A grinding and polishing quality detection method, system, device and medium

Abstract

The application discloses a kind of grinding and polishing quality detection method, system, equipment and medium, wherein, in the method, migration model is obtained by the following steps: obtaining the source model corresponding to the target pose and the adversarial training set, the adversarial training set includes source pose majority class sample and migration pose minority class sample;The feature extraction is carried out to the adversarial training set, and the adversarial training feature is obtained;Texture prediction and field discrimination are carried out to the adversarial training feature, and the training texture prediction feature and the pose domain information of the training texture prediction feature are obtained;According to the pose domain information and the training texture prediction feature, the source model is subjected to migration adversarial training, and the migration model is obtained.The method can effectively improve the accuracy and applicability of robot grinding and polishing quality detection.The application relates to the field of robot automation manufacturing technology.

Description

Technical Field

[0001] This invention relates to the field of robotic automated manufacturing technology, and in particular to a grinding and polishing quality inspection method, system, equipment, and medium. Background Technology

[0002] Robotic grinding and polishing technology is increasingly widely used in the manufacturing of large and complex components (such as high-speed rail bodies, wind turbine blades, and aircraft skins), and the stability and consistency of its processing quality directly affects product performance. However, the inherent weak and variable stiffness characteristics of industrial robots make them susceptible to chatter and force fluctuations during the grinding and polishing process, leading to fluctuations in processing quality. Therefore, online monitoring and control of processing quality is crucial.

[0003] Currently, related technologies typically establish a "vibration-texture" feature mapping relationship to indirectly characterize polishing quality through vibration information, and then use this mapping relationship to detect the robot's polishing quality. However, this method is mainly for robots in fixed poses or under single working conditions, while robots often undergo pose changes during actual processing. This makes it difficult to achieve accurate polishing quality detection for robots throughout their entire workspace, resulting in unsatisfactory detection accuracy and applicability.

[0004] Therefore, the problems with the relevant technologies still need to be solved and optimized. Summary of the Invention

[0005] The purpose of this invention is to at least partially solve one of the technical problems existing in the related art.

[0006] Therefore, one objective of this invention is to provide a grinding and polishing quality inspection method, system, device, and medium, wherein the method can effectively improve the accuracy and applicability of robotic grinding and polishing quality inspection.

[0007] To achieve the above-mentioned technical objectives, the technical solutions adopted in the embodiments of this application include: In a first aspect, embodiments of this application provide a method for testing the quality of polishing, including: Acquire the first grinding and polishing vibration data of the robot in the target pose; The first grinding and polishing vibration data is input into the target texture prediction model to predict texture features, thereby obtaining the target texture prediction features; the target texture prediction model is the transfer model that corresponds to the target pose among all transfer models. The target texture prediction features are subjected to polishing quality detection to obtain the polishing quality detection result of the robot in the target pose. The migration model is obtained through the following steps: Obtain the source model and adversarial training set corresponding to the target pose. The adversarial training set includes majority class samples of the source pose and minority class samples of the transfer pose. The majority class samples of the source pose are training samples of the source model. The minority class samples of the transfer pose are training samples in the transferable domain corresponding to the source model. Feature extraction is performed on the adversarial training set to obtain adversarial training features; The adversarial training features are subjected to texture prediction and domain discrimination to obtain training texture prediction features and pose domain information of the training texture prediction features. Based on the pose domain information and the training texture prediction features, the source model is subjected to transfer adversarial training to obtain the transfer model.

[0008] In addition, the method according to the above embodiments of this application may also have the following additional technical features: Furthermore, in one embodiment of this application, obtaining the source model corresponding to the target pose includes: Obtain a set of transferable domain spaces, which is used to represent a set of spatial domains of transferable domains for several different source poses, and each source pose corresponds to a basic mapping model. Based on the first grinding and polishing vibration data, the transferable domain space set is matched to obtain the domain pose. The domain pose is used to characterize the source pose corresponding to the matched domain. The matched domain is a transferable domain that contains the target pose space point among all the transferable domains. Based on the domain pose, all the basic mapping models are screened to obtain the source model corresponding to the target pose.

[0009] Further, in one embodiment of this application, the step of performing transferable domain matching on the transferable domain space set based on the first grinding and polishing vibration data to obtain the domain pose includes: Based on the transferable domain space set, obtain the second grinding and polishing vibration data for each of the source poses; Based on the first grinding and polishing vibration data, pose analysis is performed on all the second grinding and polishing vibration data to obtain several pose difference degrees. Based on all the pose differences, pose filtering is performed on all the source poses to obtain the domain pose.

[0010] Furthermore, in one embodiment of this application, the step of performing pose analysis on all the second grinding and polishing vibration data based on the first grinding and polishing vibration data to obtain several pose difference degrees includes: Dynamic similarity analysis was performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the dynamic similarity. Data similarity analysis was performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the data similarity. The dynamic similarity and the data similarity are measured and analyzed to obtain the pose difference.

[0011] Furthermore, in one embodiment of this application, the step of performing dynamic similarity analysis on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain dynamic similarity includes: Frequency response analysis is performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the first set of zeros and poles corresponding to the first grinding and polishing vibration data and the second set of zeros and poles corresponding to the second grinding and polishing vibration data. The dynamic similarity is obtained by performing maximum mismatch analysis on the first set of zeros and poles and the second set of zeros and poles.

[0012] Further, in one embodiment of this application, the step of performing pose filtering on all the source poses based on all the pose difference degrees to obtain the domain pose includes: Obtain the difference threshold of the source pose; Based on each of the difference thresholds, all corresponding pose differences are filtered to obtain the target difference, which is the smallest pose difference among all pose differences less than the corresponding difference threshold. Based on the target difference, all the source poses are mapped and filtered to obtain the domain pose.

[0013] Furthermore, in this embodiment of the application, the method further includes: Obtain the original network, as well as the second grinding and polishing vibration data and grinding and polishing texture image of the robot in the source pose; The polished texture image is subjected to texture quantization analysis to obtain texture quantization information; Feature extraction is performed on the second grinding and polishing vibration data to obtain the grinding and polishing vibration features; Based on the grinding and polishing vibration characteristics and the texture quantization information, the original network is trained to obtain the basic mapping model of the source pose.

[0014] Secondly, embodiments of this application provide a polishing quality inspection system, comprising: The first processing unit is used to acquire the first grinding and polishing vibration data of the robot in the target pose. The second processing unit is used to input the first grinding and polishing vibration data into the target texture prediction model to predict texture features and obtain target texture prediction features; the target texture prediction model is the transfer model corresponding to the target pose among all transfer models. The third processing unit is used to perform polishing quality detection on the target texture prediction features to obtain the polishing quality detection result of the robot in the target pose. The migration model is obtained through the following steps: Obtain the source model and adversarial training set corresponding to the target pose. The adversarial training set includes majority class samples of the source pose and minority class samples of the transfer pose. The majority class samples of the source pose are training samples of the source model. The minority class samples of the transfer pose are training samples in the transferable domain corresponding to the source model. Feature extraction is performed on the adversarial training set to obtain adversarial training features; The adversarial training features are subjected to texture prediction and domain discrimination to obtain training texture prediction features and pose domain information of the training texture prediction features. Based on the pose domain information and the training texture prediction features, the source model is subjected to transfer adversarial training to obtain the transfer model.

[0015] Thirdly, embodiments of this application also provide an electronic device, including: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0016] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a processor-executable program, which, when executed by the processor, is used to implement the above-described method.

[0017] Fifthly, embodiments of this application also provide a computer program product, which includes a computer program stored in a computer-readable storage medium. A processor of an electronic device reads the computer program from the computer-readable storage medium and executes the computer program, causing the electronic device to perform the method described above.

[0018] The advantages and beneficial effects of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application: This application discloses a grinding and polishing quality detection method, system, device, and medium. The method involves acquiring first grinding and polishing vibration data of a robot in a target pose; inputting the first grinding and polishing vibration data into a target texture prediction model for texture feature prediction to obtain target texture prediction features; the target texture prediction model is a transfer model corresponding to the target pose among all transfer models; and performing grinding and polishing quality detection on the target texture prediction features to obtain the grinding and polishing quality detection result of the robot in the target pose. The transfer model is obtained through the following steps: acquiring a source model corresponding to the target pose. The method involves a target texture prediction model corresponding to the robot's current pose (i.e., the target pose) for texture feature prediction during the robot's grinding and polishing process. This target texture prediction model is obtained based on the source model's transfer adversarial training. Combined with subsequent grinding and polishing quality detection, it can adapt to the robot's pose changes during actual processing, achieving accurate grinding and polishing quality detection across the entire workspace, effectively improving the accuracy and applicability of grinding and polishing quality detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of this application or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions in this application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 A schematic flowchart of a polishing quality testing method provided in an embodiment of this application; Figure 2 A schematic diagram of the framework of a polishing quality inspection system provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of these embodiments 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 this application, and should not be construed as limiting this application. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0023] The technical terms used in the embodiments of this application are explained below: Positioning posture, also known as fixed working condition, refers to a working condition in which the position and posture of the end effector of a robot remain fixed during the processing process.

[0024] Variable pose, also known as variable working condition, refers to a working condition in which the position and orientation of the end effector of a robot change continuously during the processing (such as moving along a complex curved surface). Under this condition, the stiffness and dynamic characteristics of the robot system will change significantly.

[0025] Currently, related technologies typically establish a "vibration-texture" feature mapping relationship to indirectly characterize polishing quality through vibration information. This approach is mainly for fixed poses or single working conditions, and does not consider the changes in system dynamics caused by continuous pose changes during actual robot processing.

[0026] Furthermore, some related technologies use the weighted cumulative spectral distribution entropy and dual-tree complex wavelet packet energy entropy of the vibration signal as features to train a support vector machine model to identify states such as stable, regenerative flutter, and modally coupled flutter, and estimate the flutter frequency. This approach focuses on the identification of specific types of processing states (flutter), but it does not construct a direct, quantitative mapping model between the vibration signal and the final surface quality (such as roughness and texture uniformity), and its model also faces the challenge of feature distribution drift when pose changes.

[0027] Based on the above, the relevant technologies have the following shortcomings: Poor adaptability to changes in working conditions: Most vibration-based quality detection or flutter recognition models in related technologies are trained based on single or fixed pose working condition data. However, during actual grinding and polishing operations, the end stiffness and dynamic response of the robot change continuously with the pose, which leads to changes in the characteristic distribution of vibration signals. This causes the performance of the model trained under fixed working conditions to deteriorate significantly under changing working conditions.

[0028] Insufficient quality characterization and model generalization capabilities: Related technologies focus on the identification of specific abnormal states (such as flutter), or the established "vibration-texture" mapping models are only validated under limited working conditions, lacking robust quality characterization capabilities for continuous and variable working conditions. When the robot's pose changes, the mapping relationship between vibration and surface quality may drift nonlinearly, making it difficult to achieve accurate monitoring across the entire workspace by simply relying on a fixed model.

[0029] There is a lack of effective cross-condition model transfer strategies: Under varying conditions, it is costly and impractical to retrain the model by collecting a large amount of labeled data (vibration and corresponding surface quality).

[0030] It should be noted that the aforementioned related technologies are only used to assist in understanding the technical solutions of this application and do not mean that they belong to the publicly disclosed prior art.

[0031] In view of this, embodiments of this application provide a grinding and polishing quality inspection method. During the grinding and polishing process of a robot, the method provides a target texture prediction model corresponding to the robot's current pose (i.e., target pose) to predict texture features. This target texture prediction model is obtained based on source model transfer adversarial training. Combined with subsequent grinding and polishing quality inspection, it can adapt to the pose changes of the robot in the actual processing process, realize accurate grinding and polishing quality inspection of the robot in the entire workspace, and effectively improve the accuracy and applicability of grinding and polishing quality inspection.

[0032] Specifically, for different robot poses and working conditions, massive amounts of vibration data are re-collected. Training new models by pairing high-quality data is costly and inefficient. This application introduces transfer adversarial training to efficiently transfer the knowledge of a pre-trained vibration texture model (i.e., the basic mapping model) under existing pose conditions (source pose) to new pose conditions (target pose). This method maintains high detection accuracy even when target pose data is limited or even unlabeled, greatly reducing the workload of data collection and labeling, and improving the engineering applicability and promotion efficiency of the detection method.

[0033] Furthermore, this method obtains the domain pose through dynamic similarity analysis and data similarity analysis, which can quantify the impact of robot pose transformation on vibration characteristics from a mechanistic perspective. Combined with subsequent transfer adversarial training of the model, it can learn invariant features in the domain pose that are insensitive to pose transformation, thereby enabling each texture prediction model to adapt to pose transformations within the corresponding transferable domain of the robot, thus realizing pose transformations throughout the robot's entire workspace and effectively improving the stability and accuracy of robot quality detection under varying working conditions.

[0034] Reference Figure 1In this embodiment of the application, a method for testing the quality of polishing includes: Step 110: Obtain the first grinding and polishing vibration data of the robot in the target pose; Step 120: Input the first grinding and polishing vibration data into the target texture prediction model to predict texture features and obtain the target texture prediction features; the target texture prediction model is the transfer model corresponding to the target pose among all transfer models; Step 130: Perform polishing quality detection on the predicted features of the target texture to obtain the polishing quality detection result of the robot in the target pose; In this embodiment of the application, the grinding and polishing quality detection method mentioned in this embodiment can be repeatedly executed during the grinding and polishing process of the robot. Specifically, for a certain cycle in the robot grinding and polishing process, the first grinding and polishing vibration data under the current pose (i.e., target pose) of the robot's workspace can be obtained. The first grinding and polishing vibration data can be collected by a vibration sensor (such as an accelerometer) installed on the robot's end effector or grinding and polishing tool.

[0035] Texture feature prediction can be achieved by inputting the first grinding and polishing vibration data into a transfer model (i.e., a target texture prediction model) corresponding to the robot's target pose. The target texture prediction model then maps and predicts the texture feature vector corresponding to the first grinding and polishing vibration data, which is denoted as the target texture prediction feature. Grinding and polishing quality detection can be achieved by classifying the target texture prediction feature into "under-grinding," "normal grinding and polishing," or "over-grinding" states, thereby obtaining the grinding and polishing quality detection result. There are various specific classification methods available, such as using a classifier or by using pre-defined threshold ranges for under-grinding, normal grinding and polishing, and over-grinding features.

[0036] The migration model is obtained through the following steps: Obtain the source model and adversarial training set corresponding to the target pose. The adversarial training set includes majority class samples of the source pose and minority class samples of the transfer pose. The majority class samples of the source pose are training samples of the source model. The minority class samples of the transfer pose are training samples in the transferable domain corresponding to the source model. In the first implementation, the source model corresponding to the target pose can be the basic mapping model of the source pose that includes the transferable domain of the target pose among all source poses. This source pose can be the working pose of the robot under multiple sets of grinding and polishing experiments in history. The majority class samples of the source poses in the adversarial training set can be the second grinding and polishing vibration data collected by the robot under multiple sets of grinding and polishing experiments based on the source pose in history, which is also the training samples of the source model.

[0037] It is understandable that if the robot performs grinding and polishing work for the first time in the target pose during the grinding and polishing process, the minority class sample of the transition pose can be the first grinding and polishing vibration data of the robot in the current target pose, and the number of the minority class samples of the transition pose is less than the majority class samples of the source pose; or, if the robot performs grinding and polishing work in the target pose for the second or more times during the grinding and polishing process, the minority class sample of the transition pose can be the first grinding and polishing vibration data of the previous target pose, or it can be a combination of the first grinding and polishing vibration data of the previous and current working poses.

[0038] In the second implementation, the robot may enter a blank region during the variable-condition grinding and polishing process. This blank region characterizes the spatial region corresponding to the target pose when none of the transferable domains of all source poses contain the target pose. In this case, the source model corresponding to the target pose can be the basic mapping model of the source pose of the nearest region adjacent to the blank region. The rest is similar to the content of the first implementation and can be easily deduced. This ensures the continuity and accuracy of grinding and polishing quality inspection when the robot performs continuous grinding and polishing tasks on complex surfaces. Simultaneously, it improves the engineering applicability of robot grinding and polishing quality inspection when faced with limited or even unlabeled target pose data.

[0039] In some embodiments, obtaining the source model corresponding to the target pose includes: Obtain a set of transferable domain spaces, which is used to represent a set of spatial domains of transferable domains for several different source poses, and each source pose corresponds to a basic mapping model. In this embodiment of the application, the transferable domain corresponding to each source pose in the robot's workspace can be obtained. The transferable domain is used to characterize the spatial region around the source pose where the model can stably transfer. Then, all transferable domains are determined as a transferable domain space set.

[0040] Based on the first grinding and polishing vibration data, the transferable domain space set is matched to obtain the domain pose. The domain pose is used to characterize the source pose corresponding to the matched domain. The matched domain is a transferable domain that contains the target pose space point among all the transferable domains. Further, the step of performing transferable domain matching on the transferable domain space set based on the first grinding and polishing vibration data to obtain the domain pose includes: Based on the transferable domain space set, obtain the second grinding and polishing vibration data for each of the source poses; In the embodiments of this application, for any one of the transferable domains in the set of transferable domains, the second grinding and polishing vibration data of the corresponding source pose can be obtained based on the correspondence between the transferable domain and the source pose.

[0041] Based on the first grinding and polishing vibration data, pose analysis is performed on all the second grinding and polishing vibration data to obtain several pose difference degrees. Furthermore, based on the first grinding and polishing vibration data, pose analysis is performed on all the second grinding and polishing vibration data to obtain several pose difference degrees, including: Dynamic similarity analysis was performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the dynamic similarity. Further, the dynamic similarity analysis of the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the dynamic similarity includes: Frequency response analysis is performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the first set of zeros and poles corresponding to the first grinding and polishing vibration data and the second set of zeros and poles corresponding to the second grinding and polishing vibration data. The dynamic similarity is obtained by performing maximum mismatch analysis on the first set of zeros and poles and the second set of zeros and poles.

[0042] In this embodiment, for a target pose and any one of the source poses, frequency response analysis can involve acquiring the response spectrum and excitation force spectrum corresponding to the first grinding and polishing vibration data. The response spectrum can be acquired using an accelerometer, while the excitation force spectrum can be acquired using a force hammer. The ratio of the response spectrum to the excitation force spectrum is then calculated to determine the frequency response function corresponding to the first grinding and polishing vibration data. Next, the dominant zero-point set and pole set of the frequency response function of the first grinding and polishing vibration data are extracted within the main frequency band, and this set is denoted as the first zero-pole set. The content of the second zero-pole set is similar to that of the first zero-pole set and can be easily deduced. Maximum mismatch analysis can involve calculating the Hausdorff distance between the first and second zero-pole sets, and determining the calculated distance as the dynamic similarity between the target pose and a source pose.

[0043] Data similarity analysis was performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the data similarity. The dynamic similarity and the data similarity are measured and analyzed to obtain the pose difference.

[0044] In this embodiment of the application, data similarity analysis may involve extracting the feature vectors of the first grinding and polishing vibration data and the feature vector sets of the second grinding and polishing vibration data, respectively, and calculating the 1-Wasserstein distance between the two feature vector distributions, denoted as the data similarity. This data similarity can be expressed as:

[0045] in, It can be the data similarity between the i-th grinding and polishing vibration data of the source pose and the j-th grinding and polishing vibration data of the target pose, or it can be the data similarity between the grinding and polishing vibration data of the source pose i and the grinding and polishing vibration data of the target pose j. This is the feature vector set of the second grinding and polishing vibration data; This is the feature vector set of the first grinding and polishing vibration data; It is a function representation of the 1-Wasserstein distance.

[0046] The metric analysis can combine dynamic similarity and data similarity to obtain a pose difference, which quantifies the difference in feature distribution caused by changes in robot pose. This pose difference can be expressed as:

[0047] in, The degree of pose difference; For dynamic similarity; and For the weighting coefficients, satisfying .

[0048] Based on all the pose differences, pose filtering is performed on all the source poses to obtain the domain pose.

[0049] Further, the step of performing pose filtering on all the source poses based on all the pose differences to obtain the domain pose includes: Obtain the difference threshold of the source pose; Based on each of the difference thresholds, all corresponding pose differences are filtered to obtain the target difference, which is the smallest pose difference among all pose differences less than the corresponding difference threshold. Based on the target difference, all the source poses are mapped and filtered to obtain the domain pose.

[0050] In this embodiment, a difference threshold for each source pose can be obtained. The difference threshold for each source pose can be the same or different. For ease of understanding, this embodiment will subsequently use the example of all source poses having the same difference threshold. This difference threshold can be a preset threshold, and its range can be [0.1, 0.5]. Threshold filtering can involve comparing the magnitude of each pose difference with the difference threshold to filter out all pose differences less than the difference threshold, denoted as intermediate differences. Then, the magnitude of all intermediate differences is compared, and the smallest intermediate difference is determined as the target difference. Combined with subsequent mapping filtering across all source poses, the source pose corresponding to the target difference is determined as the domain pose, which can filter out the optimal source pose from all source poses whose transferable domain contains the target pose.

[0051] It is understandable that, for the pose difference between the target pose and any source pose, if the pose difference is less than the corresponding difference threshold, it means that the target pose is within the transferable domain of the corresponding source pose; conversely, if the pose difference is greater than or equal to the corresponding difference threshold, it means that the target pose is not within the transferable domain of the corresponding source pose.

[0052] It should be noted that for any transferable domain of a source pose, multiple working poses can be pre-selected as source poses within the robot's workspace. By performing dynamic similarity analysis on the grinding and polishing vibration data of the source pose and several non-source poses, the pose difference between each non-source pose and the source pose can be determined. Then, the non-source pose corresponding to each pose difference less than the difference threshold is determined as a pose space point, and the smallest region in the workspace covering all pose space points is constructed as the transferable domain of the source pose.

[0053] Based on the domain pose, all the basic mapping models are screened to obtain the source model corresponding to the target pose.

[0054] In the embodiments of this application, the basic mapping model corresponding to the domain pose can be determined based on the correspondence between each source pose and the basic mapping model, and denoted as the source model.

[0055] In some embodiments, the method further includes: Obtain the original network, as well as the second grinding and polishing vibration data and grinding and polishing texture image of the robot in the source pose; The polished texture image is subjected to texture quantization analysis to obtain texture quantization information; Feature extraction is performed on the second grinding and polishing vibration data to obtain the grinding and polishing vibration features; Based on the grinding and polishing vibration characteristics and the texture quantization information, the original network is trained to obtain the basic mapping model of the source pose.

[0056] In this embodiment, the original network can be an initialized deep residual network. The polishing texture image can be a constructed polishing surface texture image corresponding to the second polishing vibration data acquired by the robot through an image acquisition device. Texture quantization analysis can be performed by using a sparse autoencoder to extract deep features of the texture, and by using a gray-level co-occurrence matrix (GLCM) to calculate texture quantization indices (such as texture contrast, correlation, energy, homogeneity, etc.) from the deep features, which are denoted as texture quantization information.

[0057] Understandably, feature extraction can be achieved by using Symmetrized Dot Pattern (SDP) or time-frequency transform methods to convert a one-dimensional vibration signal into a two-dimensional feature image, and then extracting the feature vector corresponding to the texture index, denoted as the grinding and polishing vibration feature. Network training can be performed using the grinding and polishing vibration feature as the input to the original network and the texture quantization index as the model's output label. Supervised learning is then used to train the original network to achieve end-to-end mapping prediction from vibration signal to surface texture quality, thereby obtaining the basic mapping model for the source pose.

[0058] Feature extraction is performed on the adversarial training set to obtain adversarial training features; The adversarial training features are subjected to texture prediction and domain discrimination to obtain training texture prediction features and pose domain information of the training texture prediction features. In this embodiment, an adversarial adaptive network can be constructed based on Generative Adversarial Network (GAN) technology. This adversarial adaptive network includes a feature extractor, a texture predictor, and a domain discriminator. The feature extraction layer consists of multiple fully connected layers, the texture predictor can be a generator in GAN technology, and the domain discriminator can be a discriminator in GAN technology. Feature extraction of the adversarial training set can be achieved by inputting the majority class samples of the source pose and the minority class samples of the transferred pose into the feature extractor in the adversarial adaptive network. The feature extractor extracts the corresponding sample feature vectors, denoted as adversarial training features. Texture prediction and domain discrimination can be achieved by inputting the adversarial training features into the texture predictor and the domain discriminator in the adversarial adaptive network. The texture predictor predicts texture feature vectors based on the adversarial training features (i.e., trains texture prediction features); and the domain discriminator determines whether the adversarial training features originate from the source pose or the target pose, thereby obtaining pose domain information.

[0059] Based on the pose domain information and the training texture prediction features, the source model is subjected to transfer adversarial training to obtain the transfer model.

[0060] In this embodiment, a gradient inversion layer can be set between the feature extractor and the neighborhood discriminator. This gradient inversion layer is passed identically during forward propagation and multiplied by the incoming gradient during backward propagation. ,in For adaptive trade-off coefficients, This allows the feature extractor to engage in adversarial training with the domain discriminator, thereby encouraging the learning of invariant features that are insensitive to pose changes. Transfer adversarial training can be achieved by combining the total loss function with a total loss value determined based on pose domain information and training texture prediction features. Then, the parameters of the source model at the source pose are updated using a backpropagation algorithm to obtain the transfer model. Various implementations exist for updating model parameters using backpropagation, which will not be elaborated upon here.

[0061] For example, the total loss function can be expressed as:

[0062] in, For the total loss function, The set of trainable parameters for the feature extractor; This is the set of trainable parameters for the texture predictor. This is the set of trainable parameters for the domain discriminator. For mission losses; To determine the loss in the domain; Another representation of domain-specific loss; This is another way of representing task loss; A functional representation of the feature extractor; A functional representation of a texture predictor; This is the second grinding and polishing vibration data collected by the robot in the source pose (known working conditions); It serves as an index symbol for the second grinding and polishing vibration data or source pose. For the second grinding and polishing vibration data under a certain source pose, it can indicate the i-th second grinding and polishing vibration data. Alternatively, for the second grinding and polishing vibration data of all source poses, it can indicate the second grinding and polishing vibration data of source pose i. The first grinding and polishing vibration data collected by the robot under the target pose (new working condition); This serves as the index symbol for the first grinding and polishing vibration data or the target pose. This is the feature vector of the i-th second grinding and polishing vibration data; The label (real texture feature vector) can specifically be the aforementioned texture quantization information; This represents the total number of samples for the second grinding and polishing vibration data. This represents the total number of samples for the first grinding and polishing vibration data. This is the predicted texture feature vector output by the texture predictor; Let be the probability that the second grinding and polishing vibration data under source pose i is identified as the source pose. ; A function representation of the domain discriminator; Let be the probability that the first grinding and polishing vibration data under the target pose j is identified as the source pose. ; Let be the feature vector of the j-th first grinding and polishing vibration data.

[0063] Reference Figure 2 The polishing quality detection system proposed in this application includes: The first processing unit 101 is used to acquire the first grinding and polishing vibration data of the robot in the target pose. The second processing unit 102 is used to input the first grinding and polishing vibration data into the target texture prediction model to predict texture features and obtain target texture prediction features; the target texture prediction model is the transfer model corresponding to the target pose among all transfer models. The third processing unit 103 is used to perform polishing quality detection on the target texture prediction features to obtain the polishing quality detection result of the robot in the target pose. The migration model is obtained through the following steps: Obtain the source model and adversarial training set corresponding to the target pose. The adversarial training set includes majority class samples of the source pose and minority class samples of the transfer pose. The majority class samples of the source pose are training samples of the source model. The minority class samples of the transfer pose are training samples in the transferable domain corresponding to the source model. Feature extraction is performed on the adversarial training set to obtain adversarial training features; The adversarial training features are subjected to texture prediction and domain discrimination to obtain training texture prediction features and pose domain information of the training texture prediction features. Based on the pose domain information and the training texture prediction features, the source model is subjected to transfer adversarial training to obtain the transfer model.

[0064] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0065] Reference Figure 3 This application also provides an electronic device, including: At least one processor 201; At least one memory 202 is used to store at least one program; When the at least one program is executed by the at least one processor 201, the at least one processor 201 implements the method embodiment described above.

[0066] Similarly, it can be understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0067] This application also provides a computer-readable storage medium storing a program executable by a processor 201, which, when executed by the processor 201, is used to implement the above-described method embodiments.

[0068] Similarly, the content of the above method embodiments is applicable to the present computer-readable storage medium embodiments. The specific functions implemented by the present computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0069] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0070] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0071] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0072] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0073] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0074] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0075] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0076] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0077] In the foregoing description of this specification, the references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0078] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0079] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for testing the quality of grinding and polishing, characterized in that, include: Acquire the first grinding and polishing vibration data of the robot in the target pose; The first grinding and polishing vibration data is input into the target texture prediction model to predict texture features, thereby obtaining the target texture prediction features; the target texture prediction model is the transfer model that corresponds to the target pose among all transfer models. The target texture prediction features are subjected to polishing quality detection to obtain the polishing quality detection result of the robot in the target pose. The migration model is obtained through the following steps: Obtain the source model and adversarial training set corresponding to the target pose. The adversarial training set includes majority class samples of the source pose and minority class samples of the transfer pose. The majority class samples of the source pose are training samples of the source model. The minority class samples of the transfer pose are training samples in the transferable domain corresponding to the source model. Feature extraction is performed on the adversarial training set to obtain adversarial training features; The adversarial training features are subjected to texture prediction and domain discrimination to obtain training texture prediction features and pose domain information of the training texture prediction features. Based on the pose domain information and the training texture prediction features, the source model is subjected to transfer adversarial training to obtain the transfer model.

2. The method according to claim 1, characterized in that, The step of obtaining the source model corresponding to the target pose includes: Obtain a set of transferable domain spaces, which is used to represent a set of spatial domains of transferable domains for several different source poses, and each source pose corresponds to a basic mapping model. Based on the first grinding and polishing vibration data, the transferable domain space set is matched to obtain the domain pose. The domain pose is used to characterize the source pose corresponding to the matched domain. The matched domain is a transferable domain that contains the target pose space point among all the transferable domains. Based on the domain pose, all the basic mapping models are screened to obtain the source model corresponding to the target pose.

3. The method according to claim 2, characterized in that, The step of performing transferable domain matching on the transferable domain space set based on the first grinding and polishing vibration data to obtain the domain pose includes: Based on the transferable domain space set, obtain the second grinding and polishing vibration data for each of the source poses; Based on the first grinding and polishing vibration data, pose analysis is performed on all the second grinding and polishing vibration data to obtain several pose difference degrees. Based on all the pose differences, pose filtering is performed on all the source poses to obtain the domain pose.

4. The method according to claim 3, characterized in that, The step involves performing pose analysis on all the second grinding and polishing vibration data based on the first grinding and polishing vibration data to obtain several pose difference degrees, including: Dynamic similarity analysis was performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the dynamic similarity. Data similarity analysis was performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the data similarity. The dynamic similarity and the data similarity are measured and analyzed to obtain the pose difference.

5. The method according to claim 4, characterized in that, The dynamic similarity analysis of the first and second grinding and polishing vibration data to obtain dynamic similarity includes: Frequency response analysis is performed on the first grinding and polishing vibration data and the second grinding and polishing vibration data to obtain the first set of zeros and poles corresponding to the first grinding and polishing vibration data and the second set of zeros and poles corresponding to the second grinding and polishing vibration data. The dynamic similarity is obtained by performing maximum mismatch analysis on the first set of zeros and poles and the second set of zeros and poles.

6. The method according to claim 3, characterized in that, The step of filtering all source poses based on all pose differences to obtain the domain pose includes: Obtain the difference threshold of the source pose; Based on each of the difference thresholds, all corresponding pose differences are filtered to obtain the target difference, which is the smallest pose difference among all pose differences less than the corresponding difference threshold. Based on the target difference, all the source poses are mapped and filtered to obtain the domain pose.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Obtain the original network, as well as the second grinding and polishing vibration data and grinding and polishing texture image of the robot in the source pose; The polished texture image is subjected to texture quantization analysis to obtain texture quantization information; Feature extraction is performed on the second grinding and polishing vibration data to obtain the grinding and polishing vibration features; Based on the grinding and polishing vibration characteristics and the texture quantization information, the original network is trained to obtain the basic mapping model of the source pose.

8. A polishing quality inspection system, characterized in that, include: The first processing unit is used to acquire the first grinding and polishing vibration data of the robot in the target pose. The second processing unit is used to input the first grinding and polishing vibration data into the target texture prediction model to predict texture features and obtain target texture prediction features; the target texture prediction model is the transfer model corresponding to the target pose among all transfer models. The third processing unit is used to perform polishing quality detection on the target texture prediction features to obtain the polishing quality detection result of the robot in the target pose. The migration model is obtained through the following steps: Obtain the source model and adversarial training set corresponding to the target pose. The adversarial training set includes majority class samples of the source pose and minority class samples of the transfer pose. The majority class samples of the source pose are training samples of the source model. The minority class samples of the transfer pose are training samples in the transferable domain corresponding to the source model. Feature extraction is performed on the adversarial training set to obtain adversarial training features; The adversarial training features are subjected to texture prediction and domain discrimination to obtain training texture prediction features and pose domain information of the training texture prediction features. Based on the pose domain information and the training texture prediction features, the source model is subjected to transfer adversarial training to obtain the transfer model.

9. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to implement the method as described in any one of claims 1-7.

Images