Metal plate flexible bending path dynamic planning method and system

Abstract

The invention discloses a dynamic planning method and system for a flexible bending path of a metal plate, and belongs to the field of industrial control and intelligent manufacturing. The method comprises the following steps: acquiring multi-source sensing data in a bending process; extracting feature vectors; experience is matched in the historical fragment library, and a correction vector is generated; the online self-adaptive controller generates a self-adaptive action vector; generating a final control instruction through confidence coefficient weighted fusion; and evaluating the new empirical value and updating the fragment library. According to the method, a decision-making mechanism combining online self-adaptive control and historical experience knowledge migration is adopted, real-time, high-precision and self-adaptive compensation of the rebound deviation of the plate in the bending process can be achieved, meanwhile, the safety of a control instruction and rapid starting of a new task are ensured, and the precision and process adaptability of flexible bending manufacturing are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial control and intelligent manufacturing, and in particular to a dynamic planning method and system for flexible bending paths of metal sheets. Background Technology

[0002] During the bending process of sheet metal, the nonlinear mechanical properties of the material and the unavoidable springback phenomenon cause a deviation between the final bending angle and the preset target angle. To ensure dimensional consistency between product batches and meet the requirements of high-precision flexible manufacturing, it is necessary to plan and compensate the bending machine's path and control parameters in real time and dynamically.

[0003] In related technologies, Chinese invention patent CN120421376A discloses a bending machine system based on real-time springback compensation and a method for dynamic calibration of bending angle. It includes a bending machine slider, a worktable, linear guide rails, sensors and a compensation calculation module, and an electrical control cabinet. An upper die is located at the bottom of the bending machine slider, and a lower die matching the upper die is located above the worktable. Material to be bent is placed on the lower die. Linear guide rails are symmetrically arranged on both sides of the worktable, and angle detectors are symmetrically arranged on the linear guide rails. Each angle detector incorporates a high-precision angle sensor for real-time acquisition of material bending angle data, which is then uploaded to the control system of the electrical control cabinet via a communication module. The electrical control cabinet incorporates sensors and a compensation calculation module, and a degree-of-freedom rotating arm is installed on the cabinet. Through real-time feedback, dynamic compensation, and intelligent decision-making, the system solves the accuracy and efficiency bottlenecks caused by material springback.

[0004] Regarding the aforementioned technologies, the inventors believe that although this method achieves real-time acquisition and compensation of angle data, the acquired information is singular, relying solely on bending angle data collected by the angle sensor as the basis for compensation. This single-source information lacks perception of deeper physical information such as local deformation images of the sheet metal during bending and the distribution of contact pressure in the mold. When encountering conditions such as material batch changes, mold wear, or complex bending, the single springback angle information is insufficient to support high-precision and highly robust dynamic compensation decisions, thus making it difficult to continuously improve bending accuracy and process adaptability. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a dynamic planning method and system for flexible bending paths of metal sheets. It employs a control strategy that integrates multi-source synchronous sensing, historical experience and online adaptive control, and closed-loop empirical value evaluation, enabling high-precision dynamic planning of bending paths and continuous adaptive optimization of the control system.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] Firstly, a dynamic planning method for flexible bending paths of metal sheets is provided, including:

[0008] The image sequence of the deformation area of ​​the sheet metal during the bending process, the pressure change curve of the mold contact point, and the actual motion trajectory of the slider are acquired. The image sequence, pressure change curve and actual motion trajectory are time-stamped and aligned to generate a multi-source synchronous sensing data packet.

[0009] Extract data from the multi-source synchronous sensing data packet to generate an instantaneous state feature vector;

[0010] In a fragment library containing historical bending process segments, the instantaneous state feature vector is used for matching and retrieval to generate matching results;

[0011] A real-time empirical correction vector is generated based on the matching result, and the instantaneous state feature vector is input into the online adaptive controller to generate an adaptive action vector.

[0012] Confidence weights are calculated based on the matching results, and the real-time experience correction vector and adaptive action vector are weighted and fused using the confidence weights to generate the final control command.

[0013] The final control command is sent to the bending machine servo for execution, and the actual effect data after the command is executed is collected;

[0014] The multi-source synchronous sensing data packets, final control commands, and actual effect data within the current control cycle are combined into candidate experience segments, and the value of the candidate experience segments is evaluated. Based on the evaluation results, the segment library of historical bending process segments is selectively updated.

[0015] Based on the above technical solution, in the dynamic planning method for flexible bending path of metal sheet provided in this application, a control strategy of multi-source synchronous sensing, dual-path fusion of historical experience and online adaptive control, and closed-loop experience value evaluation is adopted, which can realize high-precision dynamic planning of bending path and continuous adaptive optimization of the control system.

[0016] In conjunction with the first aspect above, in one possible implementation, acquiring the image sequence of the sheet metal deformation area, the pressure change curve of the mold contact point, and the actual motion trajectory of the slider during the bending process, and aligning the image sequence, pressure change curve, and actual motion trajectory with timestamps, includes:

[0017] A sequence of local deformation images of the bending area of ​​the sheet metal is acquired using a high frame rate vision sensor.

[0018] Real-time pressure distribution data of the contact area between the mold and the sheet metal is collected by a high-response pressure sensor array, and the pressure change curve is obtained by processing.

[0019] The real-time position information of the slider during the bending process is collected by a high-precision displacement sensor, and the actual motion trajectory is obtained through processing.

[0020] Based on a unified clock, the acquired local deformation image sequence, pressure change curve, and actual motion trajectory are synchronized and aligned in time, and then encapsulated into a multi-source synchronous sensing data packet.

[0021] In conjunction with the first aspect above, in one possible implementation, the fragment library of the historical bending process fragments stores multiple historical bending process fragments. Each historical bending process fragment consists of a historical state feature vector, a historical control action vector executed under the historical state feature vector, and historical effect data generated after executing the historical control action vector.

[0022] In conjunction with the first aspect above, in one possible implementation, generating the real-time empirical correction vector based on the matching result includes:

[0023] Determine whether there are several target historical segments in the matching results whose similarity to the instantaneous state feature vector exceeds a preset similarity threshold;

[0024] If so, the historical control action vectors contained in the aforementioned target historical segments are fused and calculated to generate a real-time experience correction vector;

[0025] If not, a zero vector or a random noise vector is generated as a real-time empirical correction vector.

[0026] In conjunction with the first aspect above, in one possible implementation, calculating the confidence weight based on the matching result and using the confidence weight to perform a weighted fusion of the real-time experience correction vector and the adaptive action vector includes:

[0027] Extract similarity values ​​and historical performance data of the target historical segment from the matching results;

[0028] The confidence weight is calculated by combining the similarity value and the performance evaluation value represented by the historical effect data of the target historical segment.

[0029] The higher the similarity value and the performance evaluation value, the greater the calculated confidence weight, thereby enhancing the influence of the real-time experience correction vector in the weighted fusion.

[0030] In conjunction with the first aspect above, in one possible implementation, the step of evaluating the value of candidate experience fragments and selectively updating the fragment library of historical bending process fragments based on the evaluation results includes:

[0031] Calculate the performance improvement of the actual effect data relative to the target effect data in the candidate experience fragment;

[0032] Calculate the minimum difference between the instantaneous state feature vector in the candidate experience segment and all historical state feature vectors in the historical bending process segment library;

[0033] If the performance improvement is greater than the experience value threshold and the minimum difference is within the range defined by the local exploration threshold and the redundancy filtering threshold, then the candidate experience fragment is compressed and stored in the historical bending process fragment library. The experience value threshold is used to measure whether the performance improvement of the new experience is sufficient to be adopted, defining the value threshold for updating the fragment library. The local exploration threshold is used to ensure that the new experience is not a redundant fragment that is too similar to the existing fragments, encouraging exploration. The redundancy filtering threshold is used to ensure that although the new experience is new, it is still related to the existing knowledge to a certain extent, and is not completely noise or outlier, filtering out meaningless exploration.

[0034] In conjunction with the first aspect above, in one possible implementation, the online adaptive controller is implemented through a policy network. The objective function of the policy network simultaneously optimizes the cumulative reward and the policy entropy. The objective of the cumulative reward is used to drive the adaptive action vector to minimize the rebound deviation, and the objective of the policy entropy is used to drive the adaptive action vector to explore actions in the face of uncertainty.

[0035] In conjunction with the first aspect above, in one possible implementation, the method further includes, after generating the final control command and before sending it to the bending machine servo system for execution:

[0036] Read the safety boundary values ​​from the database that stores the range of safe operating parameters for the device;

[0037] The action parameters contained in the final control command are compared and verified with the safety boundary value;

[0038] If any action parameter exceeds the safety boundary value, the action parameter is limited, a corrected final control command is generated, and the corrected final control command is sent to the bending machine servo system for execution.

[0039] In conjunction with the first aspect above, in one possible implementation, the method includes initialization:

[0040] Obtain task feature information for a new bending task, wherein the task feature information includes material grade, sheet thickness and mold number;

[0041] In the global policy library that stores complete initial control policies, a matching initial control policy is retrieved and loaded based on the task feature information;

[0042] The parameters of the initial control strategy are applied to the online adaptive controller to complete the controller initialization, providing a control starting point that is superior to random initialization.

[0043] Secondly, a dynamic planning system for flexible bending paths of metal sheets is provided, including:

[0044] The multi-source sensing and data alignment module is used to acquire image sequences of the deformation area of ​​the sheet metal during bending, pressure change curves of the mold contact points, and the actual motion trajectory of the slider. It also timestamps the image sequences, pressure change curves, and actual motion trajectories to generate multi-source synchronous sensing data packets.

[0045] The state feature extraction module extracts data from the multi-source synchronous sensing data packet and generates an instantaneous state feature vector;

[0046] The fragment library and experience retrieval module are used to perform matching retrieval in the fragment library containing historical bending process fragments using the instantaneous state feature vector, and generate matching results;

[0047] The experience correction and adaptive control module is used to generate a real-time experience correction vector based on the matching result, and input the instantaneous state feature vector into the online adaptive controller to generate an adaptive action vector;

[0048] The instruction fusion and generation module is used to calculate the confidence weight based on the matching result, and to use the confidence weight to perform weighted fusion of the real-time experience correction vector and the adaptive action vector to generate the final control instruction.

[0049] The instruction sending and data acquisition module is used to send the final control instruction to the bending machine servo for execution and to acquire the actual effect data after the instruction is executed;

[0050] The experience fragment value assessment and update module is used to combine the multi-source synchronous sensing data packets, final control commands and actual effect data in the current control cycle into candidate experience fragments, assess the value of the candidate experience fragments, and selectively update the fragment library of historical bending process fragments based on the assessment results.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] This invention achieves precise dynamic planning of the bending process by constructing a closed-loop control system that integrates multi-source sensing, experience reuse, and online adaptive learning. By fusing multi-dimensional information such as sheet deformation images, contact pressure, and slider position in real time, a comprehensive and profound understanding of the current physical state can be obtained, thereby accurately predicting and compensating for nonlinear effects such as springback, improving the control accuracy of the bending angle and the consistency between product batches.

[0053] The dual-path fusion decision-making mechanism of experience correction and adaptive control designed in this invention greatly enhances the system's flexibility and adaptability. It can dynamically adjust the degree of dependence on historical strategies based on the matching degree and reliability of historical experience. When handling similar tasks, mature solutions can be quickly reused to ensure efficiency. When facing unknown working conditions such as new materials and new molds, it can rely on online adaptive controllers for exploration and learning, achieving efficient support for flexible production modes of small batches and multiple varieties.

[0054] The experience fragment value assessment and selective update mechanism established in this invention enables the system to continuously self-optimize. By evaluating the value and novelty of newly generated candidate experiences, high-value control experiences can be automatically screened and absorbed, while redundant or inefficient data is eliminated, ensuring the conciseness and efficiency of the experience base. This self-learning and evolutionary capability allows the system performance to continuously improve with increased production practice, achieving long-term iterative optimization of the control strategy.

[0055] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A structural architecture diagram of a dynamic planning system for flexible bending paths of metal sheets provided in this application embodiment;

[0058] Figure 2 A flowchart illustrating a dynamic planning method for flexible bending paths of metal sheets provided in this application embodiment;

[0059] Figure 3This is a space diagram for the evaluation and acceptance of experience fragments provided in the embodiments of this application.

[0060] Figure 4 This is a comparison chart of the change in rebound error with the control cycle provided in the embodiments of this application. Detailed Implementation

[0061] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0062] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0063] The dynamic planning method for flexible bending paths of metal sheets provided in this application embodiment can be applied to, for example... Figure 1 In the dynamic planning system 100 for flexible bending paths of metal sheets shown, such as Figure 1 As shown, the system includes: a multi-source sensing and data alignment module, which is used to acquire image sequences of the deformation area of ​​the sheet metal during bending, pressure change curves of the mold contact points, and the actual motion trajectory of the slider, and to perform time stamp alignment on the image sequences, pressure change curves, and actual motion trajectories to generate multi-source synchronous sensing data packets;

[0064] The state feature extraction module extracts data from the multi-source synchronous sensing data packet and generates an instantaneous state feature vector;

[0065] The fragment library and experience retrieval module are used to perform matching retrieval in the fragment library containing historical bending process fragments using the instantaneous state feature vector, and generate matching results;

[0066] The experience correction and adaptive control module is used to generate a real-time experience correction vector based on the matching result, and input the instantaneous state feature vector into the online adaptive controller to generate an adaptive action vector;

[0067] The instruction fusion and generation module is used to calculate the confidence weight based on the matching result, and to use the confidence weight to perform weighted fusion of the real-time experience correction vector and the adaptive action vector to generate the final control instruction.

[0068] The instruction sending and data acquisition module is used to send the final control instruction to the bending machine servo for execution and to acquire the actual effect data after the instruction is executed;

[0069] The experience fragment value assessment and update module is used to combine the multi-source synchronous sensing data packets, final control commands and actual effect data in the current control cycle into candidate experience fragments, assess the value of the candidate experience fragments, and selectively update the fragment library of historical bending process fragments based on the assessment results.

[0070] like Figure 2 As shown in the figure, this application provides a dynamic planning method for flexible bending paths of metal sheets, including:

[0071] The image sequence of the deformation area of ​​the sheet metal during the bending process, the pressure change curve of the mold contact point, and the actual motion trajectory of the slider are acquired. The image sequence, pressure change curve and actual motion trajectory are time-stamped and aligned to generate a multi-source synchronous sensing data packet.

[0072] Extract data from the multi-source synchronous sensing data packet to generate an instantaneous state feature vector;

[0073] In a fragment library containing historical bending process segments, the instantaneous state feature vector is used for matching and retrieval to generate matching results;

[0074] A real-time empirical correction vector is generated based on the matching result, and the instantaneous state feature vector is input into the online adaptive controller to generate an adaptive action vector.

[0075] Confidence weights are calculated based on the matching results, and the real-time experience correction vector and adaptive action vector are weighted and fused using the confidence weights to generate the final control command.

[0076] The final control command is sent to the bending machine servo for execution, and the actual effect data after the command is executed is collected;

[0077] The multi-source synchronous sensing data packets, final control commands, and actual effect data within the current control cycle are combined into candidate experience segments, and the value of the candidate experience segments is evaluated. Based on the evaluation results, the segment library of historical bending process segments is selectively updated.

[0078] It should be noted that by aligning timestamps, high-frequency acquired images of sheet deformation, mold contact pressure, and slider motion trajectory are integrated into a multi-source synchronous sensing data packet, from which instantaneous state feature vectors are extracted to achieve a comprehensive and dynamic description of the physical state of the bending process. The core decision-making process consists of two parallel paths: one is an experience-driven path, which uses current state features to quickly match and retrieve data from a historical experience fragment library to generate a real-time experience correction vector based on past successful cases; the other is a model-driven path, which inputs current state features into an online adaptive controller to generate an adaptive action vector based on real-time model inference. These two control vectors are dynamically fused by confidence weights, which are adaptively adjusted according to the matching degree and effectiveness of historical experience. This ensures that high-quality experience is fully utilized when available, while relying more on the exploratory capabilities of the adaptive controller when facing new working conditions. After the final generated control command is sent and executed, actual effect data is collected and combined with the current cycle's state and actions to form candidate experience fragments. After value assessment, only those experiences that demonstrate improved efficiency or explore new state spaces are selectively stored in the experience fragment library, thereby enabling the continuous evolution and optimization of the experience library and forming a self-improving intelligent control closed loop.

[0079] In one possible implementation of the embodiments of this application, combined with Figure 2 The above-mentioned acquisition of image sequences of the sheet metal deformation area, pressure change curves of the mold contact points, and actual movement trajectories of the slider during the bending process, and the timestamp alignment of the image sequences, pressure change curves, and actual movement trajectories, includes:

[0080] A sequence of local deformation images of the bending area of ​​the sheet metal is acquired using a high frame rate vision sensor.

[0081] In some implementations, a high-frame-rate vision sensor deployed in the throat area of ​​the bending machine continuously captures images of the localized deformation area where the sheet metal contacts the V-groove of the lower die. A high-frame-rate vision sensor is an industrial camera characterized by its ability to capture hundreds of frames per second. It is used to capture the rapid, subtle geometric changes that occur in sheet metal under high-speed pressure. Its frame rate is typically set in the range of 200 to 500 Hz to ensure the capture of transient changes such as material yielding, corner forming, and minute springback, thereby generating a sequence of localized deformation images. Each frame in this sequence records the geometric contour of the sheet metal at a specific moment.

[0082] For example, deploy an industrial vision sensor with a frame rate of 300Hz, starting from the moment the bend begins. Begin continuously photographing the contact area between the sheet material and the V-groove. During the bending process... Instant During this period, the sensor acquired a total of 3 images, each marked with a timestamp. , , ,in This frame clearly captures the instantaneous geometric profile and minute springback of the metal sheet under pressure.

[0083] Real-time pressure distribution data of the contact area between the mold and the sheet metal is collected by a high-response pressure sensor array, and the pressure change curve is obtained by processing.

[0084] In some implementations, a high-response pressure sensor array integrated under the contact surface between the lower die or upper punch and the sheet metal is used to collect real-time pressure distribution data in the contact area between the die and the sheet metal. The high-response pressure sensor array is an array composed of multiple miniature pressure sensing units, capable of rapidly detecting the spatial distribution and temporal changes of force on the die surface at a kilohertz sampling rate. The collected multi-point pressure data is processed by an onboard processor through spatial weighted averaging or peak extraction to form a pressure change curve characterizing the overall contact force.

[0085] For example, a component is integrated inside the lower mold. A high-response pressure sensor array composed of miniature sensing units, with a sampling rate set to 2000Hz, was used to collect force data at 16 points in real time during the bending process. At that time, these 16 pressure values ​​were spatially weighted and averaged by the onboard processor to obtain the instantaneous value of the overall contact force. ; then in Then, after processing again, the instantaneous value is obtained. This results in a pressure change curve describing the overall contact force over time.

[0086] The real-time position information of the slider during the bending process is collected by a high-precision displacement sensor, and the actual motion trajectory is obtained through processing.

[0087] In some implementations, a high-precision displacement sensor, such as a grating ruler or a laser interferometer, mounted on the slider is used to monitor the slider's vertical position in real time with micron-level accuracy. Its sampling frequency is usually synchronized with that of the pressure sensor to generate the slider's actual motion trajectory.

[0088] For example, a high-precision grating ruler displacement sensor with micron-level accuracy is installed on the slider side of a bending machine, and its sampling frequency is set to 2000Hz in sync with the pressure sensor. The sensor... Record the vertical position of the slider at time. millimeters; in The time record position is Millimeters, these high-frequency position data points together depict the actual motion trajectory of the slider.

[0089] Based on a unified clock, the acquired local deformation image sequence, pressure change curve, and actual motion trajectory are synchronized and aligned in time, and then encapsulated into a multi-source synchronous sensing data packet.

[0090] In some implementations, a unified clock is used to ensure the temporal consistency of data from different sources. This clock serves as the source of a standard time reference for all data acquisition units and can be implemented via the Network Time Protocol (NTP) or hardware synchronization signals, ensuring that the timestamps of all sensor data have a unified reference system. Each data point acquired by a sensor, whether it be an image frame, pressure value, or position reading, is appended with a high-precision timestamp. A time synchronization alignment algorithm is executed to aggregate the acquired local deformation image sequences, pressure change curves, and actual motion trajectories. The alignment process is based on a preset system time step, such as 5 to 10 milliseconds. At each time step node, sensor data with different sampling rates are aligned to that time point using a lookup or interpolation algorithm, and finally encapsulated into a multi-source synchronous sensing data packet. This data packet is a structured data container tailored to the bending state at a specific moment. The generated multi-source synchronous sensing data packet is represented by the following formula:

[0091] ;

[0092] in, This represents the k-th multi-source synchronous sensing data packet. Synchronization timestamps provided for a unified clock. Represents the timestamp Locally deformed image frames acquired nearby are selected from the image sequence using the nearest neighbor algorithm. Represents the timestamp The instantaneous pressure value on the pressure change curve obtained by time processing, if the original sampling point is not accurately matched Then it is calculated using linear interpolation. Represents the timestamp The position coordinates of the slider on its actual motion trajectory are also obtained from the readings of the high-precision displacement sensor through interpolation.

[0093] For example, a unified hardware synchronization signal is used as the clock reference, and the time step for data synchronization is defined as 5 milliseconds. To generate... Real-time multi-source synchronous sensing data packets First of all To synchronize timestamps, the nearest neighbor algorithm is then used to select the image sequence with the closest timestamps. As Next, since the original sampling point just happened to record pressure value and displacement value ,therefore and Finally, according to the formula Encapsulation This is a structured data container.

[0094] In one possible implementation, combining Figure 2 The fragment library of the aforementioned historical bending process segments stores multiple historical bending process segments. Each historical bending process segment consists of a historical state feature vector, a historical control action vector executed under the historical state feature vector, and historical effect data generated after executing the historical control action vector.

[0095] In some implementations, a structured, searchable experience knowledge base is constructed, namely a fragment library of historical bending process segments. This fragment library stores past successful or representative bending experiences, providing data support for real-time experience correction. This fragment library is not a simple data listing, but is achieved by organizing it into a series of standardized historical bending process segments. Each historical bending process segment is essentially a data triple, accurately recording the complete causal chain of "state-action-effect" at a specific historical moment, providing a foundation for subsequent matching retrieval and decision fusion. The structure of this triple is as follows: First, the historical state feature vector, generated by feature extraction from the multi-source synchronous sensing data packets at that time during a historical control cycle, is a high-dimensional numerical vector, typically between 128 and 512 dimensions, comprehensively describing the deformation state, stress state, and kinematic characteristics of the sheet metal at that time. Second, the historical control action vector is a parameterized representation of the historical control commands executed following the historical state. Specifically, it contains the action commands output by the system and executed by the bending machine servo system at that time, such as the target displacement increment of the slider, speed, or pressure holding commands. Finally, there is the historical effect data, which is a quantitative evaluation of the physical effects produced after the execution of historical control action vectors. This data comes from actual effect data collected after command execution. Key indicators include, but are not limited to, the deviation between the actual angle measured after bending and the target angle (i.e., springback error), and the energy consumption or time-equivalent performance evaluation value for achieving the desired effect. Through this structure, each segment in the segment library encapsulates a complete micro-decision and its outcome. (A historical bending process segment is shown.) Formalized as:

[0096] ;

[0097] in, The first one in the fragment library A fragment of a historical turning point. The historical state feature vector corresponding to the segment is extracted from historical multi-source synchronous sensing data packets and serves as an "index key" for similarity retrieval in the segment library. The vector representing the historical control actions actually performed in this historical state is an "experience action" that can be referenced after a successful retrieval. This represents historical effect data after the action is performed, used to evaluate the "value" of the experience. For example, a scalar value, the smaller the value, the higher the bending accuracy.

[0098] For example, constructing the first fragment in a fragment library A fragment of a historical bending process First, the historical state feature vector. It is a 256-dimensional numerical vector extracted from the multi-source synchronous sensing data packets at that time during a historical control cycle, for example: This vector comprehensively describes the deformation state, stress conditions, and kinematic characteristics of the plate at that time. Secondly, the historical control motion vector... It is following closely The parameterized representation of the control commands executed based on the current state; specifically, the action commands output by the system and executed by the bending machine servo system at that time, such as the target displacement increment of the slider. Speed ​​maintenance command is and pressure fine-tuning increments are ,Right now Finally, historical performance data. Yes The quantitative evaluation of the physical effects produced after execution includes key indicators such as a deviation of 0.05 degrees between the actual angle measured after bending and the target angle, and an effectiveness evaluation value of 0.001 for achieving this effect. This is based on formal representation. The 75th historical bend Stored as: The scalar value of 0.001 represents the performance evaluation value, and the smaller the value, the higher the bending accuracy.

[0099] In one possible implementation, combining Figure 2 The above-mentioned generation of real-time empirical correction vectors based on the matching results includes:

[0100] Determine whether there are several target historical segments in the matching results whose similarity to the instantaneous state feature vector exceeds a preset similarity threshold;

[0101] In some implementations, the generated matching results are analyzed. These results include a series of historical bending process segments similar to the current instantaneous state feature vector, along with their corresponding similarity scores. First, it is determined whether there are several target historical segments whose similarity to the current instantaneous state feature vector exceeds a preset similarity threshold. Similarity is typically calculated using cosine similarity or the reciprocal of a weighted Euclidean distance. The preset similarity threshold is an adjustable parameter between 0.85 and 0.98, used to filter out truly valuable and highly relevant historical experiences.

[0102] For example, suppose the current instantaneous state feature vector The data has been generated, and the 10 most similar historical segments have been retrieved from the segment library. A preset similarity threshold has been set. The value was set to 0.90. After comparison, a segment was found... The similarity is 0.95, and the fragments The similarity is 0.92, and the fragments The similarity of one segment was 0.90, while the similarity of the remaining segments was all below 0.90. This is because the segments... , , The similarity scores are all greater than or equal to 0.90, therefore it is determined that there are three target historical segments that meet the conditions.

[0103] If so, the historical control action vectors contained in the aforementioned target historical segments are fused and calculated to generate a real-time experience correction vector;

[0104] In some implementations, if the condition is met (i.e., at least one or more target historical segments satisfy the criteria), the historical control action vectors contained in these segments are fused and calculated to generate a real-time experience correction vector. The fusion calculation typically uses a weighted average method, where the weights are directly derived from the similarity scores between each target historical segment and the current instantaneous state feature vector. This ensures that historical experiences with higher similarity scores have greater influence in generating the correction vector. The calculation formula is as follows:

[0105] ;

[0106] in, This represents the final generated real-time experience correction vector. (Summation symbol) The calculation range is all target historical segments whose similarity exceeds the preset similarity threshold. It is the first The similarity value between a target's historical fragment and its current instantaneous state feature vector. Representing the The formula uses historical control action vectors recorded in a target historical segment to ensure that the generated real-time experience correction vector can integrate the advantages of multiple optimal historical experiences.

[0107] For example, target historical fragments , , Corresponding to their respective historical control action vectors , , These represent the slider displacement increment, velocity, and pressure correction, respectively. A weighted average method is used to fuse these three motion vectors, where the weights are... We directly take their similarity scores. Let's assume their historical action vectors are respectively... , , Real-time empirical correction vector The calculation formula is: First, calculate the sum of the weights. Then calculate. The components: Calculated This is the real-time experience correction vector.

[0108] If not, a zero vector or a random noise vector is generated as a real-time empirical correction vector.

[0109] In some implementations, if the judgment is negative—that is, if insufficient similar historical experience is found in the fragment library of historical bending process segments—it indicates that the current situation is entirely new or rare, and historical experience has no direct guiding significance. In this case, a zero vector is generated as a real-time experience correction vector. The engineering significance of generating a zero vector is that it explicitly indicates that experience correction will not provide any correction opinions within the current control cycle. Subsequent control commands will be generated entirely by the online adaptive controller, avoiding the introduction of erroneous experience guidance under uncertain conditions, and ensuring the stability of the control system and its ability to explore new operating conditions. Under some exploratory strategies, a low-amplitude random noise vector can also be generated to stimulate exploration, but generating a zero vector is a more robust default choice.

[0110] For example, suppose that in another control period, among the 10 most similar historical segments retrieved, the highest similarity is only 0.80, which is below a preset threshold. The system determines that no target historical segment meets the similarity requirement. Therefore, it chooses to generate a zero vector. This zero vector serves as a real-time experience correction vector. It is then used for weighted fusion, ensuring that the final control command relies entirely on the output of the online adaptive controller, thereby encouraging the system to autonomously explore new operating conditions where reliable historical experience is lacking.

[0111] In one possible implementation, combining Figure 2 The above-mentioned calculation of confidence weights based on the matching results, and the weighted fusion of the real-time experience correction vector and the adaptive action vector using the confidence weights, includes:

[0112] Extract similarity values ​​and historical performance data of the target historical segment from the matching results;

[0113] In some implementations, all data that are identified as target historical segments are extracted from the generated matching results, specifically including the similarity value of each target historical segment and its corresponding historical effect data.

[0114] For example, suppose that in a certain control cycle, two target historical fragments that meet certain conditions are extracted from the matching results. and Fragment Similarity score The value is 0.95, based on its historical performance data. for Degree; Fragment Similarity score The value is 0.92, based on its historical performance data. for Degree. Simultaneously, set the adjustment coefficient. .

[0115] The confidence weight is calculated by combining the similarity value and the performance evaluation value represented by the historical effect data of the target historical segment.

[0116] In some implementations, confidence weights are calculated using extracted data. This requires transforming raw historical performance data, such as rebound error in degrees, into standardized performance evaluation values. Performance evaluation values ​​are dimensionless scalars used to characterize the quality of historical experience; higher values ​​indicate better performance. Commonly used transformation functions are Gaussian or negative exponential functions, ensuring that smaller errors result in higher performance evaluation values ​​that approach 1. The calculation process is as follows:

[0117] ;

[0118] in, Representing the The performance evaluation value of a target historical segment. This refers to the historical performance data of this segment, specifically the historical rebound error value. This is a positive adjustment coefficient used to control the sensitivity of the performance evaluation function to errors; its value typically ranges from 0.5 to 2.0. The similarity scores of all target historical segments and their corresponding performance evaluation scores are aggregated to calculate a single confidence weight representing the overall reliability of the current experience. This weight is obtained by weighted averaging of the performance evaluation scores of all target historical segments; the weight itself is the similarity score of each segment, and its calculation formula is:

[0119] ;

[0120] in, Σ represents the final calculated confidence weight, with a value range of 0 to 1. Σ represents the summation over all target historical segments. It is the first The similarity value between a target's historical fragment and its current instantaneous state feature vector. It is the first The formula evaluates the effectiveness of historical segments targeting a given target. This formula ensures that segments with higher similarity and better historical performance contribute more to the confidence weight, thereby enhancing the influence of high-quality, highly relevant experiences.

[0121] For example, using formula Transform historical performance data into performance evaluation values : fragment Performance evaluation value for fragment Performance evaluation value for Then, the confidence weights are calculated. : The final calculated confidence weight It is approximately 0.9985.

[0122] The higher the similarity value and the performance evaluation value, the greater the calculated confidence weight, thereby enhancing the influence of the real-time experience correction vector in the weighted fusion.

[0123] In some implementations, the calculated confidence weights are used. The real-time experience correction vector and the adaptive action vector are weighted and fused. The adaptive action vector is the initial control suggestion generated by the online adaptive controller based on the current instantaneous state feature vector. The process of weighted fusion to generate the final control command is defined by the following formula:

[0124] ;

[0125] in, This represents the final control command generated. It is a real-time experience correction vector generated based on historical experience. This is the adaptive action vector output by the online adaptive controller. Confidence weights. This is used here as a dynamic fusion coefficient. When A higher value indicates greater confidence in the retrieved historical experience, and the final control command will be more biased towards the real-time experience correction vector. Conversely, when A lower value indicates that historical experience is not very valuable for reference, and more reliance will be placed on the adaptive action vectors generated by the online adaptive controller. This achieves a dynamic balance between experience-driven and model-driven approaches.

[0126] For example, the calculated confidence weights This indicates that historical experience has a very high degree of reliability. Next, this weight will be used to adjust the real-time experience vector. and adaptive action vectors Perform weighted fusion. Assume the real-time empirical correction vector is... The original motion vector output by the online adaptive controller is Final control command The calculation formula is: .because Final control command Extremely biased : Calculation results This result is better than... Slightly close However, in reality, it is based on highly confident experience. The term "dominant" reflects the dominant role of experience-driven approaches.

[0127] In one possible implementation, combining Figure 2 The above-mentioned evaluation of the candidate experience fragments, and the selective updating of the fragment library of historical bending process fragments based on the evaluation results, includes:

[0128] Calculate the performance improvement of the actual effect data relative to the target effect data in the candidate experience fragment;

[0129] In some implementations, the evaluation process is divided into two parallel dimensions. The first dimension is performance evaluation, which calculates the performance improvement of the actual performance data in the candidate experience fragments relative to the target performance data. Here, the actual performance data is typically the actual rebound error of the current bend, while the target performance data can be the historical performance data corresponding to the most similar historical fragment retrieved from the fragment library under the current state. Through this comparison, it can be determined whether the new control strategy is superior to existing best practices. The performance improvement is calculated as follows:

[0130] ;

[0131] in, The value represents the extent of the performance improvement; a positive value indicates that the performance has improved. It is a function that maps performance data to effectiveness evaluation values. The actual effect data representing candidate experience segments. This represents historical performance data from the most similar historical clip matched from the clip library. When a new control action results in a smaller error, It will be greater than This results in a positive increase in efficiency.

[0132] For example, suppose the current candidate experience fragment The actual effect was 0.015 degrees, while the most similar historical fragment matched from the fragment library was... The historical performance data is 0.02 degrees. Define the performance function. First, calculate the performance values ​​corresponding to the two effects: The efficiency value is ,and The efficiency value is .because With smaller errors, its performance value is higher. Therefore, the performance improvement is greater. The calculation formula is: .

[0133] Calculate the minimum difference between the instantaneous state feature vector in the candidate experience segment and all historical state feature vectors in the historical bending process segment library;

[0134] In some implementations, the second dimension is novelty assessment, which calculates the minimum difference between the instantaneous state feature vector in the candidate experience fragment and all historical state feature vectors in the historical bending process fragment library. This calculation aims to quantify the similarity between the current working condition and all known working conditions in the historical database, and its calculation method is as follows:

[0135] ;

[0136] in, This represents the minimum degree of difference. It is the instantaneous state feature vector in the candidate experience fragment. The first in the representative fragment library Each historical state feature vector. This represents the Euclidean or Mahalanobis distance between two vectors. The smaller the value, the more similar the current state is to a certain historical state in the fragment library.

[0137] For example, suppose the instantaneous state feature vector in the candidate experience fragment It is a 128-dimensional vector, and the fragment library contains... Historical state feature vectors Calculate one by one With each Euclidean distance between Assume that the minimum distance among all calculated distances appears in the fragment. At that location, its value is 0.15. Therefore, the minimum difference... The calculation result is This value quantifies the distance between the current operating condition and the most similar operating condition known in the historical database.

[0138] If the performance improvement is greater than the experience value threshold and the minimum difference is within the range defined by the local exploration threshold and the redundancy filtering threshold, then the candidate experience fragment is compressed and stored in the historical bending process fragment library. The experience value threshold is used to measure whether the performance improvement of the new experience is sufficient to be adopted, defining the value threshold for updating the fragment library. The local exploration threshold is used to ensure that the new experience is not a redundant fragment that is too similar to the existing fragments, encouraging exploration. The redundancy filtering threshold is used to ensure that although the new experience is new, it is still related to the existing knowledge to a certain extent, and is not completely noise or outlier, filtering out meaningless exploration.

[0139] In some implementations, the two evaluation results are jointly judged based on a preset threshold. The trigger condition for accepting a candidate experience fragment and updating it to the fragment library is the magnitude of its performance improvement. It must be greater than the empirical value threshold, and its minimum difference must be... The new experience must be within the range defined by the local exploration threshold and the redundancy filtering threshold. The experience value threshold is a performance improvement threshold, for example, set to 0.05, used to measure whether the performance improvement brought by the new experience is significant enough. The local exploration threshold is the lower bound of the minimum difference, used to ensure that the new experience is not redundant data that is too similar to existing fragments, avoiding fragment library expansion. The redundancy filtering threshold is the upper bound of the minimum difference, used to ensure that while the new experience is novel, it still maintains a certain relevance to the existing knowledge system, filtering out meaningless data that may be caused by sensor noise or extreme abnormal operating conditions. If determined to be correct, the candidate experience fragment is compressed, for example, by downsampling or feature encoding the image data, and then stored in the historical bending process fragment library, completing one learning and iteration cycle. Figure 3As shown, the distribution of candidate experience fragments across two dimensions—performance improvement and minimum dissimilarity—is illustrated. An "acceptance region" is defined using an experience value threshold, a local exploration threshold, and a redundancy filtering threshold. Only fragments falling within this region are accepted and stored in the fragment library, ensuring that accepted fragments possess both sufficient performance improvement value and appropriate state novelty, thereby achieving both efficiency and conciseness in the experience library.

[0140] For example, setting an experience value threshold. The local exploration threshold is 0.0001. The lower limit is 0.1, which is the redundancy filtering threshold. The upper limit is 0.5. Based on the calculation results of the first two steps, the efficiency improvement is... satisfy The conditions. Simultaneously, minimum difference. satisfy The interval conditions are met. Since both conditions are satisfied, the candidate experience fragment is deemed valuable and novel to a reasonable degree. Therefore, the acceptance action is performed: the high frame rate image data in the candidate experience fragment is JPEG compressed, and the numerical data such as feature vectors are quantized and compressed to 8 bits before being stored in the historical bending process fragment library, completing one learning and knowledge iteration.

[0141] In one possible implementation, combining Figure 2 The aforementioned online adaptive controller is implemented through a policy network. The objective function of the policy network simultaneously optimizes the cumulative reward and the policy entropy. The objective of the cumulative reward is used to drive the adaptive action vector to minimize the rebound deviation, and the objective of the policy entropy is used to drive the adaptive action vector to explore actions in the face of uncertainty.

[0142] In some implementations, the core decision-making logic of the online adaptive controller is constructed by introducing a policy network that balances objective optimization and uncertainty exploration. This allows the network to generate high-quality adaptive action vectors, enabling it to make effective control decisions independently, even without high-quality historical experience for reference. The online adaptive controller is implemented through a policy network, which is a deep neural network whose function is to map the input instantaneous state feature vector to the optimal adaptive action vector. To achieve this goal, the training process of the policy network aims to maximize a composite objective function that simultaneously optimizes the cumulative reward and policy entropy. The specific form of this composite objective function can be expressed as:

[0143] ;

[0144] in, The objective function of the policy network is used to find a set of network parameters during training. To maximize the value of this function. The expected value represents the average of all possible trajectories during the entire bending process. Represents the current moment The cumulative reward from the start to the end of the bend is designed to directly drive the adaptive motion vector to minimize the rebound deviation. It is the immediate reward obtained after performing an action, calculated based on collected actual effect data, and is usually related to rebound deviation. Inversely proportional, for example ,in A positive reward scaling factor. The difference between the measured actual bending angle and the target angle is used to minimize the springback deviation, which is equivalent to maximizing the cumulative reward. This is the entropy regularization coefficient, also known as the temperature parameter. It is a key hyperparameter, typically ranging from 0.01 to 0.2, used to balance the objectives of maximizing reward and maximizing entropy. Its value determines the trade-off between pursuing the optimal solution and exploring new actions, and is crucial for action exploration under uncertainty. This represents the policy entropy term, which represents the feature vector of the current instantaneous state. At that time, policy network The entropy value of the output action distribution. Entropy is a concept in information theory, used here to measure the uncertainty of policy output. A higher entropy value indicates greater uncertainty in the policy network's action selection, i.e., stronger exploratory behavior. By jointly optimizing these two objectives, the policy network achieves a dual function. The cumulative reward objective drives policy learning to generate a control action sequence with minimal rebound bias, which is the controller's primary task. The policy entropy objective, as a regularization term, encourages the policy network to avoid premature convergence to a single deterministic action when faced with unfamiliar instantaneous state feature vectors or multiple actions with similar decision values. Instead, it maintains a certain degree of randomness in action exploration, thereby discovering potentially better control strategies and enhancing the system's robustness to environmental changes and long-term learning ability. Figure 4 The diagram shows a comparison of springback error over control cycles. It visually compares the bending accuracy achieved by the method of this invention and traditional control methods over multiple control cycles. Traditional control methods are inferior to the method of this invention in both convergence speed and final stability error. This invention, through joint optimization of cumulative reward and policy entropy via a policy network, can rapidly converge the springback error to an extremely low level within fewer control cycles, thus improving bending accuracy and production efficiency.

[0145] For example, at a certain moment during the bending process Acquire instantaneous state feature vector This vector is input into the policy network. The policy network aims to maximize the composite objective function. Set the entropy regularization coefficient. This coefficient balances the pursuit of performance with the exploration of new actions. For example, in policy networks... according to Calculate the action And assess the expected cumulative reward of this action as This is a negative value, representing the expected minimum penalty, and it also calculates the policy entropy of the action in the action space distribution. At this point, the composite objective function The immediate evaluation value The training objective is to iteratively update the network parameters. To bring it closer to 0, i.e., to maximize Assume at time... Execute action The subsequent data collected showed the instantaneous springback error of this bend. The value is 0.01 degrees. Using the instant reward formula... Calculate the instant reward, where the reward scaling factor is... Set to 1000. The instant reward received at this point. This negative reward value is immediately fed back to the policy network for updating, clearly indicating the action to be taken. A rebound deviation of 0.01 degrees was generated, and the optimization of the cumulative reward objective was achieved through continuous selection. Minimum actions to maximize This brings it close to 0, thereby directly driving the adaptive action vector to minimize the rebound bias. Assume a policy network... When faced with an unfamiliar instantaneous state feature vector At that time, the output action distribution is relatively dispersed, and the policy entropy term The value is relatively high, for example, 15.0. When the policy network finds during training that the optimal reward value for a given state is close to convergence, but the policy entropy is decreasing, the entropy target... It will come into play. Assuming this is the case... Strategy Entropy Target for If the network attempts to converge to a deterministic action, the policy entropy will drop to 5.0. Evaluation value .because , For the original The policy network will be penalized, thus being driven to maintain the diversity and randomness of the output action distribution, preventing it from getting trapped in local optima and enabling the exploration of uncertain working conditions.

[0146] In one possible implementation, combining Figure 2 The above method, after generating the final control command and before sending it to the bending machine servo system for execution, also includes:

[0147] Read the safety boundary values ​​from the database that stores the range of safe operating parameters for the device;

[0148] In some implementations, safety boundary values ​​are read from a database that stores the safe operating parameter ranges of the device. This database is a static or semi-static configuration file containing the limit values ​​of various key actuator parameters of the bending machine. Its contents are provided by the equipment manufacturer or pre-set by experienced engineers based on the specific equipment model, die configuration, and material properties. These safety boundary values ​​define the "safety envelope" for normal operation of the equipment. Key parameters typically include the maximum downward speed of the slide block, the maximum permissible bending force, the minimum safe position of the slide block, and the maximum output torque of the servo motor.

[0149] For example, suppose the system reads the target speed of the slider from a database containing the safe operating parameter ranges stored on a dedicated storage device. and target pressure Safety boundary value: The upper limit of the maximum downward speed of the slider. for lower limit for Target pressure ceiling for lower limit for .

[0150] The action parameters contained in the final control command are compared and verified with the safety boundary value;

[0151] In some implementations, the individual action parameters contained in the fused final control command are compared and verified one by one with the corresponding safety boundary values ​​read from the database. The final control command is a vector containing multiple control components, for example, ,in Represents the target speed of the slider. Represents target pressure, Represents the target downlink depth. The command contains... Compare with the boundary value of the maximum downward speed of the slider, and Compare with the boundary value of the maximum permissible bending force, and so on.

[0152] For example, the original final control command was generated. It includes three motion parameters: slider target speed. Target pressure and target downside depth millimeters. Its upper limit By comparison, we found This means the slider target speed has exceeded the safety limit; simultaneously, the target pressure... Compared with its boundary value, it was found that The target pressure is within a safe range.

[0153] If any action parameter exceeds the safety boundary value, the action parameter is limited, a corrected final control command is generated, and the corrected final control command is sent to the bending machine servo system for execution.

[0154] In some implementations, if any action parameter is found to exceed its corresponding safety boundary value during the comparison and verification process, a limiting mechanism will be immediately triggered. Limiting is a simple non-linear operation: if a parameter value exceeds the upper limit, it is forcibly set to the upper limit value; if it is below the lower limit, it is set to the lower limit value. For example, for action parameters exceeding the safety boundary... The correction process is expressed as follows:

[0155] ;

[0156] in, These are the corrected motion parameters. These are the action parameters in the original final control command. and These are the lower and upper limits of the safety boundary for the action parameter, respectively, and these values ​​are read from the database of the device's safe operating parameter range. and The function ensures that the corrected value is always constrained within a safe range. After verifying all parameters and handling any possible limiting, the corrected final control command is generated. This command, while maintaining the original command's intent, eliminates all potentially dangerous operations. This safety-verified corrected final control command is sent to the bending machine's servo system for execution, thus ensuring that the entire dynamic programming process not only pursues high precision and efficiency but also prioritizes equipment safety as the highest principle.

[0157] For example, due to the target speed of the slider Exceeded the limit Immediately apply amplitude limiting to this parameter. The correction process utilizes a formula. Perform: Corrected speed for Target pressure and downside depth Because it remains within the safety envelope, it remains unchanged. Ultimately, the revised final control command is generated. The slider speed is limited to And subsequently, this security-verified... Send to the bending machine servo system for execution.

[0158] In one possible implementation, combining Figure 2 The above method includes initialization:

[0159] Obtain task feature information for a new bending task, wherein the task feature information includes material grade, sheet thickness and mold number;

[0160] In some implementations, before executing a new bending task, an initial control strategy highly matched to the task characteristics is loaded to efficiently initialize the online adaptive controller, providing a control starting point far superior to random initialization and shortening the system's learning and adaptation time to new operating conditions. This initialization process is triggered upon receiving a new bending task. First, the task's characteristic information is acquired. This information is typically input by the operator on the human-machine interface or automatically received from the upper-level manufacturing execution system. Specifically, it includes the material grade used to uniquely identify the sheet metal's mechanical properties, such as SS304 or SPCC; the sheet thickness directly affecting the bending force, such as 1.5 mm; and the die number determining the bending geometry, such as V8 or V12.

[0161] For example, when an operator enters new bending task information on the human-machine interface, the system automatically receives and retrieves its task characteristics: the material grade is confirmed as SS304, the sheet thickness is 2.0 mm, and the die number determining the bending geometry is V12. This discrete information is then combined into a structured query key. .

[0162] In the global policy library that stores complete initial control policies, a matching initial control policy is retrieved and loaded based on the task feature information;

[0163] In some implementations, the acquired discrete task feature information is combined into a structured query key, and then a retrieval operation is performed in a global policy library storing complete initial control policies. The global policy library is a pre-built and offline optimized database containing a large number of validated and efficient initial control policies for different combinations of task feature information. Each initial control policy is essentially a complete set of weights and bias parameters constituting the policy network of the online adaptive controller. The retrieval process is formally represented as:

[0164] ;

[0165] in, This represents the retrieved set of policy network parameters used for initialization. It is a database retrieval function, and it is in the global strategy library. Search within. It is a database that stores a large number of historical best strategies. Its data structure is key-value pairs, where the key is the task feature and the value is the strategy parameter. The query key is composed of task-specific information such as the material grade, sheet thickness, and mold number of the current task. This function uses exact matching or nearest neighbor matching algorithms to find the historically optimal strategy parameters most relevant to the current task.

[0166] For example, the database retrieval function is then called. Assuming a global policy library Includes SS304 with a thickness of The historical optimal strategy parameter set for mold V12 The retrieval function found this parameter set through exact matching and loaded it as the parameter set for the policy network to be initialized. This parameter set Essentially, it is the numerical set of all weights and biases of the policy network.

[0167] The parameters of the initial control strategy are applied to the online adaptive controller to complete the controller initialization, providing a control starting point that is superior to random initialization.

[0168] In some implementations, once the parameters of the initial control strategy are matched... Once loaded, it is immediately applied to the online adaptive controller. Specifically, this application process involves replacing the weights and biases of the policy network from default random values ​​with the loaded parameter set. Through this operation, the online adaptive controller no longer engages in random exploration at the start of the first control cycle. Instead, it directly bases its behavior on a proven, mature strategy that has proven effective for similar tasks. This allows for more accurate and reasonable control decisions in the initial stages of the bending process, laying a high-level foundation for subsequent online fine-tuning and dynamic programming.

[0169] For example, once the parameter set Once loaded, it is immediately applied to the policy network in the online adaptive controller. For example, the policy network's... The equal weights and bias parameters, instead of the default random values, have been replaced with... The corresponding value in the data. Through this initialization operation, when the controller receives the first instantaneous state feature vector, it can output a mature control action based on the material properties of SS304 and the geometry of the V12 mold, avoiding the large springback error of more than 5 degrees that may be caused by random initialization controller in the early stage, and directly controlling the initial deviation within 0.5 degrees.

[0170] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0171] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A dynamic planning method for flexible bending paths of metal sheets, characterized in that, The system includes: The image sequence of the deformation area of ​​the sheet metal during the bending process, the pressure change curve of the mold contact point, and the actual motion trajectory of the slider are acquired. The image sequence, pressure change curve and actual motion trajectory are time-stamped and aligned to generate a multi-source synchronous sensing data packet. Extract data from the multi-source synchronous sensing data packet to generate an instantaneous state feature vector; In a fragment library containing historical bending process segments, the instantaneous state feature vector is used for matching and retrieval to generate matching results; A real-time empirical correction vector is generated based on the matching result, and the instantaneous state feature vector is input into the online adaptive controller to generate an adaptive action vector. Confidence weights are calculated based on the matching results, and the real-time experience correction vector and adaptive action vector are weighted and fused using the confidence weights to generate the final control command. The final control command is sent to the bending machine servo for execution, and the actual effect data after the command is executed is collected; The multi-source synchronous sensing data packets, final control commands, and actual effect data within the current control cycle are combined into candidate experience segments, and the value of the candidate experience segments is evaluated. Based on the evaluation results, the segment library of historical bending process segments is selectively updated.

2. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, The process of acquiring image sequences of the sheet metal deformation area, pressure change curves at the die contact points, and the actual movement trajectory of the slider during bending, and aligning the image sequences, pressure change curves, and actual movement trajectories with timestamps, includes: A sequence of local deformation images of the bending area of ​​the sheet metal is acquired using a high frame rate vision sensor. Real-time pressure distribution data of the contact area between the mold and the sheet metal is collected by a high-response pressure sensor array, and the pressure change curve is obtained by processing. The real-time position information of the slider during the bending process is collected by a high-precision displacement sensor, and the actual motion trajectory is obtained through processing. Based on a unified clock, the acquired local deformation image sequence, pressure change curve, and actual motion trajectory are synchronized and aligned in time, and then encapsulated into a multi-source synchronous sensing data packet.

3. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, The fragment library of the historical bending process segments stores multiple historical bending process segments. Each historical bending process segment consists of a historical state feature vector, a historical control action vector executed under the historical state feature vector, and historical effect data generated after executing the historical control action vector.

4. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, Generating a real-time empirical correction vector based on the matching results includes: Determine whether there are several target historical segments in the matching results whose similarity to the instantaneous state feature vector exceeds a preset similarity threshold; If so, the historical control action vectors contained in the aforementioned target historical segments are fused and calculated to generate a real-time experience correction vector; If not, a zero vector or a random noise vector is generated as a real-time empirical correction vector.

5. The dynamic planning method for flexible bending path of metal sheet according to claim 4, characterized in that, Calculating confidence weights based on the matching results, and then using these confidence weights to perform a weighted fusion of the real-time experience correction vector and the adaptive action vector, includes: Extract similarity values ​​and historical performance data of the target historical segment from the matching results; The confidence weight is calculated by combining the similarity value and the performance evaluation value represented by the historical effect data of the target historical segment. The higher the similarity value and the performance evaluation value, the greater the calculated confidence weight, thereby enhancing the influence of the real-time experience correction vector in the weighted fusion.

6. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, The process of evaluating the value of candidate experience fragments and selectively updating the fragment library of historical bending process fragments based on the evaluation results includes: Calculate the performance improvement of the actual effect data relative to the target effect data in the candidate experience fragment; Calculate the minimum difference between the instantaneous state feature vector in the candidate experience segment and all historical state feature vectors in the historical bending process segment library; If the performance improvement is greater than the experience value threshold and the minimum difference is within the range defined by the local exploration threshold and the redundancy filtering threshold, then the candidate experience fragment is compressed and stored in the historical bending process fragment library. The experience value threshold is used to measure whether the performance improvement of the new experience is sufficient to be adopted, defining the value threshold for updating the fragment library. The local exploration threshold is used to ensure that the new experience is not a redundant fragment that is too similar to the existing fragments, encouraging exploration. The redundancy filtering threshold is used to ensure that although the new experience is new, it is still related to the existing knowledge to a certain extent, and is not completely noise or outlier, filtering out meaningless exploration.

7. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, The online adaptive controller is implemented through a policy network. The objective function of the policy network simultaneously optimizes the cumulative reward and the policy entropy. The objective of the cumulative reward is used to drive the adaptive action vector to minimize the rebound deviation, and the objective of the policy entropy is used to drive the adaptive action vector to explore actions when facing uncertainty.

8. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, The method, after generating the final control command and before sending it to the bending machine servo system for execution, also includes: Read the safety boundary values ​​from the database that stores the range of safe operating parameters for the device; The action parameters contained in the final control command are compared and verified with the safety boundary value; If any action parameter exceeds the safety boundary value, the action parameter is limited, a corrected final control command is generated, and the corrected final control command is sent to the bending machine servo system for execution.

9. The dynamic planning method for flexible bending path of metal sheet according to claim 1, characterized in that, The method includes initialization: Obtain task feature information for a new bending task, wherein the task feature information includes material grade, sheet thickness and mold number; In the global policy library that stores complete initial control policies, a matching initial control policy is retrieved and loaded based on the task feature information; The parameters of the initial control strategy are applied to the online adaptive controller to complete the controller initialization, providing a control starting point that is superior to random initialization.

10. A dynamic planning system for flexible bending paths of metal sheets, characterized in that, The system is used in a dynamic planning method for flexible bending paths of metal sheets as described in any one of claims 1-9, and the system comprises: The multi-source sensing and data alignment module is used to acquire image sequences of the deformation area of ​​the sheet metal during bending, pressure change curves of the mold contact points, and the actual motion trajectory of the slider. It also timestamps the image sequences, pressure change curves, and actual motion trajectories to generate multi-source synchronous sensing data packets. The state feature extraction module extracts data from the multi-source synchronous sensing data packet and generates an instantaneous state feature vector; The fragment library and experience retrieval module are used to perform matching retrieval in the fragment library containing historical bending process fragments using the instantaneous state feature vector, and generate matching results; The experience correction and adaptive control module is used to generate a real-time experience correction vector based on the matching result, and input the instantaneous state feature vector into the online adaptive controller to generate an adaptive action vector; The instruction fusion and generation module is used to calculate the confidence weight based on the matching result, and to use the confidence weight to perform weighted fusion of the real-time experience correction vector and the adaptive action vector to generate the final control instruction. The instruction sending and data acquisition module is used to send the final control instruction to the bending machine servo for execution and to acquire the actual effect data after the instruction is executed; The experience fragment value assessment and update module is used to combine the multi-source synchronous sensing data packets, final control commands and actual effect data in the current control cycle into candidate experience fragments, assess the value of the candidate experience fragments, and selectively update the fragment library of historical bending process fragments based on the assessment results.

Images