A hardware bending angle dynamic compensation control method

Abstract

The present application relates to a kind of hardware bending angle dynamic compensation control method.For the problem that material springback error is difficult to accurately compensate in hardware bending process, by collecting high-frequency micro-strain signal of hardware pre-bending stage, material dynamic response fingerprint is constructed, and material state is tagged with springback tendency coefficient prediction, combined with historical bending data knowledge base to realize compensation angle adaptive optimization.Finally, the compensation angle is adjusted by model self-learning, which significantly improves the consistency and springback compensation accuracy of the same batch of materials.The scheme effectively overcomes the problem of insufficient adaptation of conventional model to material dynamic characteristics, and has high precision, real-time adaptive compensation and performance optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for bending processes of hardware parts, and in particular to a dynamic compensation control method for bending angles of hardware parts. Background Technology

[0002] Currently, the calculation of compensation angles and the identification of dynamic material properties in the field of intelligent control of bending processes for hardware parts mainly rely on the measurement and estimation of macroscopic static physical parameters, such as elastic modulus, yield strength, plate thickness, and bending radius. The mainstream technical approach obtains material parameters through offline tensile testing or mechanical parameter databases, then uses CAE simulation modeling, springback theory formulas, or finite element analysis to predict bending springback behavior, and combines this with process parameters to calculate compensation angles. Some solutions integrate multi-source sensor inputs, including pressure, torque, and angle sensors, or employ machine vision for surface microstructure identification and grading.

[0003] In recent years, to address batch fluctuations in complex materials, industry trends have included integrating higher-precision hardware, employing long-term time-series prediction algorithms, knowledge graph case retrieval, and multi-dimensional physical quantity fusion calibration, aiming to reduce the impact of batch differences on bending consistency. Some high-end production lines have already achieved closed-loop data linkage between material batch databases and real-time operating conditions. However, these methods are limited by expensive sensor deployment, system integration complexity, and computing power requirements, making them difficult to widely apply in general-purpose bending equipment or cost-sensitive production lines. Existing technologies are generally applicable to standard sheet metal and scenarios with small batch fluctuations. For batches of special materials, non-homogeneous coils, or fine-grained performance fluctuations under different rolling directions, traditional parameter estimation and static models cannot achieve high-resolution real-time identification.

[0004] The main shortcomings of existing technologies are: without cost-effective hardware expansion, they cannot quickly identify the internal state and dynamic properties of materials at the piece-by-piece level; parameter estimation and CAE modeling lack the ability to capture the dynamic response of complex materials, resulting in long adjustment cycles; multi-source data fusion and machine vision solutions have high requirements for mechanical structures and computing power, making them unsuitable for traditional production line modifications; and compensation angle generation based on static theoretical models is easily affected by batch differences and has limited self-learning capabilities. Therefore, there is an urgent need for a technical solution that can achieve high-precision, fine-grained identification of material dynamic characteristics without adding extra hardware. This solution should extract intrinsic behavior patterns to achieve an intelligent closed-loop mapping between material properties and compensation strategies, meeting the requirements for low-cost, fast-response, and highly adaptable production line deployment. Summary of the Invention

[0005] This application provides a dynamic compensation control method for the bending angle of hardware parts, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0006] This application provides a method for dynamic compensation and control of the bending angle of hardware components, specifically including: S1: Acquire the original multi-channel micro-strain-time response signal of the hardware component during the pre-bending stage by a distributed micro-strain sensor array; S2: Based on the original multi-channel micro-strain-time response signal, perform sliding window segmentation processing to generate a three-segment micro-strain evolution time series segment; S3: Extract the distribution density of the inflection point of the rising slope, the energy ratio of the oscillation frequency band of the plateau segment, and the gradient change parameters of the relaxation decay rate of the unloading segment from the three-segment micro-strain evolution time series segment to generate a dimensionless dynamic response fingerprint. S4: Input the dimensionless dynamic response fingerprint into the graph neural network model trained offline, and output the material performance cluster label that identifies the fine-grained performance cluster to which the material's current state belongs and the corresponding springback tendency coefficient; S5: Based on the material property cluster label, retrieve the preset cluster-compensation mapping knowledge base to obtain the median value of the optimal compensation angle interval directly associated with the material property cluster label; S6: The median value of the optimal compensation angle interval is superimposed on the preset basic compensation model output value to generate an adaptive dynamic compensation angle command for driving the upper mold to perform the formal bending action; S7: Use the measured springback deviation data after executing the adaptive dynamic compensation angle command and completing the formal bending action to perform online weight fine-tuning on the graph neural network model.

[0007] The method for dynamic compensation and control of bending angle of hardware parts provided in this application has the following beneficial effects: (1) This application introduces a micro-load excitation and high-frequency strain response acquisition mechanism to actively excite the intrinsic dynamic behavior of the material in the sub-springback critical region before formal forming, and constructs a dynamic response fingerprint characterization system based on multi-point temporal feature extraction. By performing refined analysis on the micro-strain evolution morphology of the three stages of initial loading, plateau maintenance and unloading relaxation, 12-dimensional dimensionless features such as the inflection point density of the rising slope, the energy ratio of the oscillation frequency band, and the attenuation rate gradient are extracted, so that materials in different states with similar macroscopic parameters can be accurately distinguished, which significantly improves the identification resolution and robustness of material properties.

[0008] (2) By adopting a graph neural network model combined with offline training and online feedback mechanisms, a paradigm shift from parameter estimation to behavior pattern recognition is achieved, avoiding high-cost paths such as CAE simulation, LSTM prediction, and multi-source sensor fusion. It can run efficiently in existing PLC embedded controllers. The fine-grained performance cluster labels output by the model are directly associated with the cluster-compensation mapping knowledge base, establishing a direct connection from fingerprint to the optimal compensation interval. The decision response speed is improved to the level of hundreds of milliseconds. The feedback of the measured deviation after the first bending triggers the online fine-tuning of the network weights, forming a closed-loop self-learning mechanism.

[0009] (3) The overall technical architecture achieves a high degree of integration and automated closed loop of material identification and process compensation, which is suitable for flexible manufacturing scenarios with frequent material changes and no prior trial molding conditions. The distributed micro strain sensor array is embedded in the key contact area of ​​the mold to ensure signal quality. The three-stage time sequence feature extraction takes into account both physical interpretability and algorithm stability. The compensation angle generated in the end has both universality and personalized adjustment capability, which significantly improves the consistency of forming accuracy without sacrificing system stability. Attached Figure Description

[0010] Figure 1 This is the main flowchart of a dynamic compensation control method for bending angle of hardware parts; Figure 2 This is a sub-flowchart of a dynamic compensation control method for bending angle of hardware parts; Figure 3 This is another sub-flowchart of a dynamic compensation control method for bending angle of hardware parts. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0013] like Figure 1 As shown, this application provides a method for dynamic compensation control of the bending angle of hardware parts, specifically including: S1: Acquire the original multi-channel micro-strain-time response signal of the hardware component during the pre-bending stage by a distributed micro-strain sensor array; S2: Based on the original multi-channel micro-strain-time response signal, perform sliding window segmentation processing to generate a three-segment micro-strain evolution time series segment; S3: Extract the distribution density of the inflection point of the rising slope, the energy ratio of the oscillation frequency band of the plateau segment, and the gradient change parameters of the relaxation decay rate of the unloading segment from the three-segment micro-strain evolution time series segment to generate a dimensionless dynamic response fingerprint. S4: Input the dimensionless dynamic response fingerprint into the graph neural network model trained offline, and output the material performance cluster label that identifies the fine-grained performance cluster to which the material's current state belongs and the corresponding springback tendency coefficient; S5: Based on the material property cluster label, retrieve the preset cluster-compensation mapping knowledge base to obtain the median value of the optimal compensation angle interval directly associated with the material property cluster label; S6: The median value of the optimal compensation angle interval is superimposed on the preset basic compensation model output value to generate an adaptive dynamic compensation angle command for driving the upper mold to perform the formal bending action; S7: Use the measured springback deviation data after executing the adaptive dynamic compensation angle command and completing the formal bending action to perform online weight fine-tuning on the graph neural network model.

[0014] Step S1: Acquire the multi-channel micro-strain-time response raw signal of the hardware component during the pre-bending stage using a distributed micro-strain sensor array, and use the multi-channel micro-strain-time response raw signal as the data input source for subsequently constructing the material dynamic response fingerprint spectrum. Specifically, this includes: S1.1: The pre-bending action trigger command of the hardware is parsed to generate a sequence of synchronous control parameters containing the micro-load application interval and sampling time window, which serves as the execution basis for starting the distributed micro-strain sensing array.

[0015] Receive the hardware pre-bending action trigger command from the host computer or PLC controller, and parse the target plate material code, thickness specification and preset micro-load level identifier contained in the command.

[0016] Based on the micro-load level identifier obtained from the analysis, the preset process parameter mapping table is queried to obtain the corresponding micro-load application duration threshold and strain sampling start delay time.

[0017] Based on the real-time displacement rate fed back by the mold closing speed sensor, the estimated physical time from the issuance of the command to the contact of the upper mold with the plate surface is calculated, thereby determining the absolute start time of the sampling time window.

[0018] By combining the microload application duration threshold and the high-frequency sampling requirement, the minimum number of sampling points required is calculated, and a sequence of synchronization control parameters including the sampling channel mask, sampling rate setting value, and data buffer size is generated.

[0019] The generated sequence of synchronous control parameters is encapsulated into a standard communication protocol data packet and sent to the control interface of the distributed micro strain sensor array via a high-speed bus as the basis for initiating data acquisition.

[0020] Through the above-described analysis and parameter generation process, abstract action triggering commands are transformed into precise timing control parameter sequences, achieving microsecond-level synchronization between the sensor array and mechanical actions, and ensuring the integrity and timing consistency of micro-strain response signal acquisition.

[0021] S1.2: Based on the synchronous control parameter sequence, drive the upper mold to perform a micro-load loading action that is 5% lower than the conventional forming force, so as to induce a small deformation physical field on the material surface in the critical contact area of ​​the mold to generate a sub-springback critical zone.

[0022] The system receives the sequence of synchronous control parameters generated by S1.1, parses the micro-load application interval threshold and sampling time window start mark contained therein, and converts them into the low-level motion instruction set of the servo drive system.

[0023] The control system calculates the precise displacement required for the upper die to descend based on the threshold of the micro-load application interval. This displacement corresponds to the contact pressure generated at 5% lower than the conventional forming force, ensuring that the material is in the sub-springback critical zone dominated by elastic deformation.

[0024] The servo driver receives the displacement command and drives the upper mold to approach the lower surface of the mold at a constant low speed until the upper mold establishes initial physical contact with the surface of the hardware. At this time, the system records the pressure feedback value at the moment of contact as the zero position reference.

[0025] After confirming the establishment of contact, the control system applies a preset micro-load increment, monitors the pressure sensor feedback in real time through closed-loop force control processing, and dynamically adjusts the downward speed of the upper mold to maintain the load stable within the set micro-load range.

[0026] During the micro-load holding stage, the upper mold applies a continuous micro-bending torque to the hardware, inducing reversible elastic distortion and a small amount of micro-plastic slip in the internal lattice of the material, forming a micro-deformation physical field with the intrinsic properties of the material.

[0027] The micro-deformation physical field is confined to the critical contact area of ​​the mold, and its strain amplitude is controlled between 0.3° and 1.2° to avoid macroscopically visible permanent deformation, while fully stimulating the material's springback response potential to subsequent formal bending.

[0028] Through the aforementioned micro-load loading action, the control parameters of the previous step are transformed into a physical field of micro-deformation in the sub-springback critical region of the material surface, realizing the non-destructive excitation of the dynamic mechanical properties of the material and providing a high signal-to-noise ratio physical basis for subsequent high-frequency strain signal acquisition.

[0029] S1.3: High-frequency synchronous sampling of the physical field of the small deformation is performed by a distributed micro-strain sensor array deployed in the key contact area of ​​the mold to obtain a set of multi-channel analog voltage signals that reflect the intrinsic dynamic response characteristics of the material.

[0030] S1.4: Perform analog-to-digital conversion and timing alignment processing on the set of multiple analog voltage signals to generate a multi-channel discrete digital strain data stream with a unified timestamp reference.

[0031] S1.5: Based on the multi-channel discrete digital strain data stream, perform noise filtering and baseline drift correction operations, output standardized multi-channel micro-strain-time response raw signals, and use them as data input sources for constructing material dynamic response fingerprint spectra.

[0032] Receive a multi-channel discrete digital strain data stream with a unified timestamp reference. This data stream contains high-frequency sampling signals induced by micro-deformation of the mold contact surface, as well as accompanying environmental noise and baseline drift interference.

[0033] A moving average filter is applied to the multi-channel discrete digital strain data stream. The window length is set to two percent of the sampling frequency to smooth high-frequency random noise and preserve the macroscopic evolution trend of micro-strain signals, generating a strain time series after preliminary denoising.

[0034] The wavelet thresholding method is used to perform in-depth purification of the strain time series after preliminary denoising. The Daubechies4 wavelet basis function is selected for three-level decomposition. The soft thresholding function is used to process the detail coefficients to remove residual pulse interference and output a pure micro-strain signal with high signal-to-noise ratio.

[0035] To address the sensor zero-point drift phenomenon present in pure micro-strain signals, a baseline estimation model based on polynomial fitting is constructed. The mean value of the initial stable segment of the signal is selected as the initial reference point, and the baseline offset trajectory in the entire time domain is calculated.

[0036] The baseline drift correction calculation is performed using the following formula: in, This represents the corrected strain value at the i-th sampling point. This represents the original strain value at the i-th sampling point. This is the baseline drift estimate corresponding to the i-th sampling point.

[0037] The amplitude normalization process is performed on the corrected multi-channel signal to map the data of each channel to the dimensionless range of zero to one, thereby eliminating the dimensional influence caused by the difference in sensitivity of different sensors and generating a standardized multi-channel micro-strain-time response original signal.

[0038] Through the above noise filtering and baseline drift correction processing methods, the multi-channel discrete digital strain data stream of the previous step is transformed into a high-fidelity, zero-baseline standardized multi-channel micro-strain-time response raw signal, achieving the expected technical effect of a high-quality data input source required for constructing a material dynamic response fingerprint spectrum.

[0039] For example, for an 8-channel strain data stream with a sampling frequency of 10kHz, a moving average filter with a window length of 200 points is used to truncate high-frequency noise. A 3-level decomposition is performed using the db4 wavelet basis, and a general threshold rule is set to process detail coefficients, significantly improving the signal-to-noise ratio. A 3rd-order polynomial baseline trajectory is fitted, and drift is subtracted according to the formula, with the maximum baseline deviation controlled within 0.5με. After Z-score normalization, the mean of each channel signal approaches zero, and the standard deviation is uniformly standardized to 1, outputting an 8×N dimensional normalized matrix to ensure that subsequent feature extraction is not affected by hardware differences.

[0040] Step S2: Perform sliding window segmentation processing based on the original multi-channel microstrain-time response signal to generate a three-segment microstrain evolution time series segment. The three-segment microstrain evolution time series segment includes standardized initial loading segment data, standardized platform segment data, and standardized unloading segment data. Specifically, it includes: S2.1: Perform first-order difference operation on the original multi-channel micro-strain-time response signal to generate a micro-strain rate gradient sequence that characterizes the trend of micro-strain rate change. This sequence is used to identify the key inflection point position of load state switching.

[0041] Receive the standardized multi-channel micro-strain-time response raw signal after processing in step S1.5. This signal contains a discrete strain data sequence of multiple sensing nodes throughout the pre-bending cycle.

[0042] A point-by-point first-order forward difference operation is performed on the discrete strain data sequence of each channel to calculate the strain increment between adjacent sampling times, thereby constructing an instantaneous micro-strain rate sequence that reflects the rate of change of local deformation.

[0043] The micro-strain rate of the i-th channel at time t is calculated using the following formula: Among them, v i,t Let ε be the micro-strain rate of the i-th channel at time t. i,t Let t be the micro-strain value of the i-th channel at time t. t At the current sampling time, t t-1 This refers to the previous sampling time.

[0044] The generated instantaneous micro-strain rate sequence is subjected to moving average filtering with a window length of 3 to 5 sampling points to suppress the interference of high-frequency quantization noise on rate abrupt changes, thereby generating a smoothed micro-strain rate gradient sequence.

[0045] Zero-crossing points in the smoothed sequence where the value changes from zero to positive are identified as candidate loading start points, zero-crossing points where the value changes from positive to negative are identified as candidate unloading start points, and continuous intervals where the absolute value of the rate is lower than a preset threshold are identified as plastic flow plateau periods.

[0046] Through the above differentiation and filtering processing, the static strain time series data of the previous step is transformed into a dynamic micro-strain rate gradient sequence, which realizes the accurate quantitative identification of the boundary inflection points of the three physical stages of elastic loading, plastic flow and elastic unloading in the pre-bending process, and provides a high signal-to-noise ratio criterion basis for subsequent adaptive threshold detection.

[0047] S2.2: Based on the micro-strain rate gradient sequence, an adaptive threshold detection process is performed to generate a three-segment boundary marker set containing the loading start point, the plastic flow stabilization point, and the unloading termination point. This set defines the precise time window range for each physical stage.

[0048] The micro-strain rate gradient sequence generated by first-order difference operation is received. This sequence characterizes the instantaneous rate fluctuation of the material surface strain over time during the pre-bending process.

[0049] A sliding window statistical filtering process is performed on the micro-strain rate gradient sequence to calculate the mean and standard deviation within the local window, so as to suppress the interference of high-frequency sampling noise on inflection point detection and generate a smoothed rate change trend curve.

[0050] Based on the smoothed rate change trend curve, dynamic threshold segmentation is used to identify the loading start point. The threshold is set to three times the standard deviation of the background noise. When the rate value first continuously exceeds the threshold, it is marked as the start boundary of the elastic loading stage.

[0051] After the loading start point, the slope of the micro-strain rate change is monitored. When the rate changes from a sharp increase to a steady fluctuation and the fluctuation amplitude is lower than the upper limit of the preset hysteresis interval, the material is determined to have entered the plastic flow stable state, and the time corresponding to this turning point is marked as the plastic flow stable moment point.

[0052] After the micro-load unloading command is issued, the process of micro-strain rate returning from negative to zero is monitored. When the absolute value of the rate drops below the background noise level and remains below the preset duration, the unloading process is determined to be over, and this moment is marked as the unloading termination point.

[0053] The loading start point, the plastic flow stabilization point, and the unloading termination point are encapsulated in chronological order to generate a three-segment boundary marker set containing three key physical state switching nodes.

[0054] By using adaptive threshold detection processing, the micro-strain rate gradient sequence of the previous step is transformed into a three-segment boundary marker set containing the loading start, plastic stabilization and unloading termination times. This achieves accurate temporal division of the physical stages of the entire pre-bending process and provides an accurate time window basis for subsequent segmented feature extraction.

[0055] S2.3: Using a three-segment boundary marker set, perform time-domain truncation on the original multi-channel micro-strain-time response signals to generate mutually independent and continuous initial loading segment signal subsets, platform segment signal subsets, and unloading segment signal subsets that cover the entire pre-bending process.

[0056] S2.4: Perform zero-mean normalization on the initial loading segment signal subset, platform segment signal subset, and unloading segment signal subset respectively to generate standardized initial loading segment data, standardized platform segment data, and standardized unloading segment data that eliminate the influence of sensor baseline drift.

[0057] S2.5: Based on the standardized initial loading segment data, standardized platform segment data, and standardized unloading segment data, perform data structure encapsulation operations to generate a three-segment micro-strain evolution time series segment containing complete phase information. This segment serves as the direct input object for the subsequent feature extraction module.

[0058] like Figure 2 As shown, step S3: Extract the distribution density of the inflection point of the rising slope, the energy proportion of the oscillation band in the plateau segment, and the gradient change parameters of the relaxation decay rate in the unloading segment from the three-segment micro-strain evolution time series segment to generate a dimensionless dynamic response fingerprint. The dimensionless dynamic response fingerprint is used to characterize the intrinsic dynamic response properties of the material. Specifically, it includes: S3.1: Perform a first-order difference operation on the standardized initial loading segment data in the three-segment micro-strain evolution time series segment to generate an instantaneous strain rate sequence, and generate a set of inflection point coordinates of the loading process based on the zero-crossing point and extreme point positions detected by the instantaneous strain rate sequence, and then count the number of inflection points in the unit angle interval to generate the distribution density of the rising slope inflection points.

[0059] S3.1 performs a first-order difference operation on the normalized initial loading segment data in the three-segment micro-strain evolution time series to generate a discrete strain rate sequence characterizing the instantaneous deformation rate. This operation converts the static displacement signal into a dynamic rate signal by calculating the ratio of the strain difference between adjacent sampling points to the time interval, thereby amplifying the nonlinear response characteristics caused by the micro-lattice slip of the material.

[0060] Based on the generated instantaneous strain rate sequence, a sliding extremum detection process is used to identify local maxima and zero-crossing points in the sequence. A dynamic threshold is set to filter out high-frequency noise interference, retaining only standard inflection points whose amplitude exceeds three times the reference noise level, generating a set of inflection point coordinates for the loading process that includes timestamps and strain rate amplitudes. This set accurately locates the critical state of the material transitioning from elastic deformation to plastic flow.

[0061] The distribution of effective inflection points within a unit bending angle interval is statistically analyzed to construct the inflection point distribution density of the rising slope. The inflection point distribution density is calculated using the following formula: Where D is the distribution density at the inflection point of the rising slope, and N cp Δθ represents the total number of effective inflection points within a unit angle interval, and Δθ is the preset width of the unit bending angle interval. This parameter quantifies the degree of non-uniform deformation of the material during the initial loading stage, reflecting the complexity of the distribution of the internal residual stress field.

[0062] Through the above chain derivation process, the time-domain waveform of the initial loading segment is transformed into the distribution density of the rising slope inflection point, which reflects the consistency of the material's microstructure. This enables high-resolution quantitative characterization of the performance fluctuations within a batch of materials, providing core dimensional data for the subsequent construction of a dimensionless dynamic response fingerprint.

[0063] S3.2: Perform Fast Fourier Transform on the standardized plateau segment data in the three-segment micro-strain evolution time series segment to generate a frequency domain power spectral density distribution, and calculate the ratio of the integral energy value to the total energy value within a preset high-frequency bandwidth based on the frequency domain power spectral density distribution to generate the plateau segment oscillation frequency band energy ratio.

[0064] We receive a time series segment of microstrain evolution of the platform segment after zero-mean normalization. This data characterizes the micro-deformation fluctuation characteristics of the material during the stable stage of plastic flow.

[0065] A windowed Fast Fourier Transform is performed on the discrete time series of the platform segment. The Hanning window function is used to suppress spectral leakage, and the time-domain strain signal is mapped to the frequency domain space to generate a set of spectral components in complex form.

[0066] Power spectral density values ​​at each frequency point are calculated based on complex spectral components, and frequency domain power spectral density distribution curves reflecting the distribution characteristics of micro-strain energy at different frequencies are constructed.

[0067] A high-frequency bandwidth threshold is set, and frequency bands above the threshold are identified and extracted. These bands correspond to the high-frequency micro-oscillation components caused by lattice slip and dislocation motion within the material.

[0068] Numerical integration is performed on the power spectral density values ​​within a preset high-frequency bandwidth to calculate the total high-frequency oscillation energy, which is used as the numerator to characterize the dynamic instability of the material.

[0069] The total energy of the signal is calculated by performing a global integral operation on the power spectral density values ​​across the entire frequency band, and is used as the denominator term to characterize the overall response intensity.

[0070] The energy proportion of the oscillation band in the plateau segment is calculated using the following formula: Among them, E ratio P(f) represents the energy proportion of the oscillation band in the plateau segment, and P(f) represents the power spectral density at frequency f. high f is the lower limit threshold for high-frequency bandwidth. max It is half the sampling frequency.

[0071] By using the frequency domain energy ratio calculation method described above, the time domain waveform data from the previous step is transformed into the energy ratio of the dimensionless plateau segment oscillation frequency band, thereby realizing the quantitative characterization of the dynamic stability differences of the internal microstructure of the material and effectively distinguishing the rebound tendency of different batches of materials.

[0072] S3.3: Perform exponential fitting regression analysis on the standardized unloading segment data in the three-segment micro-strain evolution time series segment to generate stress relaxation decay curves, and calculate the ratio of the decay rate difference to the time interval within adjacent time windows based on the stress relaxation decay curves to generate unloading segment relaxation decay rate gradient change parameters.

[0073] Receive the standardized unloading segment data generated in the preceding steps, which contains the strain relaxation time series of the material after the microload is removed.

[0074] A nonlinear least squares exponential fitting method was performed on the standardized unloading segment data to construct a stress relaxation attenuation model to characterize the hysteresis effect of lattice slip inside the material.

[0075] Using a single exponential decay function as the fitting basis function, its mathematical expression is as follows: in, Let be the strain value at time t. denoted as the residual strain reference value, A as the initial relaxation amplitude, k as the material-specific relaxation rate constant, and t as the time variable relative to the unloading start point.

[0076] The parameter k in the above equation is solved by the Levenberg-Marquardt method to obtain a quantitative index reflecting the viscoelastic recovery ability of the material.

[0077] Divide the unloading time axis into N equal-length sliding time windows, and calculate the instantaneous slope of the strain curve within each window: in, Let be the average decay rate of the i-th window. This represents the strain value at the corresponding moment.

[0078] Calculate the difference in decay rate between adjacent time windows and divide it by the time interval to obtain the relaxation decay rate gradient. The specific formula is as follows: Where G is the parameter representing the gradient change of the relaxation decay rate during the unloading phase. This represents the time step of the sliding window.

[0079] The G values ​​over the entire time period are statistically aggregated, and their mean and standard deviation are extracted as the final feature components.

[0080] By using exponential fitting regression and gradient difference operations, the time-domain waveform of the unloading segment is transformed into a dimensionless gradient parameter that characterizes the micro-damping properties of the material, thus achieving high-resolution identification of viscoelastic differences between batches of materials.

[0081] S3.4: Normalize the distribution density of the inflection point of the rising slope, the energy ratio of the oscillation frequency band of the platform segment, and the gradient change parameters of the relaxation decay rate of the unloading segment to generate a set of standardized feature components, and then perform vector concatenation of the set of standardized feature components according to the predefined feature dimension order to generate a multidimensional original dynamic response feature vector.

[0082] The S3.4 sub-step is located at the end of the feature extraction process. It is responsible for converting the multidimensional physical feature parameters generated in the preceding steps into standardized, structured data vectors to eliminate the interference of dimensional differences on the accuracy of subsequent graph neural network clustering.

[0083] Receive the rising slope inflection point distribution density, plateau segment oscillation frequency band energy ratio, and unloading segment relaxation decay rate gradient change parameters from the output of steps S3.1 to S3.3, and construct an original feature set containing multiple components.

[0084] For each feature component, a pre-set global statistical extreme value table is retrieved to obtain the minimum and maximum values ​​recorded in thousands of historical material tests for that feature dimension, which are used as the reference boundary for normalization processing.

[0085] By employing maximum-minimum normalization, a linear mapping transformation is performed on each original feature value, namely the distribution density of the inflection point of the rising slope, the energy proportion of the oscillation frequency band in the platform segment, and the gradient change parameter of the relaxation decay rate in the unloading segment. This compresses the physical quantities in any interval to the standard numerical interval of [0, 1], ensuring the consistency of the numerical magnitude of each feature dimension.

[0086] The specific calculation process follows the following mathematical logic: Where, x i Let x be the original eigenvalue of the i-th dimension. min With x max These are the historical minimum and maximum values ​​for this dimension, respectively. These are the normalized, standardized feature components.

[0087] The above normalization operation is performed sequentially on the feature parameters of multiple dimensions to generate a set of standardized feature components consisting of multiple values ​​in the [0, 1] interval, thus eliminating weight bias caused by differences in sensor sensitivity or different physical dimensions.

[0088] Based on the predefined feature dimension index order, multiple components such as the standardized rising slope inflection point distribution density, the energy proportion of the oscillation frequency band in the platform segment, and the gradient change of the relaxation decay rate in the unloading segment are serially vectorized.

[0089] By using array serialization, discrete standard components are combined into a continuous multidimensional column vector, forming a multidimensional original dynamic response feature vector.

[0090] It is understood that the normalization process is not limited to the maximum-minimum normalization method. In other embodiments of the present invention, alternative methods such as Z-score normalization (also known as standard deviation normalization) or decimal scaling normalization can also be used to standardize the feature parameters. Z-score normalization converts the data into a standard normal distribution with a mean of 0 and a standard deviation of 1 by calculating the difference between each feature parameter value and its mean, divided by the standard deviation. This is suitable for scenarios where the feature parameter distribution approximately follows a Gaussian distribution. Decimal scaling normalization maps the feature parameters to the [-1, 1] or [0, 1] interval by shifting the decimal point position. This is suitable for scenarios with a large range of numerical values. All of the above normalization methods can achieve the technical effect of eliminating the differences in the dimensions of different feature parameters and unifying the numerical scale. Those skilled in the art can make an adaptive selection based on the actual data distribution characteristics and engineering requirements.

[0091] By using maximum and minimum value normalization and vector concatenation, the multi-source heterogeneous physical parameters from the previous step are transformed into dimensionless, same-order multidimensional original dynamic response feature vectors, achieving standardized encapsulation of feature data and providing a unified input format for the subsequent initialization of node attributes in graph neural networks.

[0092] S3.5: Perform dimensionless mapping processing on the multidimensional original dynamic response feature vector to generate a dimensionless dynamic response fingerprint, wherein the dimensionless dynamic response fingerprint is directly input into the subsequent graph neural network model as the final data form characterizing the intrinsic dynamic response properties of the material.

[0093] like Figure 3 As shown, step S4: Input the dimensionless dynamic response fingerprint into the graph neural network model trained offline, and output the material performance cluster label identifying the fine-grained performance cluster to which the material's current state belongs, along with the corresponding springback tendency coefficient. The fine-grained performance cluster refers to a material performance sub-category obtained by clustering historical samples based on the material's micro-strain dynamic response characteristics using a graph neural network model. Each fine-grained performance cluster corresponds to a group of material batches or states with similar dynamic response characteristics. The acquisition method is as follows: Using the graph neural network model trained offline, the input dimensionless dynamic response fingerprint is mapped to a high-dimensional embedding space. Then, the probability distribution of the fingerprint belonging to each preset material performance cluster is calculated using a fully connected layer and a softmax classifier. The cluster with the highest probability is determined as the fine-grained performance cluster of the current material, and the corresponding cluster label is output. Specifically, this includes: S4.1: Construct a dynamic response feature vector based on the dimensionless dynamic response fingerprint, and use a preset feature node mapping rule to convert the value of each dimension in the dynamic response feature vector into the initial node attribute value in the graph neural network, thereby generating an initial feature node set containing multiple feature nodes.

[0094] The feature node mapping rule refers to the mapping criterion that establishes a one-to-one correspondence between each feature dimension in the dimensionless dynamic response fingerprint and a specific feature node in the graph neural network. This feature node mapping rule is pre-set during the offline training phase of the model. Specifically, the dimensionless dynamic response fingerprint is sequentially numbered from dimension 1 to dimension 12, and multiple feature nodes in the graph neural network are sequentially numbered from N1 to N12, establishing a fixed correspondence between dimension numbers and node numbers, i.e., the i-th feature value is mapped to node N. i The feature node mapping rule remains unchanged after model deployment, ensuring that features with the same physical meaning are always input to the same feature node during each inference, thus maintaining the input consistency of graph structure data.

[0095] Specifically, the dimensionless dynamic response fingerprint generated by step S3 is received. This fingerprint consists of characteristic components such as the distribution density of the inflection point of the rising slope, the energy ratio of the oscillation frequency band of the plateau segment, and the gradient change of the relaxation decay rate of the unloading segment, which serve as the initial input data source for the graph neural network model.

[0096] The dimensionless dynamic response fingerprint is parsed into an independent scalar sequence. Based on the preset feature node mapping rules, a one-to-one correspondence between fingerprint dimensions and graph neural network nodes is established to ensure that each physical feature dimension is uniquely mapped to a specific feature node.

[0097] A linear transformation is performed on the numerical values ​​of each dimension, mapping them to the numerical range that the graph neural network activation function is sensitive to, generating initial node attribute values ​​with a uniform scale, thus eliminating the impact of differences in physical magnitudes on the network convergence speed.

[0098] An initial set of feature nodes containing multiple independent elements is constructed, where each node carries a standardized attribute vector that fully preserves the morphological feature information of the corresponding micro-strain response stage.

[0099] The initial set of feature nodes is pre-encoded with topological structure, and the message passing priority of each node in the subsequent graph convolution process is marked, providing a structured data foundation for constructing a dynamic adjacency matrix that characterizes the internal stress coupling relationship of the material.

[0100] Through the above feature mapping and node initialization processing, the dimensionless dynamic response fingerprint from the previous step is transformed into an initial feature node set containing multiple feature nodes, realizing the paradigm shift from static feature vectors to graph structure data, and providing a standardized input object for subsequent identification of material property clusters based on graph topology relationships.

[0101] S4.2: Based on the physical coupling relationship between each feature node in the initial feature node set, perform a dynamic adjacency matrix construction operation, define edge connection weights that characterize the strong correlation between the distribution density of the inflection point of the rising slope and the energy ratio of the oscillation frequency band of the plateau segment, and generate a dynamic adjacency matrix that characterizes the stress transmission topology of the material.

[0102] Receive an initial set of feature nodes generated by S4.1, which contains multiple feature nodes, each of which carries a dimensionless micro-strain dynamic response feature component.

[0103] Based on the constitutive relations of materials mechanics and the microscopic lattice slip mechanism, a physical coupling association rule library between characteristic nodes is constructed. A strong positive correlation is defined between the distribution density node of the inflection point of the rising slope and the energy ratio node of the oscillation frequency band of the plateau segment, which is defined as the energy dissipation mapping relationship during the transition from the elastic stage to the plastic stage.

[0104] Initialize a zero matrix with a dimension of 12×12 as the basic structure of the dynamic adjacency matrix, where the row index and column index correspond to the numbers of multiple feature nodes, respectively.

[0105] Iterate through all node pairs in the initial feature node set. For node pairs with predefined physical coupling relationships, calculate the edge connection weights based on the strength of their correlation.

[0106] The Pearson correlation coefficient is used to quantify the degree of linear correlation between nodes. The edge connection weight between the i-th node and the j-th node is calculated using the following formula: Among them, w ij Let x be the edge connection weight between node i and node j. ik Let i be the feature value of node i in the k-th sample. Let be the feature mean of node i, and n be the number of historical training samples. Let j be the feature mean of node j. Let be the feature value of node j in the kth sample.

[0107] The calculated edge connection weights are filled into the corresponding positions in the dynamic adjacency matrix. For node pairs without direct physical coupling, the weights are kept to zero or given a very small default connectivity value to maintain the connectivity of the graph.

[0108] The generated dynamic adjacency matrix is ​​symmetric to ensure that the matrix satisfies A ij =A ji This reflects the bidirectional equilibrium characteristics of stress transmission within the material.

[0109] The dynamic adjacency matrix is ​​standardized by normalized Laplace transform to eliminate the influence of node degree differences on subsequent graph convolution aggregation operations, thus generating the final dynamic adjacency matrix.

[0110] Through the above processing method, the discrete feature nodes of the previous step are transformed into a dynamic adjacency matrix containing the internal stress transmission topology of the material, realizing a structured representation of the intrinsic correlation of the dynamic properties of the material, and providing an accurate topological basis for the message transmission of the subsequent graph neural network.

[0111] S4.3: Perform multi-layer graph convolution aggregation processing on the initial feature node set using the dynamic adjacency matrix, fuse the local response information of adjacent feature nodes through a message passing mechanism and update the hidden layer state representation of the central feature node to generate a deep node embedding vector containing global material dynamic property context information.

[0112] The system receives the initial set of feature nodes containing multiple feature nodes and the dynamic adjacency matrix representing the stress transfer topology generated in the previous steps, and uses them as the input data source for the first layer convolution operation of the graph neural network.

[0113] For any i-th feature node in the graph, based on the connection relationship determined by the dynamic adjacency matrix, retrieve the set of all its directly connected neighbor nodes, and extract the current hidden layer state vector and connection edge weight coefficient of each neighbor node.

[0114] A weighted summation aggregation operator is used to perform element-wise multiplication of the state vector of each neighbor node with its corresponding edge weight coefficient, and the results are accumulated to generate an aggregated message vector that reflects the local neighborhood structure information.

[0115] A learnable linear transformation matrix is ​​introduced to perform a linear mapping transformation on the original feature vector of the current central node, generating self-loop feature components that preserve the node's own attributes.

[0116] The aggregated message vector is concatenated or fused with the self-loop feature components to construct an intermediate layer joint feature representation that includes its own attributes and neighborhood interaction information.

[0117] The intermediate layer joint feature representation is subjected to element-wise threshold filtering using the nonlinear activation function ReLU to suppress negative noise interference and enhance the sparsity and nonlinear discrimination ability of the feature representation.

[0118] Repeat the above neighbor information aggregation, linear transformation and nonlinear activation operations, and sequentially pass through the preset three-layer graph convolutional network layers to expand the receptive field from direct neighbors to second-order and third-order neighbor nodes.

[0119] After each convolutional layer, a batch normalization layer is introduced to standardize the node embedding vectors, eliminating feature distribution shifts caused by batch differences in materials and accelerating model convergence.

[0120] By using the message passing mechanism of multi-layer graph convolution, the micro-strain response fingerprint features scattered across multiple dimensions are transformed into deep node embedding vectors in a high-dimensional space, thereby achieving global contextual awareness of the micromechanical behavior inside the material.

[0121] For example, the graph neural network model is set to contain three graph convolutional layers, with each hidden unit having a dimension of 64. For SPCC material samples, the initial node feature vector is 12-dimensional. In the first convolutional layer, node 1 (inflection point density of the rising slope) aggregates information from its neighbor node 2 (plateau oscillation energy), with the edge weight set to 0.85. After a linear transformation W1 (64x12 matrix) and ReLU activation, a 64-dimensional intermediate vector is obtained. After three layers of aggregation, a 64-dimensional deep node embedding vector is finally output. This vector effectively integrates the coupled features of the loading, plateau, and unloading stages. Compared to a single-layer fully connected network, it significantly improves the sensitivity to material yield strength fluctuations, ensuring the accuracy of subsequent classification labels.

[0122] S4.4: Perform a fully connected layer classification mapping operation based on the deep node embedding vector, call the softmax activation function trained offline to calculate the probability distribution value of the deep node embedding vector belonging to each preset material performance cluster, and generate a material performance cluster label with the highest identification probability for the target material performance cluster.

[0123] The system receives deep node embedding vectors generated by graph convolution aggregation, which contain global topological features and local dynamic coupling information of the material's micro-strain response, and serves as the input data source for classification mapping.

[0124] Construct a weight matrix for the fully connected layer, and linearly map the multidimensional feature space of the deep node embedding vectors to the classification space of the preset material property cluster, thereby realizing the dimensional transformation from high-dimensional hidden layer representation to low-dimensional class probability space.

[0125] The softmax activation function is applied to the mapped logits values. The dimensional differences are eliminated and the numerical range is compressed through exponential normalization. The non-negative probability distribution values ​​of the deep node embedding vectors belonging to each preset material property cluster are calculated.

[0126] The probability of belonging to the k-th material property cluster is calculated using the following formula: Among them, P k z represents the probability of belonging to the k-th material property cluster. k Let be the k-th logits value output by the fully connected layer, C be the total number of preset material property clusters, j be the index variable for traversing all property clusters, and z be the index variable. j This is the j-th logits value output by the fully connected layer.

[0127] Traverse the calculated probability distribution vector, perform a maximum value search operation, and locate the target index position with the highest probability value. This index corresponds to the fine-grained performance cluster with the highest matching degree with the current material dynamic response fingerprint.

[0128] Extract the performance cluster identifier corresponding to the maximum probability value to generate a material performance cluster label. At the same time, record the maximum probability value as a classification confidence score for subsequent weighted correction of the rebound tendency coefficient.

[0129] By using softmax probability distribution calculation and maximum value retrieval processing, the deep node embedding vector from the previous step is transformed into a material performance cluster label that identifies the target material performance cluster with the highest probability, thereby achieving the expected technical effect of high-resolution identification and classification of performance fluctuations within a batch of materials.

[0130] S4.5: Based on the material property cluster label, retrieve the pre-stored cluster center parameter table, extract the standard rebound behavior benchmark value uniquely bound to the material property cluster label, and perform linear weighted correction by combining the confidence score of the deep node embedding vector to generate a rebound tendency coefficient for guiding the calculation of compensation angle.

[0131] The cluster center parameter table refers to a pre-set data table that stores the one-to-one mapping relationship between each preset material property cluster and the standard rebound behavior benchmark value. It is obtained as follows: During the offline training phase, based on the dimensionless dynamic response fingerprint of historical samples, each material property cluster is obtained through clustering using a graph neural network model. Then, the measured rebound angle data of all historical samples within each cluster under standard operating conditions are statistically analyzed, and their mean is calculated as the standard rebound behavior benchmark value for that cluster. This benchmark value is then bound and stored with the corresponding material property cluster label to form the cluster center parameter table.

[0132] The system receives material property cluster labels and deep node embedding vectors generated by the classification mapping operation of the fully connected layer as input data sources for calculating the springback tendency coefficient.

[0133] Based on the material property cluster label, a key-value indexing operation is performed in the pre-stored cluster center parameter table to locate the standard rebound behavior benchmark value uniquely bound to the label. This benchmark value characterizes the average rebound characteristics of this type of material under ideal working conditions.

[0134] The activation probability value corresponding to the target classification node is extracted from the deep node embedding vector and defined as the confidence score of the current recognition result, which is used to quantify the degree of matching between the dynamic response fingerprint and the standard cluster center.

[0135] A linear weighted correction model is constructed, using the standard rebound behavior benchmark as the base component and the confidence score as the adjustment weight factor. The rebound tendency coefficient is calculated using the following formula: Where β is the rebound tendency coefficient, R base λ is the baseline value for standard rebound behavior, λ is the preset sensitivity adjustment coefficient, and C is the confidence score.

[0136] The calculated rebound tendency coefficient is numerically validated to ensure that it is within the physically feasible rebound compensation range, and outlier values ​​caused by extreme noise are removed.

[0137] By using a linear weighted correction method, the material property cluster labels and deep node embedding vectors from the previous step are transformed into a rebound tendency coefficient with confidence weighting, thereby achieving adaptive adaptation of the compensation angle calculation to the reliability of material dynamic characteristic identification.

[0138] Step S5: Based on the material property cluster tags, retrieve the median value of the optimal compensation angle interval directly associated with the material property cluster tags from the pre-set cluster-compensation mapping knowledge base. The cluster-compensation mapping knowledge base is a structured database storing closed-loop bending verification data of historical materials within the same cluster under the same equipment operating conditions. It is constructed from a historical bending process database trained and verified offline. Using the material property cluster tags as index keys, this knowledge base groups and aggregates different material batches, mold conditions, and corresponding measured springback compensation data in multiple dimensions according to material grade, plate thickness, bending angle, and mold V-groove width. This forms key-value pair objects containing the boundary values ​​of the optimal compensation angle interval, which are then loaded into the read-only memory of the embedded controller for real-time querying and retrieval.

[0139] Specifically, it includes: S5.1: Initialize and load the pre-set cluster-compensation mapping knowledge base to construct a multi-dimensional index structure containing closed-loop bending verification data of historical materials in the same cluster under the same equipment operating conditions, thereby generating a cluster-compensation mapping knowledge base instance that can be queried in real time.

[0140] The S5.1 sub-step aims to initialize the cluster-compensation mapping knowledge base, providing data support for subsequent rapid retrieval based on material property cluster tags. It receives a historical bending process database that has been trained and validated offline, wherein the historical bending process database contains different material batches, die conditions, and corresponding measured springback compensation data. After performing data cleaning and outlier removal operations on the different material batches, mold conditions and corresponding measured springback compensation data, the data are grouped in multiple dimensions according to material grade, sheet thickness, bending angle and mold V-groove width. For each group, the compensation angle records of materials in the same cluster during the hundred closed-loop bending verification are extracted, their distribution characteristics are calculated, and a cluster-level statistical summary is formed. A hash mapping process is used to generate a unique identifier key for each material property cluster, and the corresponding cluster-level statistical summary and the boundary value of the optimal compensation angle interval are encapsulated as key-value pair objects. All key-value pair objects are loaded into the read-only memory of the embedded controller to construct the cluster-compensation mapping knowledge base instance. The optimal compensation angle interval boundary value refers to the minimum and maximum compensation angle values ​​retrieved from the cluster-compensation mapping knowledge base, corresponding to the current material property cluster label. These two values ​​together define a compensation angle interval characterizing the springback tolerance range of the current material batch. The acquisition method is as follows: based on the knowledge base retrieval pointer, the underlying data table of the cluster-compensation mapping knowledge base is accessed, historical data record entries matching the material property cluster label are located, and springback deviation distribution boundary values ​​at a pre-set confidence level are extracted. The lower threshold representing the minimum springback compensation requirement and the upper threshold representing the maximum springback compensation requirement are obtained respectively. Then, the lower and upper thresholds are converted into absolute angular offsets relative to the basic bending angle, thereby generating the optimal compensation angle interval boundary value.

[0141] Through the above processing method, discrete historical verification data is transformed into structured, real-time queryable knowledge base instances, achieving the expected technical effect of low latency and high reliability in compensation strategy generation.

[0142] S5.2: Based on the material property cluster label, perform key-value matching retrieval processing on the cluster-compensation mapping knowledge base instance to locate historical data record entries that are completely consistent with the material property cluster label, thereby generating a knowledge base retrieval pointer pointing to a specific optimal compensation angle range.

[0143] The system receives material property cluster tags output from step S4. These tags are discrete integer identifiers representing the cluster affiliation of the current hardware material in the dynamic response feature space. The material property cluster tags are used as primary key index values ​​and input into the initialized and loaded cluster-compensation mapping knowledge base retrieval engine. The retrieval engine internally maintains a multi-dimensional index database based on a hash table structure, where key-value pairs consist of material property cluster tags and corresponding historical process data record pointers. Hash collision detection and address mapping operations are performed to calculate the storage location of the target tag in the hash bucket, quickly locating the set of all historical data record entries containing that tag. The located historical data record entries undergo integrity verification, confirming whether each record contains a complete statistical distribution of bending angle deviation, compensation angle adjustment history, and yield rate markers. Historical records with missing data or marked as abnormal operating conditions are removed to ensure that the source data participating in subsequent calculations has high confidence. Memory address pointers pointing to the optimal compensation angle interval are extracted from the verified historical data records. These pointers are directly associated with the interval boundary parameter blocks stored in the cache. By using hash index matching, the material property cluster tags from the previous step are transformed into knowledge base retrieval pointers pointing to specific optimal compensation angle intervals, achieving millisecond-level precision in recalling historical process knowledge and providing accurate data addressing basis for subsequent analysis of compensation angle intervals.

[0144] S5.3: Use the knowledge base retrieval pointer to extract the boundary data of the optimal compensation angle interval stored in the historical data record entries, so as to parse out the minimum compensation angle value and the maximum compensation angle value corresponding to the material performance cluster, thereby generating a set of optimal compensation angle intervals that characterize the rebound tolerance range.

[0145] Receive the knowledge base retrieval pointer generated in the previous step, which points to a specific optimal compensation angle range. This knowledge base retrieval pointer contains a unique identifier for the target material property cluster and its corresponding storage address index.

[0146] Based on the knowledge base retrieval pointer, access the underlying data table of the cluster-compensation mapping knowledge base, locate the historical data record entry that completely matches the unique identifier of the material property cluster label, and read the cleaned and statistically aggregated historical dataset of rebound deviation stored therein based on the corresponding storage address index.

[0147] Extract the rebound deviation distribution boundary values ​​under the preset confidence level from the historical data record entries, and obtain the lower threshold representing the minimum rebound compensation requirement and the upper threshold representing the maximum rebound compensation requirement, respectively.

[0148] The lower and upper thresholds are converted into absolute angular offsets relative to the base bending angle, and the boundary values ​​of the optimal compensation angle range are calculated using the following formula: Where, θ min For the minimum compensation angle value, θ max For the maximum compensation angle value, θ base The basic bending angle, Δθ lower Δθ is the lower limit of the rebound deviation threshold. upper The upper limit is the rebound deviation threshold.

[0149] The calculated minimum and maximum compensation angle values ​​are validated to eliminate non-physical extreme values ​​caused by historical data noise, ensuring that the range conforms to the current mold geometric constraints and material plastic deformation limits.

[0150] The verified minimum and maximum compensation angle values ​​are encapsulated into structured data objects to generate a set of optimal compensation angle intervals that characterize the rebound tolerance range of the current material batch.

[0151] Through the above data extraction and boundary analysis processing, the knowledge base retrieval results of the previous step are transformed into specific numerical compensation angle interval boundaries, realizing the quantitative definition of the fluctuation range of material dynamic characteristics, and providing data support for the subsequent determination of accurate compensation median value.

[0152] S5.4: Perform an arithmetic mean calculation on the set of optimal compensation angle intervals to eliminate random fluctuations in the upper and lower limits of the intervals and determine the center position of the intervals, thereby generating the median value of the optimal compensation angle interval for driving the upper die to perform the formal bending action.

[0153] Step S6: Superimpose the median value of the optimal compensation angle interval onto the preset output value of the basic compensation model to generate an adaptive dynamic compensation angle command for driving the upper die to perform the formal bending action. Specifically, this includes: S6.1: Obtain the preset basic compensation model output value. The basic compensation model output value is a static reference compensation amount calculated based on the target bending angle of the current hardware, the thickness of the sheet metal, and the geometric parameters of the mold through the standard springback formula, and is used as a fixed reference dependent variable for subsequent superposition calculations.

[0154] The current processing task instruction is analyzed, and key geometric constraint parameters such as the target bending angle, nominal thickness of the sheet metal, width of the mold V-groove, and opening size of the lower mold are extracted to construct the initial input vector for static compensation calculation.

[0155] The system calls the material mechanical properties database pre-stored in the controller's non-volatile memory, retrieves the corresponding elastic modulus, yield strength, and hardening index based on the workpiece material grade, and uses these as the physical constants basis for theoretical springback calculation.

[0156] Based on the Kisssinger springback theory model, a geometric mapping relationship including the material's plastic deformation zone and elastic recovery zone is established, and the theoretical springback angle is calculated using the following formula: Where Δθ is the theoretical springback angle, K is the material hardening coefficient, t is the plate thickness, E is the elastic modulus, R is the inner radius of the bend, and V is the width of the V-groove of the lower die.

[0157] Substitute the retrieved physical constants into the above formula and perform floating-point operations to calculate the theoretical springback compensation amount that depends only on macroscopic static parameters. This value represents the expected deviation of an ideal homogeneous material under standard operating conditions.

[0158] Perform boundary rationality verification on the calculated theoretical rebound compensation amount, compare it with the maximum allowable compensation stroke and minimum resolution limit of the equipment, and eliminate non-physical extreme values ​​caused by abnormal parameters.

[0159] By using the standardized formulaic calculation method described above, discrete processing geometric parameters and material static properties are transformed into definite static benchmark compensation quantities, realizing the theoretical quantification of the basic springback trend in the bending process and providing a stable reference zero point for subsequent dynamic characteristic correction.

[0160] S6.2: Call the median value of the optimal compensation angle interval generated in the previous step. The median value of the optimal compensation angle interval is a dynamic correction factor retrieved from the cluster compensation mapping knowledge base based on the material property cluster label. It is used to characterize the rebound deviation trend of the current batch of materials relative to the standard material.

[0161] Receive the median value of the optimal compensation angle range generated in step S5.4. This value represents the springback correction benchmark of the current material batch under specific working conditions.

[0162] The median value of the optimal compensation angle range is read from the shared memory buffer of the embedded controller, and a data type check is performed to ensure that it is in double-precision floating-point format.

[0163] The read median value is assigned to the dynamic correction factor variable, which serves as a floating component in subsequent superposition operations to offset systematic deviations caused by fluctuations in material properties.

[0164] Perform validity boundary checks on the dynamic correction factor to verify whether its value is within the preset physical reasonable range, and prevent extreme value injection caused by abnormal knowledge base retrieval.

[0165] If the verification passes, the dynamic correction factor is locked as the effective correction parameter for the current bending cycle, and its source is marked as the adaptive calculation result based on micro-strain fingerprint recognition.

[0166] By using the above data retrieval and verification methods, the knowledge base retrieval results from the previous step are transformed into dynamic correction factors that can be directly referenced by the control system, enabling the compensation strategy to accurately respond to performance fluctuations within material batches and significantly improving the consistency of bending angles.

[0167] S6.3: Perform linear superposition operation based on the static reference compensation amount and the dynamic correction factor, and seamlessly integrate the dynamic correction factor, which reflects the intrinsic dynamic response characteristics of the material, into the static reference compensation amount through algebraic summation to generate a preliminary synthetic compensation angle value containing real-time fluctuation information of the material.

[0168] The system receives the static reference compensation amount and the median value of the optimal compensation angle interval output from the preceding steps. The former serves as the theoretical calculation basis based on geometric parameters, while the latter serves as a real-time correction factor based on the dynamic properties of the material.

[0169] Perform data type verification on the static baseline compensation amount to confirm that it is a floating-point value and is within the allowable physical travel range of the device, to prevent calculation overflow caused by data abnormalities.

[0170] Read the median value of the optimal compensation angle range. This value represents the intrinsic dynamic response deviation trend of the current batch of materials in the sub-rebound critical region and has the real-time characteristic of high-frequency updates.

[0171] A linear superposition operation model is constructed, in which the static benchmark compensation amount is defined as a fixed reference dependent variable, and the median value of the optimal compensation angle interval is defined as a dynamic correction independent variable.

[0172] The fusion calculation is performed using algebraic summation, and the preliminary composite compensation angle values ​​are generated using the following formula: Where, θ syn To initially synthesize the compensation angle value, θ base Δθ is the static reference compensation value. mid This is the median value of the optimal compensation angle interval.

[0173] The calculated preliminary composite compensation angle values ​​are truncated to three decimal places to match the control resolution requirements of the servo drive system.

[0174] By using a linear superposition processing method, the dynamic correction factor, which reflects the intrinsic dynamic response characteristics of the material, is seamlessly integrated into the static benchmark compensation amount, generating a preliminary synthetic compensation angle value that includes real-time fluctuation information of the material, thus realizing the expected technical effect of transforming the compensation strategy from static preset to dynamic adaptive.

[0175] S6.4: Perform amplitude limiting filtering and smoothing processing on the preliminary synthesized compensation angle value, use the preset physical travel constraint threshold to remove abnormal jump components that exceed the safety range of the equipment, and use a first-order inertial filtering algorithm to eliminate high-frequency noise interference, and generate stable intermediate compensation angle data that meets the response characteristics of the actuator.

[0176] Receive the preliminary synthetic compensation angle value output from step S6.3. This value integrates the static reference and dynamic correction factor, but may contain non-physical high-frequency jitter components caused by sensor noise or model transient response.

[0177] Define the physical travel safety boundaries of the servo drive system, including the maximum permissible positive compensation angle threshold and the maximum permissible negative compensation angle threshold, and construct the upper and lower limit constraint ranges of the limiting filter to prevent the actuator from mechanical interference or overload damage due to overshoot commands.

[0178] The preliminary synthesized compensation angle value is logically compared with the preset safety boundary threshold. If the value exceeds the upper limit, it is forcibly truncated to the upper limit value; if it is lower than the lower limit, it is forcibly truncated to the lower limit value, thus generating the amplitude-limited compensation angle data that conforms to mechanical safety specifications.

[0179] To address the high-frequency random noise remaining in the compensated angle data after amplitude limiting, a first-order inertial filtering algorithm is used for smoothing, and a recursive relationship is established between the current output value, the previous output value, and the current input value.

[0180] The stable intermediate compensation angle data is calculated using the following formula: Where, θ stable For the stable intermediate compensation angle data at the current moment, θ limited θ represents the current moment's compensated angle data after amplitude limiting. prev The stable intermediate compensation angle data of the previous moment is α, which is the filtering coefficient, with a value between 0 and 1, used to balance the system response speed and smoothness.

[0181] Based on the dynamic response characteristics and sampling period of the bending machine servo motor, the value of the filtering coefficient α is set, usually between 0.2 and 0.4, in order to effectively suppress high-frequency interference while ensuring real-time performance.

[0182] The calculated stable intermediate compensation angle data is used as the final output. This data eliminates abnormal jumps and high-frequency noise, and has continuity and stability.

[0183] By using amplitude limiting filtering and first-order inertial smoothing, the preliminary synthetic compensation angle value from the previous step is transformed into stable intermediate compensation angle data that meets the response characteristics of the actuator, thereby achieving the expected technical effect of improving the stability of bending action and protecting the mechanical structure.

[0184] S6.5: Based on the stable intermediate compensation angle data, perform control command encoding conversion processing, map it into a digital pulse sequence or analog voltage signal that can be recognized by the servo drive system, and generate an adaptive dynamic compensation angle command for directly driving the upper mold to perform the formal bending action.

[0185] Step S7: Using the measured springback deviation data after executing the adaptive dynamic compensation angle command and completing the formal bending action, perform online weight fine-tuning on the graph neural network model to generate updated model weight parameters and improve the material performance cluster label recognition accuracy of subsequent materials in the same batch. Specifically, this includes: S7.1: Obtain the actual bending angle data and the target bending angle command collected by the angle sensing unit after executing the adaptive dynamic compensation angle command and completing the formal bending action. Calculate and generate a measured springback deviation vector characterizing the springback characteristics of the current material batch based on the difference between the two, and use it as a reference error signal for model optimization.

[0186] Using a high-precision absolute encoder or laser triangulation sensor, after the hardware part completes the formal bending action and unloads and holds pressure, the actual forming angle value under the closed state of the mold is collected in real time to form the original angle measurement dataset.

[0187] The original angle measurement dataset is subjected to moving average filtering to remove instantaneous abnormal jump points caused by mechanical vibration or sensor noise, and the true value of the actual bending angle that is stable and converges is extracted.

[0188] The target bending angle command value is read from the control system register and sent to the servo drive. This command value contains the adaptive dynamic compensation angle correction amount generated in the previous steps, which represents the theoretically expected ideal forming angle.

[0189] Construct a model for calculating the rebound deviation based on actual measurements, and substitute the true value of the actual bending angle and the commanded value of the target bending angle into the deviation calculation formula: Where δ is the measured rebound deviation vector, θ actual θ is the true value of the actual bending angle after filtering. target The target bending angle command value.

[0190] Based on the calculated measured rebound deviation value, and combined with the multi-channel micro-strain response fingerprint features collected during the pre-bending stage, a multi-dimensional error signal vector containing deviation amplitude, deviation direction, and confidence weight is constructed.

[0191] The multidimensional error signal vector is normalized and mapped to the interval [-1, 1] to eliminate the dimensional differences caused by sensors with different ranges, and to generate the benchmark error gradient input for subsequent backpropagation algorithm training.

[0192] Through the above data processing flow, the bending results in the physical world are quantified into digital error feedback signals, realizing the construction of a closed-loop data link from the execution result to the model optimization input, and providing high-precision supervision labels for online weight fine-tuning of lightweight graph neural networks.

[0193] S7.2: Based on the measured rebound deviation vector and the material property cluster labels and rebound tendency coefficients output in step S4, the gradient of the loss function is calculated using the backpropagation algorithm to generate the error sensitivity mapping matrix of the connection weights of each layer node in the pointing graph neural network model.

[0194] The measured rebound deviation vector, material property cluster labels, and rebound tendency coefficient are received as error feedback input sources.

[0195] A mean squared error loss function is constructed with the difference between the predicted compensation angle and the measured angle as its core, to quantify the degree of prediction deviation of the model in the current state.

[0196] The loss function L is defined using the following formula: Where N is the number of samples, y i Let y be the predicted rebound value for the i-th sample.des This corresponds to the measured rebound value of the target.

[0197] The partial derivative of the loss function with respect to the connection weights of each layer in the lightweight graph neural network is calculated, and the error signal is backpropagated layer by layer using the chain rule.

[0198] Calculate the gradient contribution of the output layer nodes to the activation values ​​of the hidden layer nodes, and generate the local error sensitivity component.

[0199] The local error sensitivity component is multiplied by the dynamic response fingerprint feature of the input layer to obtain the weight gradient matrix from the input layer to the hidden layer.

[0200] By aggregating the gradient components calculated at all levels, a complete error sensitivity mapping matrix covering the entire network topology is constructed.

[0201] By using the backpropagation algorithm, the measured deviations from the previous step are transformed into error sensitivity mapping matrix data pointing to the weights of each layer, thereby achieving the expected technical effect of accurately locating the direction of model parameter optimization.

[0202] S7.3: Perform adaptive learning rate constraint processing on the error sensitivity mapping matrix, and combine momentum optimization strategy to generate incremental adjustment step size for the hidden layer weights of the graph neural network model, so as to suppress the drastic weight oscillation caused by single sample noise.

[0203] Receive the error sensitivity mapping matrix generated in step S7.2, which contains gradient information of the connection weights of each layer node in the lightweight graph neural network relative to the measured bounce deviation.

[0204] Perform element-wise absolute value calculation on the error sensitivity mapping matrix to generate a sensitivity magnitude matrix that characterizes the contribution of each weight parameter to the current prediction error.

[0205] The global average sensitivity threshold is calculated based on the sensitivity magnitude matrix to distinguish between high-sensitivity weights and low-sensitivity weights, thus avoiding drastic updates to noise-sensitive parameters.

[0206] An adaptive learning rate decay mechanism is introduced, which dynamically calculates the base learning rate coefficient based on the ratio of the current iteration number to the preset maximum iteration period, ensuring that the model converges quickly in the early stage of training and is finely adjusted in the later stage.

[0207] By combining the momentum optimization strategy, the weight update history vector of the previous iteration cycle is read, multiplied by the preset momentum factor, and a momentum compensation term with inertial characteristics is generated.

[0208] The momentum compensation term is vector-superimposed with the current gradient vector scaled by the adaptive learning rate to generate a comprehensive incremental adjustment step size that includes historical update trends and current error guidance.

[0209] The comprehensive incremental adjustment step size is subjected to norm constraint processing. If its Euclidean norm exceeds the preset safe update threshold, the step size modulus is reduced proportionally to prevent the model from diverging due to excessive single update magnitude.

[0210] By using the adaptive learning rate constraint and momentum fusion processing method described above, the error sensitivity mapping matrix of the previous step is transformed into a lightweight graph neural network hidden layer weight increment adjustment step size that suppresses noise interference and is directionally stable, thereby improving the robustness and optimizing the convergence speed during the online fine-tuning process of the model.

[0211] S7.4: The original connection weights of the graph neural network model are iteratively updated using the incremental adjustment step size to generate updated model weight parameters containing the latest material dynamic characteristics, thereby completing the real-time transfer of knowledge state within the model.

[0212] The incremental adjustment step size of the hidden layer weights of the graph neural network model generated by the adaptive learning rate constraint and momentum optimization strategy is received. This incremental adjustment step size serves as the core driving variable for the weight update operation and is directly related to the correction magnitude of the model to the current material dynamic property error.

[0213] The original connection weight matrix between nodes in each layer of the graph neural network model is extracted. The original connection weight matrix represents the strength of the mapping relationship between the micro-strain response characteristics of the material and the performance cluster label in the previous inference cycle, and serves as the benchmark parameter set for this iteration update.

[0214] Based on the incremental adjustment step size and the original connection weight matrix, perform element-wise algebraic addition to calculate the updated connection weights using the following formula: in, For the updated connection weights, These are the original connection weights. Adjust the step size for incremental changes.

[0215] The updated connection weights generated by the operation are subjected to numerical stability verification. The maximum absolute value of the weights is limited by a preset weight pruning threshold to prevent gradient explosion caused by excessive noise in a single sample, and to ensure that the model parameters fluctuate within a reasonable physical range.

[0216] The updated connection weights, after stability verification, are remapped into the adjacency matrix and node attribute transformation matrix of the graph neural network, replacing the original static weight configuration, thus completing the real-time migration of the model's internal knowledge state from the previous time step to the current time step.

[0217] By using the above-mentioned weight iterative update processing method, the incremental adjustment step size generated in the previous step is transformed into updated model weight parameters that include the latest material dynamic characteristics, thereby realizing the continuous improvement of the adaptive capture and recognition accuracy of lightweight graph neural networks for the performance fluctuations of the same batch of materials.

[0218] S7.5: Load the updated model weight parameters into the inference engine of the graph neural network model, replace the original static weight configuration, and generate an enhanced material property cluster label recognition model with online self-learning capabilities to ensure that the material property cluster label recognition accuracy of subsequent batches of materials is improved piece by piece.

[0219] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0220] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0221] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic compensation control of bending angle of hardware parts, characterized in that, Specifically, it includes: S1: Acquire the original multi-channel micro-strain-time response signal of the hardware component during the pre-bending stage by a distributed micro-strain sensor array; S2: Based on the original multi-channel micro-strain-time response signal, perform sliding window segmentation processing to generate a three-segment micro-strain evolution time series segment; S3: Extract the distribution density of the inflection point of the rising slope, the energy ratio of the oscillation frequency band of the plateau segment, and the gradient change parameters of the relaxation decay rate of the unloading segment from the three-segment micro-strain evolution time series segment to generate a dimensionless dynamic response fingerprint. S4: Input the dimensionless dynamic response fingerprint into the graph neural network model trained offline, and output the material property cluster label and the corresponding springback tendency coefficient; S5: Based on the material property cluster label, retrieve the preset cluster-compensation mapping knowledge base to obtain the median value of the optimal compensation angle interval directly associated with the material property cluster label; S6: The median value of the optimal compensation angle interval is superimposed on the preset basic compensation model output value to generate an adaptive dynamic compensation angle command for driving the upper mold to perform the formal bending action.

2. The method for dynamic compensation control of bending angle of hardware parts according to claim 1, characterized in that, After step S6, the following also includes: S7: Use the measured springback deviation data after executing the adaptive dynamic compensation angle command and completing the formal bending action to perform online weight fine-tuning on the graph neural network model.

3. The method for dynamic compensation control of bending angle of hardware parts according to claim 1, characterized in that, The three-segment micro-strain evolution time series segment includes standardized initial loading segment data, standardized platform segment data, and standardized unloading segment data.

4. The method for dynamic compensation control of bending angle of hardware parts according to claim 3, characterized in that, Step S3 specifically includes: The first-order difference operation is performed on the standardized initial loading segment data to generate an instantaneous strain rate sequence. Based on the instantaneous strain rate sequence, the positions of zero-crossing points and extreme points are detected to generate a set of inflection point coordinates in the loading process. Then, the number of inflection points within a unit angle interval is counted to generate the distribution density of the rising slope inflection points. Perform Fast Fourier Transform processing on the standardized platform segment data to generate a frequency domain power spectral density distribution, and calculate the ratio of the integral energy value to the total energy value within a preset high-frequency bandwidth range based on the frequency domain power spectral density distribution to generate the energy proportion of the platform segment oscillation frequency band. An exponential regression analysis is performed on the standardized unloading section data to generate a stress relaxation decay curve. Based on the stress relaxation decay curve, the ratio of the decay rate difference to the time interval within adjacent time windows is calculated to generate the unloading section relaxation decay rate gradient change parameter. Normalization processing is performed on the distribution density of the inflection point of the rising slope, the energy ratio of the oscillation frequency band of the platform segment, and the gradient change parameters of the relaxation decay rate of the unloading segment to generate a standardized feature component set. The standardized feature component set is then vector-concatenated according to a predefined feature dimension order to generate a multidimensional original dynamic response feature vector. The dimensionless dynamic response fingerprint is generated by performing dimensionless mapping processing on the multidimensional original dynamic response feature vector.

5. The method for dynamic compensation control of bending angle of hardware parts according to claim 4, characterized in that, The normalization process is a maximum and minimum value normalization process.

6. The method for dynamic compensation control of bending angle of hardware parts according to claim 1, characterized in that, Step S4 specifically includes: Based on the dimensionless dynamic response fingerprint, a dynamic response feature vector is constructed. Using a preset feature node mapping rule, the value of each dimension in the dynamic response feature vector is transformed into the initial node attribute value in the graph neural network, generating an initial feature node set containing multiple feature nodes. Based on the physical coupling relationship between each feature node in the initial feature node set, a dynamic adjacency matrix construction operation is performed to define the edge connection weights that characterize the strong correlation between the distribution density of the rising slope inflection point and the energy proportion of the oscillation frequency band of the platform segment, thereby generating a dynamic adjacency matrix. The dynamic adjacency matrix is ​​used to perform multi-layer graph convolution aggregation processing on the initial feature node set. The local response information of adjacent feature nodes is fused through a message passing mechanism and the hidden layer state representation of the central feature node is updated to generate a deep node embedding vector containing global material dynamic property context information. Based on the deep node embedding vector, a fully connected layer classification mapping operation is performed. The softmax activation function, which has been trained offline, is called to calculate the probability distribution value of the deep node embedding vector belonging to each preset material property cluster, and a material property cluster label with the highest identification probability is generated. Based on the material property cluster label, a pre-stored cluster center parameter table is retrieved, a standard rebound behavior benchmark value uniquely bound to the material property cluster label is extracted, and a linear weighted correction is performed in combination with the confidence score of the deep node embedding vector to generate the rebound tendency coefficient used to guide the calculation of the compensation angle.

7. The method for dynamic compensation control of bending angle of hardware parts according to claim 6, characterized in that, The feature node mapping rule is as follows: the dimensionless dynamic response fingerprint is sequentially numbered from dimension 1 to dimension 12, and multiple feature nodes in the graph neural network are sequentially numbered from N1 to N12, establishing a fixed correspondence between dimension numbers and node numbers, that is, the i-th feature value is mapped to node N. i .

8. The method for dynamic compensation control of bending angle of hardware parts according to claim 1, characterized in that, Step S5 specifically includes: The cluster-compensation mapping knowledge base is initialized and loaded to construct a multi-dimensional index structure containing closed-loop bending verification data of historical materials in the same cluster under the same equipment conditions, and to generate a cluster-compensation mapping knowledge base instance. Based on the material property cluster tags, key-value matching retrieval processing is performed on the cluster-compensation mapping knowledge base instance to generate a knowledge base retrieval pointer pointing to a specific optimal compensation angle range; The optimal compensation angle interval boundary data stored in the historical data record entries are extracted using the knowledge base retrieval pointer to parse out the minimum and maximum compensation angle values ​​corresponding to the material performance cluster, thereby generating a set of optimal compensation angle intervals that characterize the rebound tolerance range. The arithmetic mean of the optimal compensation angle interval set is calculated to generate the median value of the optimal compensation angle interval.

9. The method for dynamic compensation control of bending angle of hardware parts according to claim 8, characterized in that, The knowledge base retrieval pointer contains a unique identifier for the target material property cluster and its corresponding storage address index.

10. The method for dynamic compensation control of bending angle of hardware parts according to claim 8, characterized in that, The cluster-compensation mapping knowledge base is initialized and loaded to construct a multi-dimensional index structure containing closed-loop bending verification data of historical materials in the same cluster under the same equipment operating conditions, generating a cluster-compensation mapping knowledge base instance, including: Receive a historical bending process database that has been trained and verified offline, wherein the historical bending process database contains different material batches, mold conditions and corresponding measured springback compensation data; After performing data cleaning and outlier removal operations on the different material batches, mold conditions and corresponding measured springback compensation data, the data are grouped in multiple dimensions according to material grade, sheet thickness, bending angle and mold V-groove width. For each group, the compensation angle records of materials in the same cluster during the hundred closed-loop bending verification are extracted, their distribution characteristics are calculated, and a cluster-level statistical summary is formed. A hash mapping process is used to generate a unique identifier key for each material property cluster, and the corresponding cluster-level statistical summary and the boundary value of the optimal compensation angle interval are encapsulated as key-value pair objects. Load all key-value pair objects into the read-only memory of the embedded controller to construct the cluster-compensation mapping knowledge base instance.

Images