Online monitoring method of grinding status based on industrial robot body information

Abstract

An online monitoring method for the grinding state based on the body information of an industrial robot, which collects the robot joint current during the pre-grinding experiment and the surface quality texture after the grinding is completed; then the joint current time series signal is converted into an image, thereby generating a grinding state label and a grinding state classification model data set, and the data set is used to train the grinding state classification model; finally, the trained grinding state classification model is used in the online stage to realize the closed-loop feedback control of the robot. The present invention uses a convolutional neural network to capture the inherent time-frequency domain characteristics of the robot joint current, establishes a feature joint selection model based on the principal component analysis of the joint current, and realizes the online monitoring of the robot grinding state.

Description

Technical Field

[0001] The present invention relates to a technology in the field of automated manufacturing, and specifically to an online monitoring method for a grinding state based on the body information of an industrial robot. Background Art

[0002] Limited by the influence of existing industrial robots and workpieces, industrial robot grinding is prone to abnormal processing conditions such as under-grinding and over-grinding, which will lead to a decrease in the surface quality of the workpiece. In order to ensure the processing quality of the workpiece being ground, it is necessary to put forward more accurate and efficient requirements for the grinding quality monitoring method, and finally realize the online monitoring of the workpiece surface quality and complete the closed-loop control of robot automated grinding in real time based on the monitoring results. The existing technology analyzes and extracts the collected processing signals, combines the evaluation indicators such as the surface texture and roughness of the workpiece, and establishes a mapping relationship between the signal characteristics and the surface quality evaluation indicators. In the online monitoring of the grinding state of industrial robots, in the article "Onlineprocess monitoring based on vibration-surface quality map for roboticgrinding", Lu et al. arranged an acceleration sensor on the support where the workpiece being ground was installed to collect the vibration signal of the robot grinding process, converted the time series signal into an image through the symmetrical point pattern diagram for feature extraction, and judged the robot grinding state as under-grinding, normal grinding, and over-grinding according to the workpiece surface image texture parameters and generated the label of the symmetrical point pattern diagram. Finally, a convolutional neural network was used to complete the image classification and realize the monitoring of the grinding state. Summary of the invention

[0003] In view of the defects that the acceleration sensors attached in the prior art are expensive and limited by the arrangement position of the sensors, the collected acceleration signals can only reflect the vibration of the workpiece to a certain extent but cannot reflect the vibration of the robot, and the monitoring accuracy depends on the size of the training data set and it is difficult to ensure the constancy of other parameters while frequently adjusting the process parameters. An online monitoring method for the grinding state based on the body information of an industrial robot is proposed, a convolutional neural network is used to capture the intrinsic time-frequency domain characteristics of the robot joint current, a feature joint selection model based on the principal component analysis of the joint current is established, and the online monitoring of the robot grinding state is realized.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to an online monitoring method for the grinding state based on the body information of an industrial robot. The method collects the robot joint current in a pre-grinding experiment and collects the surface quality texture after the grinding is completed. The joint current timing signal is then converted into an image to generate a grinding state label and a grinding state classification model data set. The data set is used to train the grinding state classification model. Finally, the trained grinding state classification model is used in the online stage to realize the closed-loop feedback control of the robot.

[0006] The pre-grinding experiment refers to: constructing the under-grinding, normal grinding and over-grinding states of the workpiece surface by setting the feed rate of the robot grinding.

[0007] The collecting of robot joint currents refers to: collecting current signals of six joints of the robot under different processing states, and normalizing the collected current signals. Preferably, the current signals of each joint of the industrial robot during grinding are collected at a frequency of 10Khz, and the percentage of the measured current to the rated current of the joint is calculated, and the collected signals are normalized.

[0008] The surface quality texture collection refers to: using a visual inspection system to collect the surface texture image of the workpiece obtained by processing. The visual inspection system includes: a linear moving slide, a high-magnification microscope and a microscope LED point light source. The computer controls the movement of the slide and the microscope to take pictures, thereby realizing the automatic collection of the surface texture information of the processed workpiece.

[0009] The robot joint current is preferably further subjected to principal component analysis, specifically: data dimension reduction processing is performed on the robot joint current information, and the joint current signal that is most obviously affected by the robot grinding state is selected in combination with the variance contribution rate, and the characteristic joint current signal is subsequently processed.

[0010] The principal component analysis specifically includes: decentralizing the current signal collected at high frequency of each joint of the robot; calculating the covariance matrix based on the decentralized robot joint current; calculating the eigenvalues ​​and eigenvectors of the covariance matrix using the singular value decomposition method; sorting the eigenvalues ​​of the covariance matrix from large to small; and selecting the robot joint current signal corresponding to the maximum eigenvalue as the characteristic joint current signal.

[0011] The conversion refers to: converting the joint current time series signal into a two-dimensional time-frequency diagram through wavelet transform, which is used as the original data set of the robot grinding state classification model, specifically:

[0012] Where: f(t) is the original signal, t is time, a is the scale factor, b is the time shift factor, ψ * is the conjugate function of ψ, ψ a,b (t) is the wavelet basis function,<x,y> is the inner product.

[0013] The polishing status label refers to: based on the grayscale co-occurrence matrix and its characteristic parameters, calculating the texture feature parameters corresponding to the collected surface image of the workpiece, completing the alignment of the robot joint current signal and the workpiece surface texture in the time domain, and automatically generating the robot polishing status label.

[0014] The texture feature parameter, i.e., the autocorrelation coefficient of the gray-level co-occurrence matrix, is obtained by first calculating the row mean of the gray-level co-occurrence matrix: and the column mean of the gray-level co-occurrence matrix Then calculate the row variance of the gray-level co-occurrence matrix and the column variance of the gray-level co-occurrence matrix Finally, the autocorrelation coefficient of the gray-level co-occurrence matrix is ​​calculated Where: G(i,j) is the element in the i-th row and j-th column of the gray-level co-occurrence matrix, μ i is the row mean of the gray-level co-occurrence matrix, μ j is the column mean of the gray-level co-occurrence matrix, S i is the row standard deviation of the gray-level co-occurrence matrix, S j is the column standard deviation of the gray-level co-occurrence matrix.

[0015] The polishing state classification model data set refers to: using a constrained autoencoder, encoding the converted two-dimensional time-frequency graph into a feature vector, using a linear interpolation method to interpolate in the feature space, and expanding the data set of the robot polishing state classification model.

[0016] The loss function of the constrained autoencoder is obtained by forming the input matrix F with the feature vectors generated by the encoder of the constrained autoencoder B×C , where: B is the batch size of network training, C is the number of channels of the feature vector, f b c is the element in the bth row and cth column of the input matrix; define the target uniform distribution as U(0,1); discretize U into B parts on the interval (0,1), and get the target vector B×1. Repeat the target vector C times and concatenate them to get the target matrix T B×C ,use is the element in the bth row and cth column of the target matrix; sort each channel of the input matrix, arrange the values ​​on the B dimension from small to large, and obtain the sorted input matrix F' B×C , is the element in the bth row and cth column of the sorted input matrix; calculate the loss function, the specific calculation formula is:

[0017] The polishing state classification model data set is preferably enhanced by feature vector encoding and original sample interpolation, wherein the original sample interpolation part is based on a hybrid idea and uses a linear interpolation method to construct a convex combination of existing samples to expand the number of original samples.

[0018] The linear interpolation specifically includes: completing the labeling of the feature vector encoded by the two-dimensional time-frequency graph based on the autocorrelation coefficient; selecting any two samples X in the same label class 1 and X 2 ; Based on X 1 and X 2 Construct a convex combination of the original samples X = γX 1 +(1-γ)X 2 ; Adjust the value of γ in the interval (0,1) to construct new and different samples.

[0019] The polishing state classification model outputs the robot polishing state such as under-polishing, normal polishing and over-polishing according to the two-dimensional time-frequency diagram obtained by wavelet transform and PCA processing of the joint current information, and iteratively trains the convolutional neural network based on the processing data set expanded by the constrained autoencoder, and finally uses the trained convolutional neural network to classify the polishing state.

[0020] The robot closed-loop feedback control refers to: conducting a grinding experiment on a new workpiece and completing surface texture information collection to verify the real-time monitoring effect of the robot grinding state classification model; based on the output results of the robot grinding state classification model, namely under-grinding, normal grinding and over-grinding states, adjusting the cutting amount, normal positive pressure and other parameters during robot grinding, changing the robot grinding state, and realizing robot closed-loop feedback control.

[0021] Technical Effects

[0022] Compared with the prior art, the present invention does not require additional external sensors and data acquisition equipment and only monitors the robot's grinding processing status based on the robot's body information, reducing the collection cost of the robot's grinding processing data set by at least 50%. It can complete efficient and accurate real-time monitoring of the robot's grinding status within 0.35 seconds, and the classification success rate of the robot's grinding status can reach up to 99.4625%. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the embodiment system;

[0024] Figure 2 This is the structure diagram of the constrained autoencoder;

[0025] Figure 3 Loss function change diagram for autoencoder training;

[0026] Figure 4The corresponding diagrams of the original collected data and the analysis results of the embodiment;

[0027] Figure 5 This is the encoding effect diagram of the constrained autoencoder;

[0028] Figure 6 is a flow chart of an embodiment;

[0029] Figure 7 This is a structural diagram of a visual measurement system for surface texture features of a workpiece processed in an embodiment;

[0030] Figure 8 A graph showing changes in network loss function and classification effect for the implementation example;

[0031] Fig. 9 This is a diagram corresponding to the online monitoring experimental platform for robot polishing status in the embodiment;

[0032] In the figure: 1 high-magnification microscope, 2 movable slide, 3 fixture, 4 measured workpiece, 5 pad, 6 end grinding head, 7 grinded workpiece, 8 fixture, 9 tooling platform, 10 industrial robot body, 11 robot control cabinet, 12 control module. DETAILED DESCRIPTION

[0033] like Figure 1 and Fig. 9 As shown, this embodiment involves an application scenario, including: a robot body, a grinding head, a control cabinet, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: the robot body performs workpiece grinding processing through the grinding head according to motion control information sent by the host computer and transmitted through the control cabinet, obtains the robot's body joint current information under different processing states, the host computer runs the robot grinding state classification model according to the high-frequency feedback of the robot body joint current information, judges the robot's grinding processing state within 0.35s and issues a motion adjustment instruction to the control cabinet, and in the host computer, performs data set expansion processing according to the collected robot grinding processing state information, expands the original processing process data by more than one time, reduces the processing data set acquisition cost, and is used for training the robot state classification model.

[0034] The six joints of the robot body are all rotary joints. The servo motor and driver are Panasonic MINASA6 series. The Beckhoff controller is used to control the Panasonic servo motor. Based on the real-time PLC kernel of the controller, the system can collect the status information of the servo motor at a maximum frequency of 20Khz.

[0035] The grinding tool selected in this embodiment is a Bosch Drummaster 4300 electric grinder, the maximum no-load speed can reach 15000r / min, and the selected grinding head is an 80-mesh flap wheel grinding head.

[0036] This embodiment relates to an online monitoring method for the grinding state based on the industrial robot body information based on the above application scenario, comprising the following steps:

[0037] Step 1: Conduct a robot pre-grinding experiment. By setting the feed rate of the robot grinding, construct the under-grinding, normal grinding and over-grinding states of the workpiece surface, and collect the current signals of the six joints of the robot under different processing states. Taking into account the number of subsequent data sets generated and the stability of the robot during operation, the PLC cycle time is set to 100us based on the Beckhoff real-time kernel. The current signal of each joint during robot grinding is collected at a frequency of 10Khz, and the percentage of the measured current to the rated current of the joint is calculated. The collected signal is normalized to avoid interference with the variance analysis in the subsequent principal component analysis due to the different magnitudes of the current of each joint. The influence of the robot grinding processing state on the current fluctuation of each joint is observed through normalization.

[0038] Step two, based on the principal component analysis method, the robot joint current information in step one is processed by data dimensionality reduction, and the joint current signal most obviously affected by the robot grinding state is selected in combination with the variance contribution rate, and the characteristic joint current signal is subsequently processed, wherein the principal component analysis of the robot joint current signal is specifically as follows: the current signal collected at a frequency of 10Khz for each joint of the robot is decentralized; the covariance matrix is ​​calculated based on the decentralized robot joint current; the eigenvalues ​​and eigenvectors of the covariance matrix are calculated using the singular value decomposition method; the eigenvalues ​​of the covariance matrix are sorted in descending order; and the robot joint current signal corresponding to the maximum eigenvalue is selected as the characteristic joint current signal.

[0039] Step 3: Divide the joint current signal selected in step 2 into 1024 points per frame. Based on wavelet transform, convert each 1024 points after division into a two-dimensional time-frequency graph to generate the original data set of the robot polishing state classification model. The one-dimensional continuous wavelet transform is specifically: Where: f(t) represents the original signal, t is time, a is the scale factor, b is the time shift factor, ψ * represents the conjugate function of ψ, ψ a,b (t) represents the wavelet basis function,<x,y> Represents the inner product.

[0040] Step 4: Use the built visual inspection system to automatically collect the surface texture information of the workpiece processed in step 1. The visual inspection system is as follows: Figure 7As shown, it includes: a linear moving slide, a high-magnification microscope and a microscope LED point light source, wherein: the computer controls the movement of the slide and the microscope to take pictures, and the visual inspection system can realize the automatic collection of the surface texture information of the workpiece. The magnification of the microscope, the running speed of the slide and the power of the light source will affect the imaging quality. The magnification of the microscope is selected as 300 times, the running speed of the moving slide is 2mm / s, and the power of the LED light source is 6W. In the design of the visual inspection system, the microscope is installed on the linear slide, and the workpiece to be measured is installed on the flat pad, which can avoid the problem of blurred microscope image acquisition caused by the deformation of the workpiece caused by gravity exceeding the depth of field of the microscope.

[0041] Step 5: Based on the gray-level co-occurrence matrix and its characteristic parameters, the texture characteristic parameters corresponding to the workpiece surface image collected in step 4 are calculated. Based on the texture characteristic parameters of the workpiece surface, after eliminating the uncertainty of manual identification, the autocorrelation coefficient is used to judge the grinding state of the robot. The robot joint current signal and the workpiece surface texture are aligned in the time domain, and the robot grinding state label is automatically generated.

[0042] like Figure 4 As shown in FIG. 1 , the original collected robot joint current signal and its analysis result are shown in FIG. 1 . The workpiece surface texture characteristic parameter in this embodiment is the autocorrelation coefficient of the gray-level co-occurrence matrix. The calculation process of the characteristic parameter is as follows: first, the row mean of the gray-level co-occurrence matrix is ​​calculated. and the column mean of the gray-level co-occurrence matrix Then calculate the row variance of the gray-level co-occurrence matrix and the column variance of the gray-level co-occurrence matrix Finally, the autocorrelation coefficient of the gray-level co-occurrence matrix is ​​calculated Where: G(i,j) represents the element in the i-th row and j-th column of the gray-level co-occurrence matrix, μ i represents the row mean of the gray-level co-occurrence matrix, μ j represents the column mean of the gray-level co-occurrence matrix, S i represents the row standard deviation of the gray-level co-occurrence matrix, S j represents the column standard deviation of the gray-level co-occurrence matrix.

[0043] Step 6: Build a constrained autoencoder in the Pytorch environment, encode the time-frequency diagram obtained in step 3 into a feature vector, use linear interpolation to interpolate in the feature space, and expand the data set of the robot polishing state classification model.

[0044] like Figure 2As shown, the dataset enhancement method designs a loss function of a constrained autoencoder and back-propagates to update the encoding layer parameters of the autoencoder, so that each channel of the encoded feature vector obeys a uniform distribution in the feature space, and adopts a linear interpolation method to construct a convex combination of existing samples to achieve dataset enhancement.

[0045] like Figure 3 As shown in the figure, the loss function of the constrained autoencoder changes with the number of training generations, and the final encoded feature vector result is as follows Figure 5 As shown, the loss function of the constrained autoencoder The specific calculation steps are as follows: the feature vectors generated by the encoder with constrained autoencoder form the input matrix F B×C , B is the batch size of network training, C is the number of channels of the feature vector, f b c Represents the element in the bth row and cth column of the input matrix; define the target uniform distribution as U(0,1); discretize U into B parts on the interval (0,1), obtain the target vector B×1, repeat the target vector C times and concatenate them, and obtain the target matrix T B×C ,use Represents the element in the bth row and cth column of the target matrix; sort each channel of the input matrix, arrange the values ​​on the B dimension from small to large, and obtain the sorted input matrix F' B×C , Represents the element in row b and column c of the sorted input matrix.

[0046] The dataset enhancement includes: feature vector encoding and original sample interpolation, wherein: the original sample interpolation part is based on the hybrid idea, and the linear interpolation method is used to construct a convex combination of existing samples to expand the number of original samples, specifically: based on the autocorrelation coefficient, the label of the feature vector encoded by the time-frequency graph is given; any two samples X in the same label class are selected 1 and X 2 ; Based on X 1 and X 2 Construct a convex combination of the original samples X = γX 1 +(1-γ)X 2 ; Adjust the value of γ in the interval (0,1) to construct new and different samples.

[0047] Step 7: Build a robot polishing state classification model based on convolutional neural network in the Pytorch environment and complete training and testing. Complete polishing test on new workpieces and collect workpiece surface images to verify the real-time monitoring effect of the robot polishing state classification model. The specific implementation process of the robot polishing state classification model is as follows: Figure 6 shown.

[0048] The input of the grinding state classification model is the time-frequency diagram obtained from the joint current information, and the output of the model is the robot grinding state such as under-grinding, normal grinding and over-grinding. The model is iteratively trained based on the expanded processing data set, and finally an effective robot grinding state classification model is obtained. The loss function and classification success rate of the robot grinding state classification model are shown in Figure 8 shown.

[0049] Step eight, apply the output results of the robot grinding state classification model to the robot closed-loop feedback control to improve the processing quality and efficiency of the robot grinding. Specifically: based on the under-grinding, normal grinding and over-grinding states output by the grinding state classification model, adjust the cutting amount, normal positive pressure and other parameters during robot grinding, change the robot grinding state, and realize the robot closed-loop feedback control.

[0050] After specific actual experiments, the above-mentioned device / method was run under the specific environment setting of 6000r / min preset speed of the electric spindle, 0.5-2.5mm preset grinding depth, and 5mm / s feed speed of the robot grinding. The experimental data obtained showed that the acquisition cost of the robot grinding processing data set can be reduced by at least 50%. In the absence of additional sensors and data acquisition equipment, the classification success rate of the robot grinding status can reach up to 99.4625%.

[0051] Compared with the prior art, the present invention can effectively reduce the collection cost of robot grinding processing data sets by more than 50%, can realize real-time monitoring of robot grinding status within 0.35 seconds at the fastest, and the classification success rate of robot grinding status can reach up to 99.4625%.

[0052] In summary, the present invention is applicable to general industrial robot grinding scenarios without additional sensors. It can extract the grinding state characteristics of the industrial robot based only on the joint current information of the robot body, and provides a characteristic joint current signal selection method based on principal component analysis. The present invention uses a constrained autoencoder to achieve uniform distribution in each dimension of the original data set in the feature space. Based on the hybrid idea, a convex combination of the original data and its labels is constructed to expand the data volume of the original data set, improve the generalization of the classification model, and reduce the acquisition cost of the processing data set. The present invention realizes real-time monitoring of the robot's grinding state, and can provide real-time feedback on the robot's motion control based on the monitoring results, thereby achieving high-quality grinding of the workpiece, which is of great significance to the large-scale application of robot grinding.

[0053] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. An online monitoring method for grinding status based on industrial robot body information, characterized in that: The robot joint current is collected during the pre-grinding experiment, and the surface quality texture is collected after the grinding is completed. The joint current time series signal is then converted into an image to generate a grinding state label and a grinding state classification model data set, and the data set is used to train the grinding state classification model. Finally, the trained grinding state classification model is used in the online stage to implement closed-loop feedback control of the robot. The collecting of robot joint currents refers to: collecting current signals of six joints of the robot under different processing states, and normalizing the collected current signals; The grinding state label refers to: based on the gray level co-occurrence matrix and its characteristic parameters, calculating the texture characteristic parameters corresponding to the surface image of the collected workpiece, completing the alignment of the robot joint current signal and the workpiece surface texture in the time domain, and automatically generating the robot grinding state label; The texture feature parameter, i.e., the autocorrelation coefficient of the gray-level co-occurrence matrix, is obtained by first calculating the row mean of the gray-level co-occurrence matrix: and the column mean of the gray-level co-occurrence matrix Then calculate the row variance of the gray-level co-occurrence matrix and the column variance of the gray-level co-occurrence matrix Finally, the autocorrelation coefficient of the gray-level co-occurrence matrix is ​​calculated Where: G(i,j) is the element in the i-th row and j-th column of the gray-level co-occurrence matrix, μ i is the row mean of the gray-level co-occurrence matrix, μ j is the column mean of the gray-level co-occurrence matrix, S i is the row standard deviation of the gray-level co-occurrence matrix, S j is the column standard deviation of the gray-level co-occurrence matrix.

2. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 is characterized in that: The pre-grinding experiment refers to: constructing the under-grinding, normal-grinding and over-grinding states of the workpiece surface by setting the feed rate of the robot grinding; The surface quality texture collection refers to: using a visual inspection system to collect the surface texture image of the workpiece obtained by processing. The visual inspection system includes: a linear moving slide, a high-magnification microscope and a microscope LED point light source. The computer controls the movement of the slide and the microscope to take pictures, thereby realizing the automatic collection of the surface texture information of the processed workpiece.

3. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 is characterized in that: The robot joint current is further subjected to principal component analysis, specifically: data dimension reduction processing is performed on the robot joint current information, and the joint current signal that is most obviously affected by the robot grinding state is selected in combination with the variance contribution rate, and the joint current signal is subsequently processed.

4. The method for online monitoring of grinding status based on industrial robot body information according to claim 3 is characterized in that: The principal component analysis specifically includes: decentralizing the current signal collected at high frequency of each joint of the robot; calculating the covariance matrix based on the decentralized robot joint current; calculating the eigenvalues ​​and eigenvectors of the covariance matrix using the singular value decomposition method; sorting the eigenvalues ​​of the covariance matrix from large to small; and selecting the robot joint current signal corresponding to the maximum eigenvalue as the characteristic joint current signal.

5. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 is characterized in that: The conversion refers to: converting the joint current time series signal into a two-dimensional time-frequency diagram through wavelet transform, which is used as the original data set of the robot grinding state classification model, specifically: Where: f(t) is the original signal, t is time, a is the scale factor, b is the time shift factor, ψ * is the conjugate function of ψ, ψ a,b (t) is the wavelet basis function,<x,y> is the inner product.

6. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 is characterized in that: The polishing state classification model data set refers to: using a constrained autoencoder, encoding the converted two-dimensional time-frequency graph into a feature vector, using a linear interpolation method to interpolate in the feature space, and expanding the data set of the robot polishing state classification model; The loss function of the constrained autoencoder is obtained by forming the input matrix F with the feature vectors generated by the encoder of the constrained autoencoder B×C , where: B is the batch number of network training, C is the number of channels of the feature vector, is the element in the bth row and cth column of the input matrix; define the target uniform distribution as U(0,1); discretize U into B parts on the interval (0,1), and get the target vector B×1. Repeat the target vector C times and concatenate them to get the target matrix T B×C ,use is the element in the bth row and cth column of the target matrix; sort each channel of the input matrix, arrange the values ​​on the B dimension from small to large, and obtain the sorted input matrix F′ B×C , is the element in the bth row and cth column of the sorted input matrix; calculate the loss function, the specific calculation formula is:

7. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 or 6, characterized in that: The polishing state classification model data set is enhanced by feature vector encoding and original sample interpolation, wherein the original sample interpolation part is based on a hybrid idea and uses a linear interpolation method to construct a convex combination of existing samples to expand the number of original samples.

8. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 is characterized in that: The polishing state classification model outputs the robot polishing state such as under-polishing, normal polishing and over-polishing according to the two-dimensional time-frequency diagram obtained by wavelet transform and PCA processing of the joint current information, and iteratively trains the convolutional neural network based on the processing data set expanded by the constrained autoencoder, and finally uses the trained convolutional neural network to classify the polishing state.

9. The method for online monitoring of grinding status based on industrial robot body information according to claim 1 is characterized in that: The robot closed-loop feedback control refers to: conducting a grinding experiment on a new workpiece and completing surface texture information collection to verify the real-time monitoring effect of the robot grinding state classification model; based on the output results of the robot grinding state classification model, namely under-grinding, normal grinding and over-grinding states, adjusting the cutting amount and normal positive pressure during robot grinding, changing the robot grinding state, and realizing robot closed-loop feedback control.

Images