A method for one-turn alignment of precision cylindrical grinding multi-source signals and cross-scale quality prediction

Abstract

The application discloses a precision cylindrical grinding multi-source signal one-turn alignment and cross-scale quality prediction method, which synchronously collects AE signals, ACC signals, workpiece rotating speed, grinding wheel rotating speed, grinding depth and quality measurement values; a one-turn physical window is constructed according to the workpiece rotating speed, so that the AE and ACC features under different rotating speeds correspond to the same rotating scale; Bayesian optimization VMD is adopted to carry out noise reduction and effective mode reconstruction; a target feature subset oriented to roughness, roundness and quality state is generated through multi-criteria feature scoring; a prediction mean value and a confidence interval are output by GPR, and early warning is output by a classification track under a low confidence state, so that the comparability of cross-rotating speed samples and the quality prediction reliability are improved.

Description

Technical Field

[0001] This invention relates to the fields of robot grasping, machine vision, digital twins, virtual-real migration, and flexible production line loading and unloading, and particularly to a method for one-turn alignment of multi-source signals and cross-scale quality prediction in precision cylindrical grinding. Background Technology

[0002] During the grinding of a dense cylindrical part, acoustic emission signals can reflect transient information about microcracks, friction, and material removal, while acceleration signals can reflect machine tool structural vibration and the interaction between the grinding wheel and the workpiece. However, the sampling frequency, bandwidth, and physical meaning of the two are significantly different.

[0003] Traditional fixed-time-window feature extraction methods can lead to inconsistencies in the physical scale of rotation for different samples when the workpiece rotation speed changes, affecting the comparability of roughness, roundness, and state recognition models. At the same time, single feature selection methods are difficult to balance relevance, redundancy, and physical interpretability.

[0004] Therefore, a cross-scale quality prediction method is needed that integrates physical window alignment, Bayesian optimized VMD denoising, multi-criteria feature selection, GPR uncertainty prediction, and regression classification dual-track state recognition. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the inconsistency in sampling frequency, physical meaning, and effective scale of AE and ACC signals in precision cylindrical grinding, the difficulty in comparing fixed time window features across rotational speeds, and the uncertainty and feature leakage risk in small sample quality prediction. This invention provides a method for multi-source signal alignment and cross-scale quality prediction with a physical window of one revolution as the anchor point.

[0006] To achieve the above objectives, the technical solution of this invention is as follows: a method for one-revolution alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, comprising: simultaneously acquiring acoustic emission (AE) signals, acceleration (ACC) signals, workpiece rotation speed, grinding wheel rotation speed, grinding depth, and quality measurement values; determining a one-revolution physical window based on the workpiece rotation speed, and extracting the AE and ACC signals respectively within the one-revolution physical window; performing Bayesian optimized VMD noise reduction and effective modal reconstruction on the extracted AE and ACC signals; extracting time-domain, frequency-domain, time-frequency-domain, and process features from the reconstructed signals; generating a target feature subset based on multi-criteria feature scoring; outputting the quality prediction mean and confidence interval using a GPR model; and outputting the quality status and updating the digital twin quality channel by combining a regression and classification dual-track model.

[0007] In the above-mentioned method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, the length of the one-turn physical window is calculated by converting the workpiece rotation speed and sampling frequency, so that the AE signal characteristics and ACC signal characteristics under different workpiece rotation speeds correspond to the same rotational physical scale.

[0008] In the above-mentioned method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, the Bayesian optimized VMD uses a weighted combination of signal-to-noise ratio, mutual information and smoothness as the objective function to determine the number of modes and the penalty factor, respectively. For AE signals, transient impact and high-frequency energy modes are preferentially retained, and for ACC signals, structural vibration-related modes are preferentially retained.

[0009] In the above-mentioned method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, the multi-criteria feature scoring includes random forest importance, Lasso coefficient, mutual information and Spearman correlation coefficient, and completes standardization, scoring, redundancy removal and target feature subset generation within the training iteration.

[0010] In the above-mentioned method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, a target feature subset is established. Target feature subsets are established for Ra, Rz, roundness, and quality status, respectively. Ra focuses on the overall energy and statistical distribution, Rz focuses on local peaks and valleys and frequency band energy ratio, roundness focuses on initial geometry and steady-state process parameters, and quality status focuses on the fusion of dynamic signals and process parameters.

[0011] In the aforementioned method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, when the confidence interval width of the GPR model output exceeds the threshold, the system marks the prediction result as a low-confidence prediction, and the classification track outputs a high-error state or a low-error state as the main warning result.

[0012] In the above-mentioned method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, the digital twin quality channel selects information sources based on the scale observable boundary. When the dynamic sensing features do not contribute enough to the target quality index, the steady-state process features or initial geometric features are preferentially used to update the corresponding quality model.

[0013] The technical solution adopted in this invention includes the following steps: window alignment, non-stationary signal reconstruction, feature selection, and low-confidence early warning: S1. Synchronously acquire AE, ACC, workpiece speed, grinding wheel speed, grinding depth, feed parameters and quality measurement values, and write the data to a timestamp.

[0014] S2. Calculate the length of the physical window for one revolution based on the workpiece rotation speed, and align and extract the AE and ACC signals within this window.

[0015] S3. Bayesian optimization VMD is used to process AE and ACC signals separately, and effective modes are selected and the noise-reduced signals are reconstructed.

[0016] S4. Extract time-domain, frequency-domain, time-frequency-domain, statistical, and steady-state process features from the reconstructed signal and process data.

[0017] S5. Through multi-criteria feature scoring and redundancy removal, target feature subsets for Ra, Rz, roundness and quality status are generated respectively.

[0018] S6. Train the GPR model to output the mean and confidence interval of the quality prediction, and train the classification track to identify high error or abnormal quality states.

[0019] S7. Based on the observable boundary of the scale, write back the prediction results, credibility and key information sources to the digital twin quality channel.

[0020] 1. By aligning the physical window in one revolution, the AE and ACC characteristics under different speed conditions are made comparable.

[0021] 2. By optimizing VMD using Bayesian methods, noise suppression and effective information preservation are balanced, thereby improving the feature quality of non-stationary grinding signals.

[0022] 3. Reduce redundancy and accidental correlations by using multi-criteria feature selection, thereby improving the model's generalization ability and physical interpretability.

[0023] 4. By using GPR confidence intervals and classification trajectory warnings, it can provide confidence markers in cases where samples are sparse or at the edge of distribution.

[0024] 5. Update the digital twin quality channel through scale-observable boundaries to avoid model drift caused by blindly stacking all features. Attached Figure Description

[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The schematic structures and processes in the drawings are used to explain the invention and do not constitute an undue limitation on the scope of protection of the invention.

[0026] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention; Figure 2 This is a schematic diagram of key modules or structural relationships in an embodiment of the present invention; The reference numerals in the figures include: Figure 1 S101 AE / ACC synchronous acquisition; S102 One-turn physical window alignment; S103 BO-VMD noise reduction and reconstruction; S104 Multi-criteria feature selection; S105 GPR and dual-track recognition; S106 Twin quality channel update; Figure 2 The input layer includes AE high-frequency signals, three-dimensional ACC signals, and speed / process parameters; the core processing layer includes one-revolution window alignment, BO-VMD modal reconstruction, target feature subset, GPR and dual-track identification; the output layer includes quality prediction values, confidence intervals, and quality status warnings; and the feedback layer includes scale-observable boundary update twin quality channels. Figure 1 and Figure 2 The arrows in the diagram indicate the direction of data, control commands, or execution feedback. Detailed Implementation

[0027] The invention will be further described in conjunction with the accompanying drawings.

[0028] Please see Figure 1 , Figure 2 This invention provides a method for one-revolution alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, comprising: simultaneously acquiring acoustic emission (AE) signals, acceleration (ACC) signals, workpiece rotation speed, grinding wheel rotation speed, grinding depth, and quality measurement values; determining a one-revolution physical window based on the workpiece rotation speed, and extracting the AE and ACC signals within the one-revolution physical window; performing Bayesian optimized VMD noise reduction and effective modal reconstruction on the extracted AE and ACC signals; extracting time-domain, frequency-domain, time-frequency-domain, and process features from the reconstructed signals; generating a target feature subset based on multi-criteria feature scoring; outputting the quality prediction mean and confidence interval using a GPR model; and outputting the quality status and updating the digital twin quality channel by combining a regression-classification dual-track model.

[0029] Furthermore, the length of the physical window is calculated from the workpiece rotation speed and sampling frequency, so that the AE signal characteristics and ACC signal characteristics at different workpiece rotation speeds correspond to the same rotational physical scale.

[0030] Furthermore, the Bayesian optimized VMD uses a weighted combination of signal-to-noise ratio, mutual information, and smoothness as the objective function to determine the number of modes and the penalty factor, respectively. For AE signals, transient impact and high-frequency energy modes are preferentially retained, and for ACC signals, structural vibration-related modes are preferentially retained.

[0031] Furthermore, the multi-criteria feature scoring includes random forest importance, Lasso coefficient, mutual information, and Spearman correlation coefficient, and standardization, scoring, redundancy removal, and target feature subset generation are completed within the training iteration.

[0032] Furthermore, target feature subsets are established for Ra, Rz, roundness, and quality status, respectively. Ra focuses on overall energy and statistical distribution, Rz focuses on local peaks and valleys and frequency band energy ratio, roundness focuses on initial geometry and steady-state process parameters, and quality status focuses on dynamic signal and process parameter fusion.

[0033] Furthermore, when the confidence interval width of the GPR model output exceeds the threshold, the system marks the prediction result as a low-confidence prediction, and the classification track outputs either a high-error state or a low-error state as the main warning result.

[0034] Furthermore, the digital twin quality channel selects information sources based on the scale observable boundary. When dynamic sensing features do not contribute enough to the target quality index, steady-state process features or initial geometric features are used to update the corresponding quality model.

[0035] To facilitate understanding and explanation, the present invention will be further described below with reference to the accompanying drawings and specific embodiments: In one embodiment, the AE sensor is installed near the grinding contact area, the triaxial accelerometer is installed on the spindle or workpiece support structure, and the acquisition system synchronously records the workpiece rotation speed, grinding wheel rotation speed, grinding depth, and cooling status.

[0036] The physical window is dynamically calculated based on the workpiece rotation speed. When the workpiece rotation speed changes, the AE and ACC signals may differ in the number of sampling points, but they both correspond to the same circumferential physical scale, thus avoiding the feature incomparability problem caused by a fixed time window.

[0037] During VMD noise reduction, the system searches for the number of modes and penalty factors for the AE and ACC signals, respectively. When reconstructing the AE signal, high-frequency modes related to abrasive impact, friction, and material removal are retained, while low- and mid-frequency modes related to machine tool vibration, grinding wheel imbalance, and workpiece support are retained when reconstructing the ACC signal.

[0038] During feature selection, the system completes standardization and scoring within each training iteration, without leaking test sample information into the training process. If two features are highly correlated and have similar physical meanings, the feature with better cross-condition stability is preferentially retained.

[0039] When making quality predictions, GPR provides the mean and confidence interval. If a sample is located at the edge of the training distribution, the confidence interval automatically widens, the system marks it as a low-confidence prediction, and prompts the user to supplement with adjacent working condition samples.

[0040] When the regression and classification track outputs conflict, the system selects the primary output based on the confidence level. If the regression track has a low confidence level while the classification track has a high confidence level for high error states, the classification warning will be prioritized, with the regression value used as a reference.

[0041] In a further embodiment, the AE signal's rotation window can be subdivided into several angular domain sub-windows according to the workpiece rotation speed, and each sub-window extracts energy and impact features to identify local grinding anomalies.

[0042] When a sudden change in signal distribution occurs after the grinding wheel is dressed, the system temporarily reduces the weight of the samples before dressing and recalibrates the kernel function hyperparameters of the GPR model using the samples after dressing.

[0043] For the roundness prediction channel, if dynamic sensing features conflict with initial geometric features, the system prioritizes the main information source after the observable boundary determination to avoid model drift caused by the stacking of all features.

[0044] In a further embodiment, a physical window is calculated based on the workpiece rotation speed and sampling frequency; when the workpiece rotation speed changes, the AE signal and ACC signal may have different numbers of sampling points, but they both correspond to the same circumferential contact scale, thereby avoiding the feature incomparability problem caused by a fixed time window.

[0045] When reconstructing AE signals, high-frequency transient modes related to abrasive impact, friction, and material removal are prioritized for retention. When reconstructing ACC signals, low- and mid-frequency modes related to machine tool structure vibration, grinding wheel imbalance, and workpiece support are prioritized for retention. The VMD mode count and penalty factor are searched for the two types of signals, respectively.

[0046] The target feature subset is independently standardized, random forest importance, Lasso sparsity coefficient, mutual information and Spearman correlation score within each training fold, and then redundancy removal and model training are performed to avoid leakage of test sample information to the feature selection process.

[0047] When the GPR confidence interval widens as the sample moves away from the training distribution, the system will mark the corresponding prediction as low confidence. If the regression track outputs low confidence while the classification track has a high confidence level for high error states, the classification warning will be the primary focus, and the regression prediction value will be used as a reference.

[0048] The observable boundary at scale is used to distinguish the information sources of quality indicators: AE / ACC dynamic features mainly describe roughness and transient machining state, while roundness indicators are preferentially combined with the initial geometry and steady-state process parameters before grinding to avoid mistaking local dynamic disturbances for macroscopic geometric errors.

[0049] The foregoing has provided a detailed description of a method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, as provided in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. It should be understood that the above embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make equivalent substitutions for specific equipment models, parameter ranges, data interfaces, and algorithm models without departing from the concept of the present invention. Any modifications, equivalent substitutions, improvements, or combinations made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding, characterized in that, include: The acoustic emission (AE) signal, acceleration (ACC) signal, workpiece rotation speed, grinding wheel rotation speed, grinding depth, and quality measurement values ​​are acquired simultaneously. A physical window of one revolution is determined based on the workpiece rotation speed, and the AE and ACC signals are extracted within the physical window. Bayesian optimized VMD noise reduction and effective mode reconstruction are performed on the extracted AE and ACC signals. Time domain, frequency domain, time-frequency domain, and process features are extracted from the reconstructed signals. A subset of target features is generated based on multi-criteria feature scoring; the mean and confidence interval of quality prediction are output using the GPR model; and the quality status is output and the digital twin quality channel is updated by combining a regression and classification dual-track model.

2. The method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding according to claim 1, characterized in that, The length of the physical window is calculated from the workpiece rotation speed and sampling frequency, so that the AE signal characteristics and ACC signal characteristics at different workpiece rotation speeds correspond to the same rotational physical scale.

3. The method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding according to claim 1, characterized in that, The Bayesian optimized VMD uses a weighted combination of signal-to-noise ratio, mutual information, and smoothness as the objective function to determine the number of modes and the penalty factor, respectively. For AE signals, transient impact and high-frequency energy modes are preferentially retained, and for ACC signals, structural vibration-related modes are preferentially retained.

4. The method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding according to claim 1, characterized in that, The multi-criteria feature scoring includes random forest importance, Lasso coefficient, mutual information, and Spearman correlation coefficient, and completes standardization, scoring, redundancy removal, and target feature subset generation within the training iteration.

5. The method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding according to claim 1, characterized in that, Establish target feature subsets for Ra, Rz, roundness and quality status respectively. Ra focuses on overall energy and statistical distribution, Rz focuses on local peaks and valleys and frequency band energy ratio, roundness focuses on initial geometry and steady-state process parameters, and quality status focuses on dynamic signal and process parameter fusion.

6. The method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding according to claim 1, characterized in that, When the confidence interval width of the GPR model output exceeds the threshold, the system marks the prediction result as a low-confidence prediction, and the classification track outputs a high-error state or a low-error state as the main warning result.

7. The method for one-turn alignment and cross-scale quality prediction of multi-source signals in precision cylindrical grinding according to claim 1, characterized in that, The digital twin quality channel selects information sources based on the scale observable boundary. When dynamic sensing features do not contribute enough to the target quality index, steady-state process features or initial geometric features are used to update the corresponding quality model.

Images