Industrial robot failure diagnosis device and method using vibration data

Abstract

The present invention relates to a robot fault diagnosis device and method, and more specifically, to an industrial robot fault diagnosis device and method utilizing vibration data. According to one embodiment of the present invention, the present invention can diagnose a robot failure in real time. According to one embodiment of the present invention, the present invention can diagnose a robot failure in real time and provide a real-time notification to the user.

Description

Technology Field

[0001] The present invention relates to a robot fault diagnosis device and method, and more specifically, to an industrial robot fault diagnosis device and method utilizing vibration data. Background Technology

[0003] Industrial robots play an essential role in implementing automation processes across various fields, including general manufacturing, automotive, and semiconductors. As the utilization of industrial robots increases, troubleshooting and predictive maintenance become increasingly important.

[0004] Industrial robots diagnose and predict robot failures using signal data generated by the drive.

[0005] However, industrial robots have difficulty collecting failure data via drive signals due to robot security issues.

[0006] To solve these problems, it is necessary to provide a universal fault diagnosis device and method that collects external data and can be applied to various robots and diverse industrial environments.

[0007] The background technology of the present invention is disclosed in Korean Published Patent No. 10-2024-0074171. The problem to be solved

[0009] The present invention provides an industrial robot fault diagnosis device and method utilizing vibration data to perform real-time fault diagnosis based on various robot movements occurring in various work environments. means of solving the problem

[0011] According to one aspect of the present invention, a robot fault diagnosis device is provided.

[0012] A robot fault diagnosis device according to one embodiment of the present invention may include a data collection unit that collects data transmitted from a robot, a data preprocessing unit that preprocesses the data, and a fault diagnosis unit that analyzes the characteristic values ​​of the data to determine the state of the robot.

[0013] According to another aspect of the present invention, a robot fault diagnosis method is provided.

[0014] A robot fault diagnosis method according to one embodiment of the present invention may include the steps of collecting data transmitted from a robot, preprocessing the data, and determining the state of the robot by analyzing the characteristic values ​​of the data. Effects of the invention

[0016] According to one embodiment of the present invention, the present invention can diagnose a robot failure in real time.

[0017] According to one embodiment of the present invention, the present invention can diagnose a robot failure in real time and provide a real-time notification to the user.

[0018] The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the composition of the invention described in the description or claims of the present invention. Brief explanation of the drawing

[0020] FIGS. 1 and FIGS. 2 are drawings briefly illustrating a robot fault diagnosis device according to an embodiment of the present invention. FIG. 3 is a flowchart briefly illustrating a robot fault diagnosis method according to one embodiment of the present invention. FIG. 4 is a flowchart illustrating in detail a robot fault diagnosis method according to one embodiment of the present invention. FIG. 5 is a flowchart showing the data collection and preprocessing process in a robot fault diagnosis method according to one embodiment of the present invention. FIG. 6 is a diagram showing a vibration data feature value formula according to an embodiment of the present invention. FIG. 7 is a diagram showing sensor setting values ​​for vibration data collection according to an embodiment of the present invention. FIGS. 8 and 9 are graphs showing vibration characteristic values ​​according to normal and abnormal states according to an embodiment of the present invention. FIG. 10 is a flowchart illustrating a KNN method according to an embodiment of the present invention. FIG. 11 is a diagram showing a scatter plot comparing actual values ​​and predicted values ​​of a KNN method according to an embodiment of the present invention. FIG. 12 is a graph showing the training and test accuracy of a KNN method according to an embodiment of the present invention. FIG. 13 is a diagram showing the accuracy of the KNN method according to one embodiment of the present invention. FIG. 14 is a flowchart illustrating a CNN method according to an embodiment of the present invention. FIG. 15 is a diagram showing the accuracy of a CNN method according to an embodiment of the present invention. FIG. 16 is a diagram showing the structure of a CNN model according to one embodiment of the present invention. FIG. 17 is a diagram showing the optimal hyperparameter values ​​selected to increase the accuracy of the CNN method according to one embodiment of the present invention. FIG. 18 is a flowchart illustrating an improved CNN method according to an embodiment of the present invention. FIG. 19 is a diagram showing the accuracy of an improved CNN method according to an embodiment of the present invention. FIGS. 20 to 23 are drawings showing the learning results of a normal state and an abnormal state according to an embodiment of the present invention. Specific details for implementing the invention

[0021] The present invention is susceptible to various modifications and may have various embodiments; therefore, specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. In describing the present invention, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions may unnecessarily obscure the essence of the invention. Furthermore, singular expressions used in this specification and claims should generally be interpreted as meaning "one or more" unless otherwise stated.

[0022] Throughout the specification, when it is stated that a part is "connected (connected, in contact, combined)" with another part, this includes not only cases where they are "directly connected," but also cases where they are "indirectly connected" with other members interposed between them. Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but rather allows for the inclusion of additional components.

[0023] The terms used herein are used merely to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0024] The present invention will be described below with reference to the attached drawings. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification have been given similar reference numerals.

[0025] FIGS. 1 and FIGS. 2 are drawings briefly illustrating a robot fault diagnosis device according to one embodiment of the present invention.

[0026] Referring to FIG. 1, the robot fault diagnosis device (10) collects data from the robot (50), preprocesses the data, and diagnoses the fault of the robot (50).

[0027] The robot (50) collects data in real time from the sensor unit (110). The sensor unit (110) may include, for example, a vibration sensor. The vibration sensor may be attached to the 4th, 5th, and 6th axes of the robot (50) where the load is highest and movement is frequent.

[0028] The robot fault diagnosis device (10) acquires vibration data collected from the robot (50).

[0029] The robot fault diagnosis device (10) preprocesses the acquired vibration data on an embedded PC.

[0030] The robot fault diagnosis device (10) analyzes the key feature values ​​of the preprocessed data, organizes them into a table, and stores them in a database.

[0031] The robot fault diagnosis device (10) can diagnose a fault in the robot (50) by utilizing data and display the normal and abnormal status of the robot (50) on a screen to determine it in real time.

[0032] Referring to FIG. 2, the robot fault diagnosis device (10) includes a data collection unit (210), a data preprocessing unit (220), and a fault diagnosis unit (230).

[0033] The data collection unit (210) collects vibration data transmitted from the robot (50).

[0034] The data collection unit (210) collects vibration data generated at each axis of the robot (50) and receives real-time vibration signals of the robot (50) through the sensor unit (110) of the robot (50).

[0035] The data collection unit (210) detects the operating state of the robot (50) and provides data to distinguish between a normal state and an abnormal state.

[0036] Abnormal conditions include motion errors and end effector overloads.

[0037] A driving error means that an abnormal stop occurs during the operation of the robot (50), and an end effector overload indicates a state where the robot (50) is subjected to an excessive load during processing.

[0038] The data collection unit (210) transmits vibration data and delivers it to the data preprocessing unit (220).

[0039] The data preprocessing unit (220) preprocesses the vibration data collected from the data collection unit (210) on an embedded PC.

[0040] The data preprocessing unit (220) extracts key feature values ​​through filtering, noise removal, data normalization, and Fast Fourier Transform (FFT).

[0041] The Fast Fourier Transform (FFT) converts time-domain data into frequency-domain data to derive key feature values.

[0042] Key feature values ​​include RMS (Root Mean Square), Peak, Crest Factor, etc., and the feature values ​​are used for learning and analysis of the fault diagnosis unit (230).

[0043] The fault diagnosis unit (230) diagnoses a fault by analyzing the key feature values ​​of the data preprocessed by the data preprocessing unit (220) and determining the normal and abnormal state of the robot (50).

[0044] Abnormal conditions include drive errors and end effector overloads, and are classified based on the characteristic values ​​of vibration data according to each condition.

[0045] The fault diagnosis unit (230) learns vibration data using the KNN method, the CNN method, and the improved CNN method and diagnoses faults in real time.

[0046] The fault diagnosis unit (230) determines an abnormal state based on the analysis results and visualizes the diagnosis results to provide real-time notifications to the user.

[0047] The fault diagnosis unit (230) may include a CNN-based structure to learn and analyze feature values.

[0048] A CNN-based structure for analyzing feature values ​​may include a normalization layer and a data augmentation layer, followed by multiple convolution layers and pooling layers, and then a flatten layer and a dense layer.

[0049] For example, convolution layers and pooling layers can be placed repeatedly 6 times.

[0050] Through a CNN-based structure, complex patterns in data can be effectively learned, and fault diagnosis accuracy can be improved.

[0051] FIG. 3 is a flowchart briefly illustrating a robot fault diagnosis method according to one embodiment of the present invention.

[0052] In step S310, the robot fault diagnosis device (10) collects vibration data transmitted from the robot (50).

[0053] The collected data includes vibration data occurring at each axis of the robot (50).

[0054] In step S320, the robot fault diagnosis device (10) preprocesses the collected vibration data.

[0055] In step S330, the robot fault diagnosis device (10) diagnoses a fault by analyzing the characteristic values ​​of the preprocessed data and determining the state of the robot (50).

[0056] FIG. 4 is a flowchart illustrating in detail a robot fault diagnosis method according to one embodiment of the present invention.

[0057] In step S405, the robot fault diagnosis device (10) collects data of the robot (50). For example, the data may be vibration data of the 4th, 5th, and 6th axes of the robot (50).

[0058] In step S410, the robot fault diagnosis device (10) extracts feature values ​​of the data.

[0059] The characteristic values ​​of the data include RMS, Peak, Crest Factor, etc.

[0060] In step S415, the robot fault diagnosis device (10) transmits the characteristic value of the data to the embedded computer.

[0061] In step S420, the robot fault diagnosis device (10) performs communication and data preprocessing on an embedded computer.

[0062] In step S425, the robot fault diagnosis device (10) analyzes the characteristic values ​​of the data in the embedded computer.

[0063] In step S430, the robot fault diagnosis device (10) transmits the characteristic values ​​of the data analyzed by the embedded computer to the database.

[0064] In step S435, the robot fault diagnosis device (10) labels the feature values ​​of the data received from the database according to the state.

[0065] In step S440, the robot fault diagnosis device (10) collects fault diagnosis data from the database.

[0066] In step S445, the robot fault diagnosis device (10) analyzes data in real time to determine normal and abnormal states.

[0067] FIG. 5 is a flowchart illustrating the data collection and preprocessing process in a robot fault diagnosis method according to one embodiment of the present invention.

[0068] In step S505, the robot fault diagnosis device (10) collects the analog signal of the robot (50).

[0069] That is, in step S505, the robot fault diagnosis device (10) collects data of the robot (50).

[0070] For example, the data is data for learning about fault diagnosis, and the robot fault diagnosis device (10) can collect data by repeating 100 times for each of the three states.

[0071] The three states are normal state, robot drive error state, and end effector overload state. The robot drive error state is a state in which the robot (50) stops at any point during operation. The end effector overload state is an overload state caused by foreign matter or friction during processing of the robot (50), and may occur during material processing after processing has started.

[0072] In step S510, the robot fault diagnosis device (10) performs analog filtering.

[0073] The robot fault diagnosis device (10) can use low-frequency filtering as a filtering method to remove noise.

[0074] Low-pass filtering is designed using first and second-order Laplace transforms and is represented by Equations 1 and 2 below.

[0075] [Mathematical Formula 1]

[0076]

[0077] [Mathematical Formula 2]

[0078]

[0079] Here, is the Laplace transform variable and is the cutoff frequency.

[0080] The cutoff angular frequency is expressed by the following mathematical formula 3.

[0081] [Mathematical Formula 3]

[0082]

[0083] Here, is the cutoff frequency.

[0084] Analog filtering is applied before the ADC (Analog-to-Digital Converter) stage to remove high-frequency components and prevent aliasing.

[0085] In step S515, the robot fault diagnosis device (10) performs sampling.

[0086] The robot fault diagnosis device (10) converts a continuous signal into a discrete signal through Nyquist sampling.

[0087] In step S520, the robot fault diagnosis device (10) performs quantization.

[0088] The robot fault diagnosis device (10) quantizes the sampled signal to approximate the discrete value to an integer.

[0089] In step S525, the robot fault diagnosis device (10) performs encoding.

[0090] The robot fault diagnosis device (10) converts quantized data into binary numbers in a digital format.

[0091] In step S530, the robot fault diagnosis device (10) performs digital signal preprocessing.

[0092] That is, the robot fault diagnosis device (10) performs digital filtering and applies low-frequency filtering to the digitized data.

[0093] In step S535, the robot fault diagnosis device (10) extracts signal analysis and features.

[0094] Signal analysis is Fast Fourier Transform (FFT) analysis.

[0095] FFT analysis is a frequency-based method that converts vibration data into the frequency domain for analysis.

[0096] In other words, FFT analysis analyzes a Nyquist-sampled signal by converting it from the time domain to the frequency domain.

[0097] The formula for the FFT is represented by mathematical equation 4 below.

[0098] [Mathematical Formula 4]

[0099]

[0100] Here, is of the signal in the frequency domain It is the nth frequency component, and is of a signal in the time domain This is the nth sample.

[0101] The FFT result is expressed as amplitude or phase in the frequency spectrum.

[0102] The results of FFT analysis are used as data to derive key feature values ​​such as RMS, Peak, and Crest Factor.

[0103] FIG. 6 is a diagram showing a vibration data feature value formula according to one embodiment of the present invention.

[0104] Referring to FIG. 6, the robot fault diagnosis device (10) can analyze vibration data by separating it into feature values ​​in the time domain and the frequency domain.

[0105] The feature value facilitates analysis when the robot fault diagnosis device (10) processes a large amount of data.

[0106] Referring to Figure 6, the feature values ​​in the time domain have formulas for RMS, Peak, Peak-to-Peak, Mean, Variance, Crest Factor, Skewness, Kurtosis, and Standard Deviation.

[0107] RMS (Root Mean Square) is the average magnitude of a vibration signal.

[0108] Peak is the largest fluctuation value of the vibration signal.

[0109] Peak-to-peak is the difference between the maximum and minimum amplitudes of a vibration signal.

[0110] Mean represents the average center position of the vibration signal.

[0111] Variance is dispersion and indicates how spread out the vibration signal values ​​are from the mean value.

[0112] The Crest Factor evaluates the degree of distortion of a vibration signal.

[0113] Skewness is a value representing the asymmetry of the vibration signal distribution.

[0114] Kurtosis is a value that indicates the degree of peaking of a vibration signal distribution.

[0115] Standard Deviation is the standard deviation.

[0116] Referring to Figure 6, in the frequency domain, the feature values ​​have formulas for Frequency Center, Band Power, Mean Frequency, and Frequency Standard Deviation.

[0117] The frequency center is the average location of the frequency components of a vibration signal.

[0118] Band Power evaluates band-specific energy by calculating the energy of a vibration signal within a specific frequency band.

[0119] Mean Frequency represents the weighted average based on the energy of the frequency components.

[0120] Frequency Standard Deviation is the frequency standard deviation.

[0121] The robot fault diagnosis device (10) uses data obtained by performing a Fast Fourier Transform (FFT) on 1024 points in the frequency domain.

[0122] Frequency resolution is represented by the following mathematical formula 4.

[0123] [Mathematical Formula 4]

[0124]

[0125] Here, is frequency resolution, is the sampling frequency, is the number of samples.

[0126] Each frequency component appears at intervals of approximately 1.5625 Hz.

[0127] The present invention can analyze frequency components up to 800 Hz by considering the Nyquist frequency.

[0128] The low-frequency band is suitable for analyzing imbalances, alignment errors, mechanical loosening, bearing defects, wear, and friction in robot rotating parts.

[0129] FIG. 7 is a diagram showing sensor setting values ​​for vibration data collection according to one embodiment of the present invention.

[0130] Referring to Fig. 7, the Measuring Range is ±8g, the Sampling Frequency is 1600Hz, the Sampling Time is 300ms, and the Frequency Peak and Frequency Band are 0~600Hz.

[0131] The robot fault diagnosis device (10) saves vibration data collected in real time into a CSV file.

[0132] The robot fault diagnosis device (10) can calculate Peak, RMS, Crest Factor, and FFT values ​​as key feature values.

[0133] FIGS. 8 and 9 are graphs showing vibration characteristic values ​​according to normal and abnormal states according to an embodiment of the present invention.

[0134] Referring to FIGS. 8 and 9, the vibration characteristic value (Magnitude) is the vibration characteristic value of the 4 axes (4axis) and 5 axes (5axis) of the robot (50).

[0135] Referring to Fig. 8, the normal state and the abnormal state in which a driving error occurred over time are shown.

[0136] Figure 8 (a) shows a normal state, and Figure 8 (b) shows an abnormal state where a driving error has occurred. The driving error occurs due to an abnormal stop during robot operation, and the error can be detected if the Peak value and RMS value change rapidly.

[0137] Referring to FIG. 8, it can be observed that the robot (50) stops due to a driving error between 6 and 8 seconds. The Peak value and RMS value change clearly in this interval.

[0138] That is, the robot fault diagnosis device (10) can extract characteristics of driving errors through Peak values ​​and RMS values.

[0139] Referring to Fig. 9, it shows a normal state over time and an abnormal state where an end effector overload occurs.

[0140] Figure 9 (a) shows a normal state, and Figure 9 (b) shows an abnormal state where an end effector overload occurs. The end effector overload is caused by shock or friction during processing, and the overload can be detected when the Crest Factor value becomes abnormally high.

[0141] Referring to Fig. 9, it can be seen that an impact occurred due to overload between 18 and 21 seconds. The Peak value and Crest Factor value clearly change in this interval.

[0142] That is, the robot fault diagnosis device (10) can extract features of the end effector overload through the Peak value and Crest Factor value.

[0143] FIG. 10 is a flowchart illustrating a KNN method according to one embodiment of the present invention.

[0144] Referring to Figure 10, the KNN (K-Nearest Neighbors) method is a general machine learning method.

[0145] In step S1005, the robot fault diagnosis device (10) loads normal data and abnormal data from a CSV file.

[0146] In step S1010, the robot fault diagnosis device (10) divides the data into training data and test data.

[0147] In step S1015, the robot fault diagnosis device (10) checks whether the data is training data.

[0148] If the data is training data, proceed from step S1020, and if the data is test data, proceed from step S1025.

[0149] In step S1020, the robot fault diagnosis device (10) learns KNN with neighbor values.

[0150] In step S1025, the robot fault diagnosis device (10) predicts new data using a learned KNN model.

[0151] In step S1030, the robot fault diagnosis device (10) saves the learned KNN model to a file. Once the robot fault diagnosis device (10) saves the learned KNN model, it can be reused in a subsequent prediction step.

[0152] In step S1035, the robot fault diagnosis device (10) checks whether the neighbor value setting is finished.

[0153] If the neighbor value setting is finished, end the step, and if the neighbor value setting is not finished, return to step S1020.

[0154] FIG. 11 is a scatter plot showing the comparison of actual values ​​and predicted values ​​of a KNN method according to one embodiment of the present invention.

[0155] Referring to Figure 11, each point represents the actual value of the test data and the value predicted by the model.

[0156] The X-axis represents each test data sample, and the Y-axis represents the actual and predicted values ​​of that sample. The actual values ​​are the blue dots, which are the true values ​​that the model must predict. The predicted values ​​are the red dots, which are the values ​​predicted by the model for the test data. The closer the red dots are to the blue dots, the better the model predicted.

[0157] Referring to Figure 11, it can be seen that the prediction performance is good at values ​​of 0.02 or less.

[0158] FIG. 12 is a graph showing the training and test accuracy of a KNN method according to one embodiment of the present invention.

[0159] Referring to Figure 12, the value of k is set from 1 to 10, and the X-axis represents the number of neighbors, and the Y-axis represents accuracy.

[0160] The blue line is Train Accuracy, and the red line is Test Accuracy.

[0161] FIG. 13 is a diagram showing the accuracy of the KNN method according to one embodiment of the present invention.

[0162] Referring to Fig. 13, normal data, abnormal data, Mean Squared Error (MSE), and Accuracy are listed.

[0163] Accuracy is calculated by considering a value accurate when the difference between the actual value and the predicted value is within the tolerance.

[0164] Referring to Figure 13, the training accuracy is 84.34% and the test accuracy is 79.66%.

[0165] FIG. 14 is a flowchart illustrating a CNN method according to one embodiment of the present invention.

[0166] Referring to Figure 14, the CNN method is used to improve the accuracy of the KNN method. The CNN method takes image data as input, learns patterns, and can classify abnormal states with high accuracy.

[0167] Referring to Fig. 14, the CNN method is a convolutional neural network method and a deep learning method. Deep learning is based on a multilayer neural network and learns complex patterns through multiple layers.

[0168] The CNN method extracts features through convolution layers and pooling layers and performs classification based on them.

[0169] Convolution layers extract important features from images, and pooling layers increase computational efficiency by reducing dimensionality to select the most important information among these features.

[0170] The present invention sets the size of the image used for training to 640x480 and labels the normal state as 0 and the abnormal state as 1.

[0171] After training, the output values ​​are distinguished as 0 or 1, and all images are normalized to between 0 and 1 to unify the range of image values.

[0172] Image filtering may include a total of 6 filtering processes.

[0173] The CNN method automatically extracts image features through weight operations and is designed by considering the shape of the graph.

[0174] The CNN model consists of 6 image filtering layers. Each layer is computed using a 3x3 convolution filter.

[0175] The CNN method uses ReLU (Rectified Linear Unit) as the activation function.

[0176] The Max Pooling 2x2 layer is used to prevent overfitting and extracts only important features.

[0177] The present invention finds the point where the loss is minimized by differentiating the value of the loss function using a backpropagation method. To do this, the number of training iterations is set to 250.

[0178] In step S1405, the robot fault diagnosis device (10) receives an image as input data.

[0179] Steps S1410 through S1465 repeat the process of performing convolution operations, applying the activation function ReLU, and then performing max pooling operations 6 times.

[0180] In step S1410, the robot fault diagnosis device (10) performs a convolution operation and applies a ReLU activation function. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0181] In step S1415, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0182] In step S1420, the robot fault diagnosis device (10) performs a convolution operation and applies a ReLU activation function. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0183] In step S1425, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0184] In step S1430, the robot fault diagnosis device (10) performs a convolution operation and applies a ReLU activation function. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0185] In step S1435, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0186] In step S1440, the robot fault diagnosis device (10) performs a convolution operation and applies a ReLU activation function. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0187] In step S1445, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0188] In step S1450, the robot fault diagnosis device (10) performs a convolution operation and applies a ReLU activation function. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0189] In step S1455, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0190] In step S1460, the robot fault diagnosis device (10) performs a convolution operation and applies a ReLU activation function. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0191] In step S1465, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0192] In step S1470, the robot fault diagnosis device (10) adds a dense layer.

[0193] In step S1475, the robot fault diagnosis device (10) adds another dense layer.

[0194] In step S1480, the robot fault diagnosis device (10) returns a result through the output layer of the model.

[0195] FIG. 15 is a diagram showing the accuracy of a CNN method according to one embodiment of the present invention.

[0196] Referring to Figure 15, the accuracy of the CNN method for the rms 6-axis and crest 6-axis is shown.

[0197] Training Accuracy is the accuracy on the training dataset, and Validation Accuracy is the accuracy on the validation dataset not used for training.

[0198] If Training Accuracy is high and Validation Accuracy is low, it is highly likely that overfitting has occurred. If both values ​​are low, it may indicate an underfitting state where the model has not been trained sufficiently.

[0199] Training Loss is the loss calculated from the training data, and Validation Loss is the loss calculated from the validation data.

[0200] If Training Loss decreases but Validation Loss increases, it is highly likely that overfitting has occurred.

[0201] The blue line is Training, and the orange line is Validation.

[0202] The 6-axis accuracy of the CNN method is 95.63% RMS and 97.50% Crest.

[0203] FIG. 16 is a diagram showing the structure of a CNN model according to one embodiment of the present invention.

[0204] Referring to Fig. 16, the layer types include convolution, max pooling, and fully connected, and there are kernel size, stride, kernel number, output size, and parameters.

[0205] The number of kernels in the convolution layer and the max pooling layer is 16 each.

[0206] The output size of the convolution layer and the max pooling layer is 33×33×16 or 16×16×16 or 8×8×16 or 4×4×16.

[0207] The kernel size of the convolution layer is 3×3 or 11×11.

[0208] The stride of the convolution layer is 1.

[0209] The parameters of the convolution layer are 2320 or 1952.

[0210] The kernel size of the max pooling layer is 2×2.

[0211] The stride of the max pulling layer is 2.

[0212] The kernel size of the fully connected layer is 100, the number of kernels is 1, the output size is 100×1, and the parameters are 25700.

[0213] Referring to Fig. 16, the CNN model consists of six image filtering layers. Each layer is computed using a 3x3 convolution filter.

[0214] The CNN method uses ReLU (Rectified Linear Unit) as the activation function.

[0215] The Max Pooling 2x2 layer extracts only important features to prevent overfitting.

[0216] The present invention finds the point where the loss is minimized by differentiating the value of the loss function using a backpropagation method. To do this, the number of training iterations is set to 250.

[0217] Referring to FIG. 16, the CNN model structure may include a normalization layer and a data augmentation layer, a plurality of convolution layers and a plurality of pooling layers, and then a flatten layer and a dense layer.

[0218] For example, convolution layers and pooling layers can be arranged repeatedly 6 times.

[0219] For example, a plurality of convolution layers may include a first convolution layer, a second convolution layer, and a third convolution layer.

[0220] For example, a plurality of pooling layers may include a first pooling layer disposed between a first convolution layer and a second convolution layer, and a second pooling layer disposed between a second convolution layer and a third convolution layer.

[0221] For example, a convolution layer can include a ReLU activation function.

[0222] For example, a convolution layer may include L2 normalization and batch normalization.

[0223] FIG. 17 is a diagram showing the optimal hyperparameter values ​​selected to increase the accuracy of the CNN method according to one embodiment of the present invention.

[0224] Referring to Figure 17, the hyperparameters include the optimizer, learning rate, activation function, and batch size.

[0225] Referring to FIG. 17, the robot fault diagnosis device (10) uses Adam as an optimizer. Adam has high computational efficiency and excellent image processing performance, and operates stably even with unbalanced data or complex data sets.

[0226] The robot fault diagnosis device (10) uses ReLU as the activation function. ReLU effectively handles non-linearity and is suitable for learning complex data structures.

[0227] The robot fault diagnosis device (10) uses le-3 as the learning rate.

[0228] The robot fault diagnosis device (10) uses 32 as the batch size.

[0229] The present invention loads graph images of normal and abnormal states to apply a robot fault diagnosis method, divides them into a training set and a test set, and assigns normal and abnormal labels to each.

[0230] The training data consists of 200 normal-state data, 200 abnormal-state drive error data, and 200 abnormal-state end-effector overload data, of which 20% is used as test data to evaluate model performance.

[0231] FIG. 18 is a flowchart illustrating an improved CNN method according to one embodiment of the present invention.

[0232] Referring to Fig. 18, the improved CNN method applies a normalization technique during the convolution operation to improve the performance of the conventional method. The normalization technique includes L2 normalization, batch normalization, and dropout.

[0233] In step S1805, the robot fault diagnosis device (10) receives an image as input data.

[0234] In step S1810, the robot fault diagnosis device (10) normalizes the input data.

[0235] In step S1815, the robot fault diagnosis device (10) performs data augmentation on the input data.

[0236] Steps S1820 through S1865 repeat the process of performing convolution operations, applying the rectified linear function (ReLU) as an activation function, performing L2 normalization and batch normalization, and then performing max pooling operations 5 times.

[0237] In step S1820, the robot fault diagnosis device (10) performs a convolution operation, applies a ReLU activation function, and performs L2 normalization and batch normalization. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0238] In step S1825, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0239] In step S1830, the robot fault diagnosis device (10) performs a convolution operation, applies a ReLU activation function, and performs L2 normalization and batch normalization. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0240] In step S1835, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0241] In step S1840, the robot fault diagnosis device (10) performs a convolution operation, applies a ReLU activation function, and performs L2 normalization and batch normalization. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0242] In step S1845, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0243] In step S1850, the robot fault diagnosis device (10) performs a convolution operation, applies a ReLU activation function, and performs L2 normalization and batch normalization. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0244] In step S1855, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0245] In step S1860, the robot fault diagnosis device (10) performs a convolution operation, applies a ReLU activation function, and performs L2 normalization and batch normalization. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0246] In step S1865, the robot fault diagnosis device (10) performs a Max Pooling operation, and the pooling size is 2×2.

[0247] In step S1870, the robot fault diagnosis device (10) performs a convolution operation, applies a ReLU activation function, performs L2 normalization and batch normalization, and applies dropout. The number of kernels is 16, the kernel size is 3×3, and the stride is 1.

[0248] In step S1875, the robot fault diagnosis device (10) performs a flattening step of unfolding a 2D feature map into a 1D vector.

[0249] In step S1880, the robot fault diagnosis device (10) adds a dense layer and performs L2 normalization and batch normalization.

[0250] In step S1885, the robot fault diagnosis device (10) returns a result through the output layer of the model.

[0251] FIG. 19 is a diagram showing the accuracy of an improved CNN method according to one embodiment of the present invention.

[0252] Referring to Figure 19, the accuracy of the improved CNN method for the rms 6-axis and crest 6-axis is shown.

[0253] Training accuracy is the accuracy on the training dataset, and validation accuracy is the accuracy on the validation dataset not used for training.

[0254] If Training Accuracy is high and Validation Accuracy is low, it is highly likely that overfitting has occurred. If both values ​​are low, it may indicate an underfitting state where the model has not been trained sufficiently.

[0255] Training loss is the loss calculated from the training data, and validation loss is the loss calculated from the validation data.

[0256] If Training Loss decreases but Validation Loss increases, it is highly likely that overfitting has occurred.

[0257] Referring to Figure 19, the blue line is Training and the orange line is Validation.

[0258] The 6-axis accuracy of the improved CNN method is improved with an RMS of 98.75% and a Crest of 99.37%.

[0259] FIGS. 20 to 23 are drawings showing the learning results of a normal state and an abnormal state according to an embodiment of the present invention.

[0260] Training accuracy refers to the model's accuracy with respect to the training data, and as training progresses, it gradually increases, allowing the model to predict more accurately.

[0261] Test accuracy is a metric that evaluates how well a trained model predicts new data.

[0262] It is desirable that the difference in accuracy between the training data and the test data is not large, and in such cases, it can be said that the model has been well trained without being overfitted.

[0263] Training loss is the value of the loss function calculated from the training data, quantifying the difference between the values ​​predicted by the model and the actual values. The smaller this value, the more accurate the model's prediction.

[0264] Test loss is generally similar to or slightly larger than training loss, and if overfitting occurs, test loss increases rapidly.

[0265] Referring to Figure 20, the learning results for a normal state and an abnormal state in which a driving error occurred are shown.

[0266] An abnormal state in which a driving error occurs is when the Train Loss and Test Loss decrease rapidly during the early stages of training and then stabilize, which means that it is operating stably without overfitting.

[0267] Referring to Figure 20, the blue line is Training and the orange line is Validation.

[0268] Referring to Figure 21, the training accuracy and test accuracy of an abnormal state in which a driving error occurred are shown.

[0269] Referring to Fig. 21, normal data, abnormal data, accuracy, and loss are listed.

[0270] Referring to Figure 21, the training accuracy is 99.51% and the test accuracy is 98.67%.

[0271] Referring to Fig. 22, the learning results for a normal state and an abnormal state where end effector overload occurs are shown.

[0272] Referring to Fig. 22, the abnormal state where end effector overload occurs is that both Train Loss and Test Loss values ​​are very small, less than 0.01, which means that the model prediction was performed very accurately.

[0273] Referring to Figure 22, the blue line is Training and the orange line is Validation.

[0274] Referring to Figure 23, the training accuracy and test accuracy in an abnormal state where end effector overload occurs are shown.

[0275] Referring to Fig. 23, normal data, abnormal data, accuracy, and loss are listed.

[0276] Referring to Figure 23, the training accuracy is 99.81% and the test accuracy is 99.87%.

[0277] The above-described autonomous driving control transfer notification device and method may be implemented as computer-readable code on a computer-readable medium. The computer-readable recording medium may be, for example, a removable recording medium (CD, DVD, Blu-ray disc, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-equipped hard disk). The computer program recorded on the computer-readable recording medium may be transmitted to another computing device via a network such as the Internet and installed on the other computing device, thereby allowing it to be used on the other computing device.

[0278] Although it has been described above that all components constituting an embodiment of the present invention are combined or operate as a single unit, the present invention is not necessarily limited to such an embodiment. That is, within the scope of the purpose of the present invention, all components may be selectively combined in one or more ways to operate.

[0279] Although operations are depicted in a specific order in the drawings, it should not be understood that the operations must necessarily be executed in the specific order depicted or in a sequential order, or that all depicted operations must be executed to obtain the desired result. In certain situations, multitasking and parallel processing may be advantageous. Furthermore, the separation of various configurations in the embodiments described above should not be understood as necessarily required, and it should be understood that the described program components and systems can generally be integrated together into a single software product or packaged into multiple software products.

[0280] The present invention has been described above with reference to its embodiments. Those skilled in the art will understand that the present invention may be embodied in modified forms without departing from the essential characteristics of the invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the invention is defined by the claims, not by the foregoing description, and all variations within the scope of the claims should be interpreted as being included in the invention. Explanation of the symbols

[0283] 10: Robot fault diagnosis device 50: Robot 110: Sensor section 210: Data Collection Unit 220: Data Preprocessing Section 230: Fault Diagnosis Unit

Claims

Claim 1 A robot fault diagnosis device comprises: a data collection unit that collects vibration data generated from a plurality of axes of the robot through a vibration sensor installed in the robot; and a data preprocessing unit that preprocesses the vibration data. A robot fault diagnosis device comprising a fault diagnosis unit that determines the state of the robot by analyzing feature values ​​of preprocessed data, wherein the data preprocessing unit performs at least one of filtering, noise removal, data normalization, and Fast Fourier Transform (FFT) to extract feature values ​​including RMS (Root Mean Square), Peak, and Crest Factor, wherein the fault diagnosis unit classifies the state of the robot into normal and abnormal states using an improved CNN model that takes the feature values ​​as input, wherein the abnormal state includes motion error and end effector overload, and provides notification to the user by visualizing the judgment result in real time, wherein the improved CNN model has a structure in which a normalization layer and a data augmentation layer are arranged, and a plurality of convolution layers and pooling layers are repeatedly arranged, followed by a flatten layer and a density layer, and is configured to perform L2 normalization and batch normalization during the convolution operation and apply dropout to prevent data overfitting and improve the accuracy of learning feature patterns of vibration data. Claim 2 delete Claim 3 delete Claim 4 A robot fault diagnosis device according to claim 1, characterized in that the vibration data is collected from at least one of the 4th, 5th, and 6th axes of the robot. Claim 5 delete Claim 6 delete Claim 7 delete Claim 8 A robot fault diagnosis method performed by a robot fault diagnosis device comprises: a step of collecting vibration data generated from a plurality of axes of a robot through a vibration sensor; and a step of performing preprocessing on the vibration data. A robot fault diagnosis method comprising a step of determining the state of the robot by analyzing feature values ​​of preprocessed data, wherein the step of performing preprocessing on the vibration data includes a step of extracting feature values ​​including RMS, Peak, and Crest Factor by performing at least one of filtering, noise removal, data normalization, and Fast Fourier Transform (FFT), wherein the step of determining the state of the robot by analyzing feature values ​​of the preprocessed data classifies the state of the robot into normal and abnormal states using an improved CNN model that takes the feature values ​​as input, wherein the abnormal state includes motion error and end effector overload, and the determination result is visualized in real time to provide notification to the user, wherein the improved CNN model has a structure in which a normalization layer and a data augmentation layer are arranged, and a plurality of convolution layers and pooling layers are repeatedly arranged, followed by a flatten layer and a dense layer, and is configured to perform L2 normalization and batch normalization during the convolution operation and apply dropout to prevent data overfitting and improve the accuracy of learning feature patterns of vibration data. Claim 9 delete Claim 10 delete

Images