

[0006] Specifically, the phrase "acquiring multimodal timing electrical signals collected in real time during the bending process of the bending machine" includes: The hydraulic system pressure signal, slider displacement signal, and angle sensor signal installed on the slider or lower die are synchronously acquired during the movement of the slider of the bending machine. The acquired raw signals are time-aligned and filtered for noise reduction. The key forming stage signal from the slider contacting the workpiece to reaching the bottom dead center and ending the pressure holding is extracted from the complete bending cycle signal. The multimodal timing electrical signal includes at least the pressure-time curve, displacement-time curve, and angle-time curve of the key forming stage.

[0007] Specifically, the step of "extracting dynamic process features related to material plastic deformation from the multimodal time-series electrical signal" includes: The maximum forming force and the integral area of ​​force with displacement are extracted from the pressure-time curve of the key forming stage; the bending angle change rate is extracted from the angle-time curve; the cross-correlation feature of the pressure signal and the angle signal in the time domain is constructed to characterize the coupling relationship between force and deformation; the extracted maximum forming force, integral area, angle change rate and cross-correlation feature are combined to form a multi-dimensional dynamic process feature vector to characterize the deformation behavior of the material during the bending process.

[0008] Specifically, the step of "triggering the machine vision system to capture an image of the workpiece's bending area and calculating the actual springback angle based on the image" includes: A binocular vision sensor fixed above the bending machine's workspace is triggered to capture left and right images of the unloaded workpiece from a direction perpendicular to the bending line. The images are then subjected to distortion correction and stereo matching to identify and locate the three-dimensional spatial point cloud of the sheet metal planes on both sides of the workpiece's bending area. The three-dimensional spatial equations of the two planes are calculated using a plane fitting algorithm, and the dihedral angle between the two planes is then calculated as the actual springback angle. This method can compensate for two-dimensional measurement errors caused by workpiece posture or camera perspective.

[0009] Specifically, the pre-training and update method of the "bounce prediction model pre-trained based on meta-learning framework" is as follows: The springback prediction model employs a neural network that takes the dynamic process feature vector as input and the predicted springback angle as output. Its meta-learning framework includes a meta-learner, which learns initial parameters of a model with good generalization ability using historical bending data containing various materials and thicknesses during the pre-training phase. During the rapid update phase for a specific batch, the sample data pairs described above are used as the support set for that batch. The meta-learner guides the neural network to perform a limited number of gradient iterations, thereby quickly adjusting the model parameters to adapt to the characteristics of the new batch of materials.

[0010] Specifically, the method also includes an optimization step for the initial bending test: When bending the first workpiece of a brand new batch of materials, the initial process parameters based on the process database are used for bending. After completing the visual measurement and rapid model update of the first workpiece, the springback compensation amount predicted by the updated model is fed back to the bending machine CNC system. The system automatically calculates and corrects the target bending angle or bottom dead center depth of subsequent workpieces to compensate for the predicted springback amount, so that qualified angles can be obtained from the second workpiece onwards.

[0011] Specifically, the step of "calculating the actual rebound angle" further includes the detection and processing of asymmetric rebound: After fitting the two planes based on the 3D point cloud, the actual springback angles of the left and right regions of the workpiece bending line are calculated respectively. When the difference between the left and right springback angles exceeds a preset threshold, it is determined that there is asymmetric springback. The left and right springback angles and their difference are used as an additional feature and associated with the dynamic process feature vector for subsequent model updates and compensation decisions, so as to support indirect compensation for the parallelism of the bending machine slider or the workpiece positioning deviation.

[0012] Specifically, the method is executed by an integrated control system, which includes: The signal acquisition module is used to connect to the sensors of the bending machine and acquire the multimodal timing electrical signals in real time. The machine vision measurement module includes an industrial camera, lens, light source, and image processing unit, used to perform image acquisition, processing, and actual rebound angle calculation; The meta-learning model processing module is used to run the pre-trained rebound prediction model and meta-learning update algorithm, and to store historical data. The process control interface module is used to receive the compensation amount output by the model and communicate with the bending machine CNC system to issue corrected process parameter instructions.

[0013] Specifically, the signal acquisition module and the machine vision measurement module coordinate through a synchronization triggering unit: the synchronization triggering unit receives bending cycle status signals from the bending machine CNC system; when it detects that the bending process is completed and the slider returns to the trigger position, it sends an image acquisition trigger signal to the machine vision measurement module; at the same time, this unit provides a unified timestamp reference for the signal acquisition module to ensure the precise correspondence between the electrical signal sequence and the image acquisition time, so as to ensure the accuracy of the temporal correlation of the sample data pairs.

[0014] Specifically, the meta-learning framework employs a model-independent meta-learning algorithm during rapid updates, and its rapid update process is as follows: The sample data generated from the first bending of the current batch of materials is used as the support set; the gradient of the loss with respect to the model parameters is calculated with the loss between the predicted and actual springback angles of the model on the current support set as the optimization objective; instead of directly using this gradient to significantly update the model parameters, the meta-learner generates a parameter update direction and step size based on pre-trained knowledge; the parameter update amount guided by the meta-learner is applied to the neural network model, enabling it to quickly adapt to the characteristics of the current batch of materials while retaining general knowledge, thus completing an incremental update.

[0015] Beneficial effects: Compared with existing technologies, the bending springback element learning compensation method and system integrating machine vision and electrical signal features provided by this invention have the following significant advantages: First, this invention fundamentally changes the adaptation mode for dealing with fluctuations in material properties, achieving a leap from "repeated trial and error" to "learning with one piece, benefiting the whole batch." By introducing a meta-learning framework, the system learns general prior knowledge about springback patterns from rich historical tasks in the offline phase. When facing a new batch of materials, it can use measured data from a single piece or a very small number of workpieces to rapidly update the model parameters, enabling the model to instantly adapt to the new material characteristics. This perfectly meets the core needs of multi-variety, small-batch production, compressing the traditionally time-consuming process debugging process that requires multiple materials into a process completed with only the first workpiece. This significantly shortens production preparation time, substantially reduces trial production material costs, and greatly improves production efficiency and flexibility.

[0016] Secondly, this invention constructs a more robust and accurate "process-result" mapping relationship by deeply integrating multimodal perception information. The method not only acquires the true three-dimensional springback angle after unloading through a high-precision binocular vision system, providing the model with an indisputably high-precision supervision signal; more importantly, it innovatively extracts feature vectors closely related to the physical mechanism of material plastic deformation from the dynamic temporal electrical signals of the bending process, such as maximum forming force, plastic work, deformation rate, and force-deformation cross-correlation. These features profoundly encode the instantaneous response behavior of the material under specific processes, enabling the meta-learning model to establish an interpretable and robust causal relationship from the physically rich dynamic process to the final static result, rather than simple data fitting. Therefore, when facing new batches of materials, even with very few samples, it can make more physically reasonable predictions and inferences, significantly improving the model's generalization ability and prediction accuracy.

[0017] Finally, this invention constructs a complete "perception-decision-execution" intelligent closed-loop system, promoting the automation and intelligent upgrading of the bending process. The system seamlessly integrates high-frequency signal acquisition, high-precision machine vision measurement, intelligent meta-learning algorithms, and the bending machine's CNC system, achieving full-process automation from automatic data acquisition and rapid model adaptation to automatic compensation and distribution of process parameters. This not only reduces reliance on operator experience and minimizes human error but also transforms the experience and knowledge of skilled workers into replicable and iterative digital models. Furthermore, the system's process status perception and feature extraction capabilities provide a valuable data foundation for further process optimization, quality traceability, and even equipment health prediction, laying a key technological foundation for realizing truly intelligent factories and flexible manufacturing units. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the bending springback element learning compensation method integrating machine vision and electrical signal features provided in a specific embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0020] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0021] Please see Figure 1 This invention provides a bending springback element learning compensation method integrating machine vision and electrical signal features, comprising: S1. Acquire multimodal timing electrical signals collected in real time when the bending machine performs the bending process.

[0022] It should be further explained that this step is fundamental to achieving process awareness, aiming to synchronously capture key physical quantities reflecting material deformation behavior and machine status during bending. In implementation, three types of sensors need to be installed on the bending machine: a high-frequency response dynamic pressure sensor (monitoring the pressure of the main cylinder or working oil chamber), a high-precision linear displacement sensor (such as a magnetic scale, monitoring the slider position), and a high-response tilt sensor (monitoring the real-time angle of the upper die or slider). These sensors are connected to a multi-channel synchronous data acquisition card via shielded cables. The acquisition card is activated by a digital trigger signal issued by the bending machine's CNC system at the start of the bending action, ensuring that all signal acquisitions have a unified time starting point. The acquisition frequency needs to be high enough, typically not lower than 500Hz, to capture deformation dynamics. The raw signal is first low-pass filtered (a cutoff frequency of 100Hz is recommended) to suppress electrical noise. Then, a signal segment representing the "critical forming stage" is precisely extracted from the complete signal cycle, defined as the moment the slider contacts the workpiece (…). The timeframe is determined by a combination of the pressure signal at the inflection point where the no-load baseline rises significantly and the sudden change in the displacement rate, until the slider reaches the preset bottom dead center and begins its return stroke after completing the pressure holding process. End. The acquired multimodal time-series electrical signals must include at least the pressure-time series within that time period. Displacement-time series and angle-time series .

[0023] Understandably, synchronously and with high fidelity acquiring multi-source electrical signals during this stage lays the foundation for subsequent extraction of features that characterize the essence of the material's plastic deformation process. This dynamic information is far more reflective of the material's constitutive properties than a single final pressure or angle value.

[0024] S2. After unloading and springback, trigger the machine vision system to capture an image of the workpiece bending area, and calculate the actual springback angle based on the image.

[0025] It should be further explained that this step aims to acquire the actual springback angle after bending with high precision and non-contact, serving as a supervisory signal for model learning. In implementation, a rigorously calibrated binocular stereo vision system is used. Two industrial cameras are fixedly mounted above the bending machine's workspace, their optical axes approximately perpendicular to the vertical plane containing the bending line, and equipped with bar light sources to enhance the contrast of the bending area. The triggering timing is crucial: when the bending action is completed and the slider returns to a preset safe height (ensuring the workpiece is completely detached from the lower die and the field of view is unobstructed), the CNC system or a dedicated synchronous trigger unit sends a hardware trigger signal to the vision system, driving the binocular cameras to synchronously expose and acquire images. Subsequently, 3D reconstruction is performed: distortion correction and stereo matching are performed on the acquired left and right images to generate a disparity map. Then, the three-dimensional point cloud of the part of interest in the bending region is calculated using the principle of triangulation. ,in , Focal length (in pixels). Baseline distance Let be the coordinates of the object point. Next, the RANSAC algorithm is used to robustly segment and fit the sheet metal planes on both sides of the workpiece's bending region from the point cloud, obtaining the equation of the left plane. and the equation of the right-hand plane And calculate its unit normal vector. and Ultimately, the actual rebound angle We obtain the angle between the two normal vectors by calculating: ,in This represents the dot product of vectors.

[0026] Understandably, this 3D vision measurement method directly acquires spatial geometric relationships, eliminating errors caused by perspective and workpiece posture in 2D measurement. It can provide stable and accurate (typically ±0.05°) springback angle true values, providing reliable training labels for the model.

[0027] S3. Extract the dynamic process features related to the plastic deformation of the material from the multimodal time-series electrical signal, and associate them with the actual springback angle to form a sample data pair.

[0028] It should be further explained that this step aims to extract information-dense, physically meaningful, low-dimensional feature vectors from the original time-series signal. This is specifically for the extracted key forming stage signals. ,in The following features are extracted: (1) Maximum forming force This reflects the peak value of the material's deformation resistance.

[0029] (2) Force-displacement integral area In discrete form, it can be calculated as Its physical meaning is the plastic work done on the workpiece during the bending process, which is related to the material hardening behavior.

[0030] (3) Average rate of change of angle It reflects the average deformation rate of a material under load.

[0031] (4) Peak value of pressure-angle cross-correlation ,in A value of 0.1 seconds can be used; this feature characterizes the degree of synchronous coupling between "force" and "deformation" in time. These four features are combined into a four-dimensional dynamic process feature vector. Finally, this feature vector The actual rebound angle calculated in step S2 Combined, they form a complete sample data pair. .

[0032] Understandably, the extracted features are not simple statistics, but descriptors closely related to the physical mechanism of material plastic deformation. They collectively encode key information of a single bending process, establishing an interpretable mapping bridge from dynamic process to static result for the meta-learning model.

[0033] S4. Input the sample data into a rebound prediction model pre-trained based on a meta-learning framework to drive the model to quickly update the model parameters.

[0034] It should be further explained that this step is the core of achieving rapid adaptation with a small sample size (usually 1). The rebound prediction model is a multilayer perceptron. Its input is the dynamic process feature vector. The output is the predicted rebound angle. The meta-learning framework employs a model-independent meta-learning algorithm, which utilizes a historical bending task dataset covering various materials and thicknesses during the pre-training (meta-training) phase to optimize model initialization parameters with strong generalization capabilities. When faced with a new batch of materials (a new task), the rapid update process is as follows: The sample data generated by the first piece of the current batch is compared with... This serves as the support set for the task. First, the computational model is set with the current parameters. (initialized to) The prediction loss for that sample, such as mean squared error. Next, the gradient of the loss function with respect to the model parameters is calculated. Then, instead of performing large-scale iterative training, a single (or a few) gradient descent update is performed: Among them, the inner loop learning rate It is a small positive number (e.g., 0.01) learned during the meta-training phase, which controls the adjustment step size when the model adapts to new tasks. This process is called the "inner loop" update.

[0035] Understandably, the key to meta-learning lies in the initialization parameters of its pre-training. The state is optimized to be sensitive to the task distribution, allowing for small gradient updates based on individual samples. This will enable the model parameters It enables rapid adaptation to the characteristics of new batches of materials while retaining general knowledge learned from historical data, thus achieving "one-step learning".

[0036] S5. Using the updated springback prediction model, predict and compensate the bending process parameters for subsequent workpieces made of the same batch of materials.

[0037] It should be further explained that this step transforms learning outcomes into practical productivity. It yields a model that is rapidly updated for new batches. Subsequently, for any subsequent workpiece in this batch, before its bending, the system collects the electrical signals of its bending process in real time (step S1) and extracts the dynamic process feature vector. (Step S3). Input the feature vector into the updated model. This allows us to predict the springback angle that the workpiece will produce under the current set process parameters. According to the principle of springback compensation, in order to obtain the target bending angle... The forming angle that the mold needs to achieve when closing should be: The CNC system of the bending machine determines the target forming angle based on the geometric relationship of the die. Converted into actual control commands, as shown below: dead point depth setting value This compensation value can be automatically sent to the bending machine CNC via the process control interface module. Starting with the second workpiece in this batch, the bending will be performed using this compensated process parameters, thus theoretically making the angle of the workpiece after unloading and springback close to the target value. .

[0038] Understandably, this step forms a closed-loop control system of "perception-prediction-compensation". This allows subsequent workpieces to avoid the "trial bending-measurement-manual adjustment" cycle after the initial trial bending and rapid model update, thus significantly reducing debugging time and material waste, and enabling rapid and precise operations in small-batch production.

[0039] Furthermore, the core of this invention lies in constructing a closed loop of "perception-learning-compensation." The "perception" stage integrates the dynamic electrical signals of the bending process and visual measurements of the resulting state. The "learning" stage employs a meta-learning framework, the core idea of ​​which is "learning to learn," meaning that during the pre-training phase (meta-training), the model learns general prior knowledge about springback patterns from numerous historical bending tasks involving different materials and thicknesses. When faced with a new batch of materials (i.e., a new task), the model can utilize this general prior knowledge, combined with a very small number of samples specific to that batch of materials (such as measurement data from the first workpiece), to quickly adapt and update the model parameters. This allows the system to adapt to the characteristic fluctuations of new material batches with minimal trial bending. The "compensation" stage converts the updated model's predicted output into process parameter correction instructions executable by the bending machine, such as the lower dead center depth compensation value. Or target angle compensation value The purpose of this method is to use the "measurement-learning" cost of a single bend (first piece) to improve the bending accuracy of all subsequent workpieces in the same batch, thereby fundamentally solving the bottleneck of debugging efficiency in multi-variety, small-batch production.

[0040] Understandably, this method deeply integrates online monitoring (electrical signals), high-precision result measurement (machine vision), and efficient few-shot learning algorithms (meta-learning) to create a dedicated rapid adaptation system for the sheet metal bending springback problem. Its beneficial effects lie in overcoming the limitations of traditional fixed formula-based or big data-driven models in dealing with material performance fluctuations, achieving "one-time learning, batch benefit," significantly reducing the number of trial bends, debugging time, and material waste. It is particularly suitable for the production scenarios of small-batch, multi-variety, high-precision sheet metal parts in aerospace, high-end equipment, and other fields.

[0041] Specifically, the "acquiring multimodal timing electrical signals collected in real time during the bending process of the bending machine" includes: The hydraulic system pressure signal, slider displacement signal, and angle sensor signal installed on the slider or lower die are synchronously acquired during the movement of the slider of the bending machine. The acquired raw signals are time-aligned and filtered for noise reduction. The key forming stage signal from the slider contacting the workpiece to reaching the bottom dead center and ending the pressure holding is extracted from the complete bending cycle signal. The multimodal timing electrical signal includes at least the pressure-time curve, displacement-time curve, and angle-time curve of the key forming stage.

[0042] It should be further noted that the integrity and synchronization of signal acquisition are fundamental to subsequent feature extraction. The implementation plan for this step specifically includes: (1) Sensor Configuration and Signal Synchronization: The hydraulic system pressure signal is obtained through a high-frequency response pressure sensor installed in the main cylinder inlet or working oil chamber. The range must cover the maximum working pressure of the bending machine, and the sampling frequency must be no less than 500Hz. The slider displacement signal is obtained through a high-precision magnetic grating ruler or optical grating ruler installed on the slider. The resolution is usually better than 1μm. The angle sensor is preferably a high-precision tilt sensor with good dynamic response performance, which is directly installed on the slider or a component rigidly connected to the lower die to directly reflect the real-time angle change of the die. All sensor signals are synchronously acquired through a multi-channel synchronous data acquisition card. The acquisition card is started by a unified trigger signal issued by the bending machine CNC system to ensure that there is a unified time reference between the signals and to eliminate the phase error caused by asynchronous acquisition.

[0043] (2) Signal preprocessing: The original signal is first low-pass filtered to suppress high-frequency noise. The cutoff frequency is set according to the working frequency of the bending machine, for example, it can be set to 100Hz. Since there may be a slight fixed delay in the signal transmission of different sensors, it is necessary to perform a slight time shift compensation on the signals of each channel through calibration experiments (such as when the slider contacts the workpiece, the pressure signal and displacement signal will show a characteristic inflection point) to achieve accurate timing alignment.

[0044] (3) Key forming stage segmentation: Define the starting moment of "slider contact with workpiece". This moment can be determined by monitoring the inflection point where the pressure signal begins to rise significantly from the no-load pressure value (close to 0), and by combining this with the abrupt change point of the displacement signal's rate of change. The moment of "end of pressure holding at the bottom dead center" is defined. This moment occurs when the slider reaches the preset bottom dead center depth, the pressure holding timer ends, and the slider begins its return journey. (Time window captured) This segment of signals records the entire process of material deformation from elastic deformation and plastic deformation to the end of forming, containing the most critical information that determines springback. The holding time is usually set between 0.5s and 2s to ensure that the material stress relaxation tends to stabilize. This time range is determined based on the creep relaxation characteristics of most metal materials. If the time is too short, the stress relaxation will be insufficient, and if it is too long, it will affect production efficiency.

[0045] Understandably, the precise synchronous acquisition and key stage extraction of multi-source electrical signals provide a high-quality data foundation for the subsequent construction of feature vectors that reflect the instantaneous deformation state of materials. Compared to acquiring only the final state signal (such as dead-point pressure), the time-series signal contains dynamic process information of material deformation and has a stronger physical correlation with the springback result.

[0046] Specifically, the "extraction of dynamic process features related to material plastic deformation in the multimodal time-series electrical signals" includes: extracting the maximum forming force and the integral area of ​​force with displacement from the pressure-time curve of the key forming stage; extracting the bending angle change rate from the angle-time curve; constructing the cross-correlation features of the pressure signal and the angle signal in the time domain to characterize the coupling relationship between force and deformation; and combining the extracted maximum forming force, integral area, angle change rate, and cross-correlation features to form a multidimensional dynamic process feature vector to characterize the deformation behavior of the material during the bending process.

[0047] It should be further explained that the goal of feature design is to transform the original time-series signal into a low-dimensional and information-rich descriptor that can characterize the material's mechanical properties and deformation mechanism. The specific calculations and physical meanings of each feature are as follows: (a) Maximum forming force From the pressure-time curve Find the maximum value in the middle. ,in This force is directly related to the material's yield strength, tensile strength, thickness, and bending length, and is a core indicator reflecting the material's resistance to deformation.

[0048] (b) Force-displacement integral area Calculate the pressure signal For displacement signal The points, In practical discrete signal processing, the trapezoidal rule can be used to calculate: ,in This represents the number of sampling points. The physical meaning of bending is the plastic work done on the workpiece during the bending process. It comprehensively reflects the material's ability to absorb energy during deformation and is closely related to the material's hardening behavior.

[0049] (c) Rate of change of bending angle Calculate the angle-time curve In the critical forming stage The average rate of change within, This characteristic reflects the deformation rate of a material under a given pressure, and is indirectly related to the material's fluidity or viscoplasticity. For the same process, a softer material may reach the target angle in a shorter time, i.e. It could be even bigger.

[0050] (d) Pressure-angle cross-correlation characteristics Calculate the pressure signal With angle signal Cross-correlation function in the time domain Near zero latency (e.g.) The maximum value of ). This feature characterizes the degree of coupling between "force" and "deformation" over time. In ideal elastoplastic deformation, force and deformation should be highly synchronized, with a high cross-correlation value. If the material exhibits significant viscous effects or abnormal mold friction, the force and deformation responses may become asynchronous, reducing this feature value. Therefore, this feature can be used to detect anomalies in material properties or process conditions.

[0051] Finally, the dynamic process feature vector can be represented as This is a 4-dimensional vector that contains both key physical quantities and dynamic characteristics of the process, providing an information-dense input for the meta-learning model.

[0052] Understandably, the extracted dynamic process feature vectors are not simple signal statistics, but descriptors closely linked to the physical mechanisms of material plastic deformation. Maximum forming force and plastic work are directly related to material strength and energy consumption; the rate of change of angle is related to strain rate sensitivity; and cross-correlation features relate to the dynamic response relationship between force and deformation. These features collectively provide the meta-learning model with physically rich clues needed to infer the springback angle, enabling the model to establish a robust mapping from process to outcome, rather than a "black box" fit, thus improving the model's interpretability and generalization ability.

[0053] Specifically, the step of "triggering the machine vision system to capture images of the workpiece bending area and calculating the actual springback angle based on the images" includes: triggering a binocular vision sensor fixed above the bending machine's workspace to capture left and right images of the unloaded workpiece from a direction perpendicular to the bending line; performing distortion correction and stereo matching on the images to identify and locate the three-dimensional spatial point cloud of the sheet metal planes on both sides of the workpiece bending area; calculating the three-dimensional spatial equations of the two planes using a plane fitting algorithm, and then calculating the dihedral angle between the two planes as the actual springback angle; this method can compensate for two-dimensional measurement errors caused by workpiece posture or camera perspective.

[0054] It should be further explained that visual measurement is used to obtain high-precision true rebound angle. The key is... The specific implementation steps are as follows: 1) System calibration: First, perform high-precision calibration on the binocular vision system to obtain the intrinsic parameter matrices of the left and right cameras. Distortion coefficient And the extrinsic parameter matrix (rotation matrix) describing the relative positional relationship between the left and right cameras. Translation vector A high-precision checkerboard calibration plate was used, and the calibration was completed using the Zhang Zhengyou calibration method.

[0055] 2) Image Acquisition and Triggering: Two industrial cameras are symmetrically mounted above the bending machine, with their optical axes approximately perpendicular to the vertical plane containing the bending line. Once the workpiece bending is complete and the slider returns to a safe position (e.g., the upper limit), the bending machine's CNC system sends a trigger signal to the vision measurement module. The vision module controls the two cameras to simultaneously capture images of the workpiece's bending area. To ensure image quality, a suitable strip light source is required to illuminate the bending area from the side, enhancing the contrast between the bending edges and the side surfaces.

[0056] 3) Image Processing and 3D Reconstruction: Distortion correction is performed on the captured left and right images. Subsequently, a disparity map is calculated using a stereo matching algorithm (such as semi-global matching, SGM). According to the principle of binocular vision, the midpoint of space 3D coordinates It can be calculated using the following formula: , , in, The focal length of the camera (in pixels). The baseline distance of the binocular camera. Principal point coordinates These are the pixel coordinates in the left image. This step converts the pixels of the region of interest (ROI) within the bend into a 3D point cloud. .

[0057] 4) Point cloud segmentation and plane fitting: In the 3D point cloud, based on region growing or RANSAC algorithm, the point cloud belonging to the left and right planes of the workpiece is segmented separately to obtain two point cloud sets. and To each and Plane fitting is performed using the least squares method. Let the equation of the left-hand plane be... Its normal vector is The equation of the plane on the right is Its normal vector is The normal vector needs to be normalized.

[0058] 5) Springback angle calculation: Actual springback angle That is, two plane normal vectors and The angle between them. The calculation formula is: Since the normal vector has been normalized, the calculation simplifies to the inverse cosine of the dot product. This method calculates the true dihedral angle in three-dimensional space, unaffected by whether the workpiece is absolutely horizontal in the field of view or by the camera's angle of view. It offers high measurement accuracy, good repeatability, and a typical accuracy of ±0.05°.

[0059] Understandably, the biggest advantage of using binocular stereo vision technology for 3D measurement compared to monocular vision or 2D projection measurement is its ability to directly obtain the 3D geometric information of the bending area, thereby calculating the true spatial angle and eliminating systematic errors caused by workpiece posture, lens perspective distortion, etc. This provides accurate and reliable supervision signals (i.e., labels) for the meta-learning model, which is a prerequisite for the model to learn effectively and converge quickly. High-precision measurement results are the cornerstone for achieving high-precision compensation.

[0060] Specifically, the pre-training and updating method of the "springback prediction model based on meta-learning framework pre-training" is as follows: the springback prediction model adopts a neural network with the dynamic process feature vector as input and the predicted springback angle as output; its meta-learning framework includes a meta-learner, which learns a model initial parameter with good generalization ability using historical bending data containing multiple materials and thicknesses during the pre-training stage; in the rapid update stage for a specific batch, the sample data pairs are used as the support set for that batch, and the meta-learner guides the neural network to perform a limited number of gradient iterations, thereby quickly adjusting the model parameters to adapt to the characteristics of the new batch of materials.

[0061] It should be further noted that the specific implementation of the meta-learning framework preferably employs a model-independent meta-learning algorithm. The entire system is divided into two phases: offline "meta-training" and online "meta-testing" (rapid adaptation). The specific design includes: ① Meta-training phase (offline): Objective: To learn a general model initialization parameter from a large number of historical bending tasks. This allows the model to quickly adjust (i.e. learn) when faced with a new task (a new batch of materials) using a small number of samples from that batch.

[0062] Furthermore, in the data preparation phase, data collection... A dataset of historical bending tasks Each task Represents a specific combination of materials / thicknesses, and its dataset It contains multiple samples, each of which is a feature vector of a dynamic process. and the corresponding true rebound angle (Obtained by visual measurement).

[0063] Furthermore, regarding the model structure, the specific design of this invention includes: Rebound prediction model It could be a simple multilayer feedforward neural network (MLP), for example, with two hidden layers, each with 32 neurons, using the ReLU activation function. The number of nodes in the input layer is equal to the dimension of the dynamic feature vector (e.g., 4), and the output layer has one node (predicting the bounce angle). Parameters This represents all the weights and biases of the neural network. Training process: The MAML (Model-Agnostic Meta-Learning) algorithm or a variant thereof is used.

[0064] Furthermore, in each meta-training iteration: (1) Randomly sample a batch of tasks .

[0065] (2) For each sampled task From its dataset Randomly sample a support set (Typically very small, e.g., K=1 or 5 samples, simulating a small sample scenario) and a query set. Used for evaluation.

[0066] (3) Using the current model parameters In the mission Support set Perform one or more steps of gradient descent to obtain task-specific parameters. For example, a single gradient update: .in, The learning rate within the task (e.g., 0.01). Let be the mean squared error loss function.

[0067] (4) Calculate task-specific parameters In the task query set Losses .

[0068] (5) For all sampling tasks, sum the query loss from step (4) and calculate the loss against the original parameters. The gradient (i.e., the second gradient) is then used to update the original parameters. : .in, The meta-learning rate is (e.g., 0.001). Through repeated iterations on numerous tasks, the model parameters... It is optimized to a state where, when faced with any new task, only a very small number of samples from that new task are needed for one or a few gradient updates to obtain model parameters that perform well on that new task. After this training phase is completed, the optimal meta-initialization parameters are saved. .

[0069] ②Rapid update phase (online, for new batches): When the incoming material is a new batch, the model initialization parameters learned during the meta-training phase are used. As initial parameters for the springback prediction model, the first workpiece in this batch is bent. After bending, the system automatically acquires electrical signals and calculates the dynamic process feature vector. The actual rebound angle was obtained through visual measurement. This constitutes the first support set sample for the new task. This single sample (or several accumulated samples) serves as the support set. Following the method of step (3) in the meta-training phase, using Perform gradient descent in one or a few steps (e.g., 1-5 steps) to quickly update the model parameters adapted to the new batch of materials. This process involves minimal computation and takes extremely short time (seconds), and can be completed online. The updated model will then be... Used to predict the springback angle of subsequent workpieces in this batch, and then to make compensation.

[0070] Understandably, the application of the meta-learning framework is the key innovation of this invention. Traditional machine learning models require collecting large amounts of data for retraining each new material, which is impractical in multi-variety, small-batch production. In contrast, the meta-learning model, having already gained extensive experience during the pre-training stage, has learned general rebound patterns extracted from diverse historical tasks. When faced with new materials, it acts like an experienced craftsman, quickly adjusting its "experienced judgment" based on the "trial bending" results of the first workpiece, achieving rapid adaptation. This solves the model generalization problem under small sample sizes, enabling the high-precision rebound prediction model to be practically applied in dynamically changing production environments.

[0071] Specifically, the method also includes an optimization step for the first trial bending: when bending the first workpiece of a brand new batch of materials, the initial process parameters based on the process database are used for bending; after completing the visual measurement and rapid model update of the first workpiece, the springback compensation amount predicted by the updated model is fed back to the bending machine CNC system, and the system automatically calculates and corrects the target bending angle or bottom dead center depth of subsequent workpieces to compensate for the predicted springback amount, so as to obtain a qualified angle from the second workpiece onwards.

[0072] It should be further explained that this step describes a seamless process from "first-piece trial bending" to "mass production": First, the process database stores successful process parameters for various materials, thicknesses, bending angles, and die combinations produced in the past. For a new batch, the system searches the database for the most similar historical records based on its material grade, thickness, target angle, and other information, and then applies its process parameters (primarily the target bottom dead center depth). The initial parameters are used for the first-piece bending. After the first-piece bending is completed, the system immediately executes the aforementioned visual measurement and rapid model update process. The updated model... Used to predict the theoretical springback angle of this batch of material under current process parameters. .

[0073] Further, compensation calculations are performed: Assuming the target bending angle is Therefore, the required forming angle (i.e., the angle when the mold is closed) should be: The bending machine control system, based on the die geometry, will... Convert to new target bottom dead center depth Alternatively, you can directly calculate the depth compensation value. Parameter distribution: The calculated parameters will be sent to the relevant departments. or The data is automatically sent to the bending machine's CNC system to update the bending program. Starting with the second workpiece in this batch, the corrected process parameters are used for bending, and theoretically, the springback angle should be close to the target value. After bending the second workpiece, its actual springback angle can be measured again to verify the compensation effect, and this new data can be used to... By adding a support set, the model can be further fine-tuned to achieve continuous optimization.

[0074] Understandably, this step represents a shift from the traditional manual debugging model of "trial and error-correction" to an automated intelligent compensation model of "perception-learning-decision-making." Its core value lies in compressing the debugging process—which previously required wasting multiple pieces of material and repeatedly adjusting process parameters—into a process completed with only the first piece. This not only saves materials and debugging time but also achieves automated and intelligent upgrades to the production process, making it particularly suitable for the deployment of flexible manufacturing cells (FMC) or smart factories.

[0075] Specifically, the step of "calculating the actual springback angle" further includes the detection and processing of asymmetric springback: after fitting the two planes based on the three-dimensional point cloud, the actual springback angles of the left and right regions of the workpiece bending line are calculated respectively; when the difference between the left and right springback angles exceeds a preset threshold, it is determined that asymmetric springback exists; the left and right springback angles and their difference are used as an additional feature and associated with the dynamic process feature vector for subsequent model updates and compensation decisions, so as to support indirect compensation for the parallelism of the bending machine slider or the workpiece positioning deviation.

[0076] It should be further explained that in actual production, due to factors such as synchronization errors of the left and right hydraulic cylinders of the bending machine slide, uneven die wear, or inaccurate workpiece positioning, the springback angles on both sides of the workpiece may be inconsistent, i.e., asymmetrical springback. This method enhances the ability to detect and handle this phenomenon: after fitting the overall left and right planes using 3D point clouds, the left point cloud is further processed... Divide into left front region along the bend line (Y-axis). and left rear area Similarly, the right-side point cloud Divided into the right anterior region and right rear area To each and Fit the plane and calculate the angle between the normal vectors of the two planes as the local rebound angle on the left. Similarly, calculate the local rebound angle on the right side. Define asymmetry. Set an asymmetry threshold. ,For example This threshold is set based on the tolerances allowed by the actual process and the repeatability accuracy of the measurement system, and is used to distinguish between normal measurement noise and genuine asymmetry. When When significant asymmetric springback is detected, the workpiece is determined to have such a condition. , as well as As additional information.

[0077] Furthermore, when constructing sample data pairs for updating the meta-learning model, not only are dynamic process feature vectors associated... and overall rebound angle Furthermore, asymmetric information is also associated. During the learning process, the model may gradually associate certain dynamic process characteristic patterns (such as the imbalance of left and right hydraulic pressure signals) with asymmetric springback. During predictive compensation, if the model predicts asymmetric risk, in addition to providing the overall angle compensation value, the system can also issue a warning to the operator, prompting them to check the equipment status or workpiece positioning. A more advanced compensation strategy, on high-end bending machines with independent left and right closed-loop control, can attempt to provide differentiated fine-tuning values ​​for the bottom dead center depth on both sides to actively correct asymmetric springback.

[0078] Understandably, this extended functionality enhances the system's adaptability to complex operating conditions and its process insight. It not only focuses on the overall springback amount but also diagnoses potential left-right imbalances during the bending process. Incorporating asymmetric information into the learning framework allows the model to learn richer process knowledge and even indirectly perceive equipment status to some extent. This improves the system's robustness and practicality, extending its value from simple springback compensation to process quality monitoring and preliminary diagnosis.

[0079] Specifically, the method is executed by an integrated control system, which includes: The signal acquisition module is used to connect to the sensors of the bending machine and acquire the multimodal timing electrical signals in real time. The machine vision measurement module includes an industrial camera, lens, light source, and image processing unit, used to perform image acquisition, processing, and actual rebound angle calculation; The meta-learning model processing module is used to run the pre-trained rebound prediction model and meta-learning update algorithm, and to store historical data. The process control interface module is used to receive the compensation amount output by the model and communicate with the bending machine CNC system to issue corrected process parameter instructions.

[0080] It should be further noted that this integrated control system is a hardware and software entity, typically deployed on an industrial computer or embedded industrial PC. The detailed structure and functions of each module are as follows: (1) Signal Acquisition Module: This module consists of a multi-channel synchronous data acquisition card and its driver software. Hardware-wise, it connects to the signal output terminals of the pressure sensor, magnetic scale / optical scale, and angle sensor of the bending machine via shielded cables. Software-wise, it configures the range, sampling frequency (e.g., 1kHz), and trigger mode (usually using the "bending start" signal from the bending machine's CNC as an external trigger). This module is responsible for recording the original electrical signal with high fidelity during each bending cycle and performing preliminary time-scale alignment and buffering.

[0081] (2) Machine Vision Measurement Module: The hardware includes two identical industrial cameras (such as 5-megapixel CMOS cameras), lenses with appropriate focal lengths, LED strip light sources, and light source controllers. The cameras are connected to the industrial computer via Gigabit Ethernet or USB 3.0 interfaces and fixed on a rigid bracket to ensure relative stability. The software includes a camera SDK, an image processing library (such as OpenCV), and a dedicated 3D measurement algorithm library. This module receives measurement trigger signals from the process control interface module, controls the cameras to take pictures synchronously, and executes the entire process from image acquisition and 3D reconstruction to springback angle calculation, ultimately outputting a high-precision actual springback angle value.

[0082] (3) Meta-learning model processing module: This is the "intelligent core" of the system, existing in software form. It contains a model management submodule, which is responsible for loading the model initialization parameter file obtained by pre-completion meta-training on the server. An online learning engine executes the MAML fast adaptation algorithm (i.e., inner loop update) when new data is received. A database stores the feature vectors, measurement results, model parameter versions, and compensation records for each bending task. This module is typically implemented using a Python deep learning framework (such as PyTorch or TensorFlow) and provides an API for other modules to call.

[0083] (4) Process Control Interface Module: This module is responsible for bidirectional communication with the bending machine's CNC system. On one hand, it receives bending status signals (such as process start / end) and current process parameters from the CNC system. On the other hand, it calculates the process parameter correction amount (such as the dead point compensation value). The data is sent to the CNC system according to a predefined protocol (such as Modbus TCP / IP or OPCUA). This module also needs to implement a human-machine interface (HMI) to display measurement results, model status, compensation suggestions, and allow the operator to confirm or fine-tune them.

[0084] Understandably, this integrated control system seamlessly connects the perception, decision-making, and execution stages, forming a complete intelligent control closed loop. It tightly integrates advanced machine learning algorithms with mature industrial automation components (sensors, vision systems, PLC / CNC), significantly improving the process adaptability and intelligence level of the bending machine through "empowerment" without altering its main structure. The modular design makes the system easy to deploy, maintain, and upgrade, providing a feasible solution for the digital and intelligent transformation of sheet metal bending processes.

[0085] Specifically, the signal acquisition module and the machine vision measurement module coordinate through a synchronization triggering unit: the synchronization triggering unit receives bending cycle status signals from the bending machine CNC system; when it detects that the bending process is completed and the slider returns to the trigger position, it sends an image acquisition trigger signal to the machine vision measurement module; at the same time, this unit provides a unified timestamp reference for the signal acquisition module to ensure the precise correspondence between the electrical signal sequence and the image acquisition time, so as to ensure the accuracy of the temporal correlation of the sample data pairs.

[0086] It is important to further explain that precise synchronization is crucial to ensuring that the "process signals" and "result measurements" are correctly correlated to the same bending action. The synchronization trigger unit is typically implemented using both hardware triggering circuitry and software logic. On the hardware side, the bending machine's CNC system provides digital output signals (DOs) reflecting key nodes in the bending cycle, such as "bending start," "bottom dead center reached," "slider return," and "top dead center reached." The synchronization trigger unit acquires these signals via a digital input card (DI). On the software side, the synchronization trigger unit runs a state machine that monitors the bending cycle in real time. Its core logic is as follows: When the "slider return" signal is detected as valid, and the encoder feedback indicates that the slider position has returned to the preset safe imaging height (this height must ensure the workpiece is completely detached from the lower mold and the field of view is unobstructed), a hardware trigger pulse signal is immediately sent to the two industrial cameras via the digital output card (DO) to drive the cameras to synchronously expose and acquire images. This triggering timing is crucial; it must be performed after the rebound has stabilized (usually 1-2 seconds after unloading) and must avoid vibrations caused by the slider movement. For the signal acquisition module, acquisition is typically triggered by the "bending start" signal and continues for a sufficient time to cover the entire bending cycle. The synchronous trigger unit records a high-precision timestamp simultaneously with the image acquisition trigger signal. The signal acquisition module also records the timestamp of each sampling point when recording data. In subsequent data processing, this unified time base can be used to determine the image acquisition time. This corresponds to a specific time point in the electrical signal sequence, thereby ensuring the dynamic process feature vector extracted from that bending process. The actual springback angle calculated from the image taken from the bending result. These data are strictly derived from the same physical event (the same bend of the same workpiece), forming an accurate and valid sample data pair. Used for model learning.

[0087] Understandably, the synchronization triggering unit solves the time alignment problem of multimodal data, which is a prerequisite for effective data fusion. Without precise synchronization, electrical signal characteristics and visual measurement results may correspond to different workpieces or different bending times, leading to data mismatch. This causes the model to learn incorrect correlations, severely affecting learning effectiveness and compensation accuracy. This design ensures the accuracy of data correlation and the reliability of the system through hardware triggering and a unified time base.

[0088] Specifically, the meta-learning framework employs a model-independent meta-learning algorithm during rapid updates. The rapid update process involves: using the sample data pairs generated from the first bending of the current batch of materials as the support set; calculating the gradient of the loss with respect to the model parameters, with the loss between the predicted and actual springback angles on the current support set as the optimization objective; instead of directly using this gradient to significantly update the model parameters, the meta-learner generates a parameter update direction and step size based on pre-trained knowledge; and applying the parameter update amount guided by the meta-learner to the neural network model, enabling it to quickly adapt to the characteristics of the current batch of materials while retaining general knowledge, thus completing an incremental update.

[0089] It should be further explained that, using the MAML algorithm as an example, the specific mathematical steps of fast update (inner loop update) are detailed here. It is assumed that the optimal meta-initialization parameters have been obtained through meta-training. Let the first piece sample data of the new batch of materials be the support set. The model is The loss function is the mean squared error (MSE). The specific design includes: The first step is to calculate the loss function with respect to the model parameters. At the initial point gradient at : This gradient The direction indicates how to change Only then can the model In this specific sample The prediction error becomes smaller.

[0090] The second step is to apply the meta-learning update rule. In MAML, this gradient is directly used, multiplied by a small inner loop learning rate. To update the parameters: Among them, the inner loop learning rate It is a hyperparameter that is determined during the meta-training phase (e.g., Its smaller value ensures that updates are "minor adjustments" rather than "retraining," thus preserving general knowledge learned from a large number of historical tasks. .this This is a concrete manifestation of "the meta-learner updating the direction and step size based on parameters generated from pre-trained knowledge": the direction is determined by the gradient. The decision is made that the step size is determined by... control. The value is obtained through meta-training on a large number of tasks, and it represents the most reasonable "adjustment strength" when adapting to new tasks.

[0091] Step 3, updated parameters This means that the model parameters have been adapted to the current batch of materials. These parameters are then used to predict subsequent workpieces. If necessary, after obtaining samples of the 2nd and 3rd workpieces, the support set can be expanded, and the above process can be repeated for multi-step gradient updates to achieve incremental optimization. Typically, 1-5 updates are sufficient to achieve good results, demonstrating the characteristic of "rapid" adaptation.

[0092] Understandably, the rapid update mechanism of the MAML algorithm is the core of its efficiency. Unlike traditional transfer learning, which fixes low-level features and only adjusts the top level, it allows gradient-based adjustments to all model parameters. Crucially, the meta-training process has already adjusted the model parameters... The optimization has reached a sensitive "position": starting from this position, for any new task, it is only necessary to follow the gradient direction indicated by the task's data (by...). (Indicates) move a small step (by) By controlling the parameters, a point can be reached that performs well for the new task. This means the model possesses a very strong potential to "generalize to new tasks." This mechanism enables the invention to achieve efficient adaptation with very few samples (even a single sample), perfectly meeting the stringent requirements for rapid debugging in small-batch production of sheet metal bending.

[0093] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the bending springback element learning compensation method integrating machine vision and electrical signal features. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0094] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0095] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0096] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0097] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A bending springback element learning compensation method integrating machine vision and electrical signal features, characterized in that, Includes the following steps: S1. Acquire multimodal timing electrical signals collected in real time when the bending machine performs the bending process; S2. After unloading and springback, trigger the machine vision system to capture an image of the workpiece bending area, and calculate the actual springback angle based on the image; S3. Extract the dynamic process features related to the plastic deformation of the material from the multimodal time-series electrical signal, and associate them with the actual springback angle to form a sample data pair; S4. Input the sample data into a rebound prediction model pre-trained based on a meta-learning framework to drive the model to quickly update the model parameters. S5. Using the updated springback prediction model, predict and compensate the bending process parameters for subsequent workpieces made of the same batch of materials.

2. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 1, characterized in that, The phrase "acquiring multimodal timing electrical signals collected in real time during the bending process of the bending machine" specifically includes: The hydraulic system pressure signal, slider displacement signal, and angle sensor signal installed on the slider or lower die are synchronously acquired during the movement of the slider of the bending machine. The acquired raw signals are time-aligned and filtered for noise reduction. The key forming stage signal from the slider contacting the workpiece to reaching the bottom dead center and ending the pressure holding is extracted from the complete bending cycle signal. The multimodal timing electrical signal includes at least the pressure-time curve, displacement-time curve, and angle-time curve of the key forming stage.

3. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 2, characterized in that, The "extraction of dynamic process features related to material plastic deformation in the multimodal time-series electrical signal" specifically includes: The maximum forming force and the integral area of ​​force with displacement are extracted from the pressure-time curve of the key forming stage; the bending angle change rate is extracted from the angle-time curve; the cross-correlation feature of the pressure signal and the angle signal in the time domain is constructed to characterize the coupling relationship between force and deformation; the extracted maximum forming force, integral area, angle change rate and cross-correlation feature are combined to form a multi-dimensional dynamic process feature vector to characterize the deformation behavior of the material during the bending process.

4. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 1, characterized in that, The phrase "triggering the machine vision system to capture an image of the workpiece's bending area and calculating the actual springback angle based on the image" specifically includes: A binocular vision sensor fixed above the bending machine's workspace is triggered to capture left and right images of the unloaded workpiece from a direction perpendicular to the bending line. The images are then subjected to distortion correction and stereo matching to identify and locate the three-dimensional spatial point cloud of the sheet metal planes on both sides of the workpiece's bending area. The three-dimensional spatial equations of the two planes are calculated using a plane fitting algorithm, and the dihedral angle between the two planes is then calculated as the actual springback angle. This method can compensate for two-dimensional measurement errors caused by workpiece posture or camera perspective.

5. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 3, characterized in that, The pre-training and update method of the "bounce prediction model based on meta-learning framework pre-training" is as follows: The springback prediction model employs a neural network that takes the dynamic process feature vector as input and the predicted springback angle as output; its meta-learning framework includes a meta-learner, which learns a model initial parameter with good generalization by using historical bending data containing various materials and thicknesses during the pre-training stage. During the rapid update phase for a specific batch, the sample data pairs described herein are used as the support set for that batch. The meta-learner guides the neural network to perform a limited number of gradient iterations, thereby rapidly adjusting the model parameters to adapt to the characteristics of the new batch of materials.

6. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 5, characterized in that, The method also includes an optimization step for the initial bending test: When bending the first workpiece of a brand new batch of materials, the initial process parameters based on the process database are used for bending. After completing the visual measurement and rapid model update of the first workpiece, the springback compensation amount predicted by the updated model is fed back to the bending machine CNC system. The system automatically calculates and corrects the target bending angle or bottom dead center depth of subsequent workpieces to compensate for the predicted springback amount, so that qualified angles can be obtained from the second workpiece onwards.

7. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 1 or 4, characterized in that, The step of "calculating the actual rebound angle" further includes the detection and processing of asymmetric rebound: After fitting the two planes based on the 3D point cloud, the actual springback angles of the left and right regions of the workpiece bending line are calculated respectively. When the difference between the left and right springback angles exceeds a preset threshold, it is determined that there is asymmetric springback. The left and right springback angles and their difference are used as an additional feature and associated with the dynamic process feature vector for subsequent model updates and compensation decisions, so as to support indirect compensation for the parallelism of the bending machine slider or the workpiece positioning deviation.

8. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 1, characterized in that, The method is executed by an integrated control system, which includes: The signal acquisition module is used to connect to the sensors of the bending machine and acquire the multimodal timing electrical signals in real time. The machine vision measurement module includes an industrial camera, lens, light source, and image processing unit, used to perform image acquisition, processing, and actual rebound angle calculation; The meta-learning model processing module is used to run the pre-trained rebound prediction model and meta-learning update algorithm, and to store historical data. The process control interface module is used to receive the compensation amount output by the model and communicate with the bending machine CNC system to issue corrected process parameter instructions.

9. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 8, characterized in that, The signal acquisition module and the machine vision measurement module work together through a synchronization triggering unit: the synchronization triggering unit receives the bending cycle status signal from the bending machine CNC system; when the bending process is detected to be completed and the slider returns to the trigger position, it sends an image acquisition trigger signal to the machine vision measurement module; at the same time, the unit provides a unified timestamp reference for the signal acquisition module to ensure the precise correspondence between the electrical signal sequence and the image acquisition time, so as to ensure the accuracy of the temporal correlation of the sample data pairs.

10. The bending springback element learning compensation method integrating machine vision and electrical signal features according to claim 5, characterized in that, The meta-learning framework employs a model-independent meta-learning algorithm during rapid updates, and its rapid update process is as follows: The sample data generated from the first bending of the current batch of materials is used as the support set; the gradient of the loss with respect to the model parameters is calculated with the loss between the predicted and actual springback angles of the model on the current support set as the optimization objective; instead of directly using this gradient to significantly update the model parameters, the meta-learner generates a parameter update direction and step size based on pre-trained knowledge; the parameter update amount guided by the meta-learner is applied to the neural network model, enabling it to quickly adapt to the characteristics of the current batch of materials while retaining general knowledge, thus completing an incremental update.

Images