CN114282577A - Industrial robot fault diagnosis method based on current signal analysis - Google Patents
Industrial robot fault diagnosis method based on current signal analysis Download PDFInfo
- Publication number
- CN114282577A CN114282577A CN202111596134.0A CN202111596134A CN114282577A CN 114282577 A CN114282577 A CN 114282577A CN 202111596134 A CN202111596134 A CN 202111596134A CN 114282577 A CN114282577 A CN 114282577A
- Authority
- CN
- China
- Prior art keywords
- data
- layer
- fault
- neural network
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Manipulator (AREA)
Abstract
The invention relates to a fault diagnosis method for an industrial robot for current signal analysis, which expands the application range of a convolutional neural network in the field of robot fault analysis; the data source is a real fault in the running process of the robot, and the analysis result is more real; the structure setting of one-dimensional convolution on the original data is changed into a two-dimensional convolution structure, so that the correlation among different sections of signals is increased, the elasticity of the network is increased, the operation speed is increased, and the settings of convolution kernels and step length parameters are reduced. According to actual experimental test results, network performances before and after introduction are compared, a good fault prediction effect can be achieved without introducing a discarding layer and an activating layer, the structure of the convolutional neural network can be further simplified, and a better diagnosis effect is achieved. The method can effectively solve the problem of fault diagnosis of the industrial robot, and provides a new idea for subsequent industrial robot health management research.
Description
Technical Field
The invention relates to an industrial robot fault diagnosis technology, in particular to an industrial robot fault diagnosis method based on current signal analysis.
Background
In the field of fault diagnosis of industrial robots, fault conditions of the robots can be distinguished through signal characteristic value extraction and analysis. However, most fault analysis algorithms need to use expert knowledge in the field of signal analysis to manually extract appropriate feature values, that is, the feature values are selected with a certain subjectivity. In practical application, the selected fault diagnosis method cannot be guaranteed to be most suitable for the applied scene. The neural network can automatically extract feature values in the signal, and can perform multi-classification based on the extracted feature values.
Chinese application patent CN 112651491a discloses a fault prediction method for an industrial robot based on a CNN convolution model (hereinafter referred to as patent a), where patent a performs one-dimensional convolution on a vibration signal of the robot, thereby well implementing fault prediction for the robot, but there are some improved spaces.
Disclosure of Invention
In order to further improve the accuracy of the convolutional neural network applied to robot fault diagnosis, an industrial robot fault diagnosis method based on current signal analysis is provided.
The technical scheme of the invention is as follows: a fault diagnosis method for an industrial robot based on current signal analysis specifically comprises the following steps:
1) setting sampling frequency according to the cycle duration of single action of the robot to collect current signals of normal operation and fault operation of the experimental robot, and collecting the signals as original data;
2) signal data extraction: extracting data from the original data by adopting a random sliding windowing method, wherein the set window length n is greater than the number of sampling data corresponding to the period duration of a single action of the robot; the same amount of normal operation data and fault operation data is extracted to form a data set,
3) data preprocessing: dividing each one-dimensional data in the data set in the step 2) into n/2 sections, wherein each section comprises n/2 sampling point data, and each data is changed into n/2 x n/2 two-dimensional data; processing all data in the data set into two-dimensional data as input of a convolutional neural network;
4) building and training a convolutional neural network model:
building a convolutional neural network model: the convolutional neural network sequentially comprises a first convolutional layer, a first pooling layer treatment, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer treatment, an expansion layer, a first full-connection layer and a second full-connection layer;
dividing the two-dimensional gray scale image preprocessed in the step 3) into a training set and a testing set, inputting training set data into an established convolutional neural network model for training, setting a loss function as a cross entropy loss function, setting an optimizer as an Adam optimizer, training until a training error reaches a set value, verifying the generalization capability of the convolutional neural network model by using the testing set data, continuously iterating until the network performance is qualified, and storing parameters of the convolutional neural network model;
5) and (3) robot fault diagnosis:
and (3) after the preprocessing operation of the step 2) and the step 3) is carried out on the newly acquired robot current data, inputting the current data into the well trained convolutional neural network model in the step 4), and outputting the fault condition of the robot.
Further, the period duration of a single action of the robot in the step 1) is about 2s, and the set signal sampling frequency is 2.5khz, so that each period contains about 5000 sampling points.
Further, in the step 2), the window length is 10000, 20 groups of data are selected in the normal operation daily window, 80 groups of data are selected in the fault operation daily window, and data collected in 4 days of normal operation and data collected in 1 day of fault operation are selected as a data set.
Further, in the step 4), the number of input channels of the first convolution layer is 1, the number of output channels is 32, the convolution kernel size is 5 × 5, the step size is 1, and the boundary is filled with 0 for 2 pixel lengths;
the first pooling layer mode is maximum pooling, the size of a pooling core is 2 x 2, and the step length is 2;
the number of input channels of the second convolution layer is 32, the number of output channels of the second convolution layer is 32, the convolution kernel size is 5 x 5, the step length is 1, and the boundary is filled with 0 for 2 pixel lengths;
the second pooling layer pattern is maximal pooling, the pooled core size is 2 x 2, and the step size is 2;
the number of input channels of the third convolution layer is 32, the number of output channels is 64, the convolution kernel size is 5 x 5, the step size is 1, and the boundary is filled with 0 for 2 pixel lengths;
the third pooling layer mode is maximum pooling, the size of a pooling core is 2 x 2, and the step length is 2;
the expansion layer expands the multi-channel two-dimensional data output by the third pooling layer into a one-dimensional vector;
the number of output channels of the one-dimensional vector output by the expansion layer sequentially passes through the first layer of full-connection layer and the second layer of full-connection layer is 2.
Further, the numerical range of the two-dimensional data in the step 3) is mapped to 0-255, and a visible n/2 × n/2 two-dimensional gray scale map is obtained and serves as the input of the convolutional neural network.
The invention has the beneficial effects that: the industrial robot fault diagnosis method based on current signal analysis expands the application range of the convolutional neural network in the field of robot fault analysis; the data source is a real fault in the running process of the robot, and the analysis result is more real; the structure of the convolutional neural network is improved, so that more information can be reserved, and a better diagnosis effect is achieved. The method can effectively solve the problem of fault diagnosis of the industrial robot, and provides a new idea for subsequent industrial robot health management research.
Drawings
FIG. 1 is a flow chart of a fault diagnosis method for an industrial robot based on current signal analysis according to the present invention;
FIG. 2 is a graph of training errors over 100 training periods in the method of the present invention;
FIG. 3 is a graph of test set accuracy over 100 training periods of the method of the present invention;
FIG. 4 is a diagram of a convolutional neural network structure in the method of the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
The invention provides an industrial robot fault diagnosis method using a convolutional neural network for current signal analysis, which mainly comprises the following steps: the specific flow of the method comprises the steps of current signal acquisition, signal prediction processing, neural network model construction and fault diagnosis, and is shown in figure 1.
S1: acquisition of the relevant current signals using special instrumentation
The data acquisition equipment is a signal acquisition box independently developed by Shanghai robot industrial and technical research institute, the acquisition device supports 16 signal input channels at most, and the sampling frequency can be set to 10khz at most; the data acquisition object is an industrial robot which performs Mean Time Between Failure (MTBF) experiments in a research institute. As the robot has faults in the experimental process or after the service life is over, the current signal data of the normal operation and fault operation of the robot can be acquired.
S2: signal preprocessing method
Setting a robot running program to obtain the cycle duration of a single action of the robot about 2s, and setting the sampling frequency of a signal acquisition box to be 2.5khz, so that each cycle comprises about 5000 sampling points; the method adopts a random sliding windowing method to extract data from an original signal, and in order to ensure that sampling point data in a window can contain data of a complete period, the set window length needs to be longer than a single period, so the window length is set to 10000.
According to the parameter setting, selecting data collected in 24 days of normal operation and data collected in 3 days of fault operation as training set data, wherein 20 groups of data are selected in a window of each day of normal operation, 160 groups of data are selected in a window of each day of fault operation, and then the training set can obtain 24 × 20-480 normal data and 3 × 160-480 fault data; selecting data collected in 4 days of normal operation and data collected in 1 day of fault operation as test set data, wherein 20 groups of data are selected in normal operation every day, 80 groups of data are selected in fault operation every day, and the test set can obtain 4 × 20-80 normal data and 1 × 80-80 fault data;
the dimensionality of each datum is 1 x 10000, each datum is divided into 100 sections in order to enable the datum to be in accordance with the input format of the convolutional neural network, each section comprises two-dimensional data of 100 sampling points, and the numerical range is mapped to 0-255. A 100 x 100 two-dimensional gray scale map can be obtained. For such 1 channel, a 100 x 100 gray scale map is used as an input to the convolutional neural network.
S3: building and training a convolutional neural network model:
building a convolutional neural network model:
step S31: performing a first layer convolution layer processing, wherein the number of input channels of the first convolution layer is 1, the number of output channels of the first convolution layer is 32, the size of a convolution kernel is 5 x 5, the step size is 1, and the boundary is filled with 0 for 2 pixel lengths; data with dimension 1 × 100 is input, and the data with dimension 32 × 100 is obtained after the first layer of convolution layer processing.
Step S32: a first layer of pooling layer treatment, wherein the first layer of pooling layer mode is maximum pooling, the size of a pooling core is 2 x 2, and the step size is 2; data with dimension 32 x 100 was input, and processing through the first pooled layer resulted in data with dimension 32 x 50.
Step S33: a second convolution layer processing, the number of input channels of the second convolution layer is 32, the number of output channels of the second convolution layer is 32, the convolution kernel size is 5 x 5, the step size is 1, and the boundary is filled with 0 for 2 pixel lengths; data with dimension 32 x 50 was input, and the second layer convolution processing resulted in 32 x 50 data.
Step S34: a second layer of pooling layer treatment, the second layer of pooling layer mode being maximum pooling, a pooling core size of 2 x 2, a step size of 2; data with dimension 32 x 50 was input, and processing through the second pooled layer resulted in data with dimension 32 x 25.
Step S35: a third layer of convolution layer processing, wherein the number of input channels of the third convolution layer is 32, the number of output channels of the third convolution layer is 64, the size of a convolution kernel is 5 x 5, the step size is 1, and the boundary is filled with 0 for 2 pixel lengths; data with dimension 32 × 25 is input, and 64 × 25 data is obtained through the third layer of convolution processing.
Step S36: processing a third pooling layer, wherein the mode of the third pooling layer is maximum pooling, the size of a pooling core is 2 x 2, and the step length is 2; data with dimension 64 × 25 was input, and processing through the third pooled layer resulted in data with dimension 64 × 12.
Step S37: and (3) unfolding layer treatment: the processed multi-channel two-dimensional data is unfolded into a one-dimensional vector; data with dimension 64 × 12 is input, and the data of 1 × 9216 is obtained through expansion layer processing.
Step S38: first full tie layer: the number of input channels is 9216, and the number of output channels is 1024;
step S39: second layer full connection layer: the input channel number is 1024, and the output channel number is 2.
Training: dividing the preprocessed data into a training set and a test set, inputting the training set data into the network for training, setting a loss function as a cross entropy loss function, setting an optimizer as an Adam optimizer, training until a training error reaches a set value, verifying the generalization capability of the network by using the test set data, continuously iterating until the network performance is qualified, and storing the parameters of a convolutional network model. FIG. 2 is a graph of training errors during 100 training periods in the method of the present invention; FIG. 3 is a graph of test set accuracy over 100 training periods of the method of the present invention. The convolutional neural network structure diagram is shown in fig. 4.
S4: robot fault diagnosis
And inputting the newly acquired data into the stored trained neural network after the preprocessing operation, and outputting the fault condition of the robot.
The method is applied to robot fault recognition in the existing convolutional neural network, and is improved by the following steps:
sources of fault data: further verifying the effectiveness of the convolutional neural network by using the real fault data;
the structure and parameter setting method of the neural network comprises the following steps: from the structural point of view, the method changes the structural setting of performing one-dimensional convolution on the original data into a two-dimensional convolution structure, and compared with the two-dimensional convolution, the two-dimensional convolution has the following advantages: (1) the operation speed is faster: taking the data dimension provided in the specific implementation in the application of the existing convolutional neural network as an example, the data dimension is calculated 500 times by using a convolutional kernel with the size of 16 and performing one-dimensional convolution on 8000 data points with the step size of 16, and if the two-dimensional convolution is used, the 16 × 16 convolutional kernel performs two-dimensional convolution on 96 × 96 data (edge expansion is required) with the step size of 16 × 16 and requires 36 times of calculation, so that the operation speed of the two-dimensional convolution can be improved, and the improvement effect is more obvious when the data amount is larger; (2) the network model is more flexible: because the two-dimensional convolution needs to stack original signals into two-dimensional data in a segmented mode, association among different segment signals is increased, so that the neural network can extract features from the relationship which is artificially increased, and the elasticity of the network is increased. (3) Data visualization: original data are converted into a two-dimensional gray scale map, so that the data are convenient to visualize. The method also simplifies a discarding layer and an activating layer in the application of the conventional convolutional neural network, compares the network performances before and after introduction according to the actual experimental test result, obtains the application scene of current signal analysis of the industrial robot, and can realize a better fault prediction effect without introducing the discarding layer and the activating layer, so that the method does not introduce more network structures, simplifies a convolutional neural network model, and can still achieve a better effect. From the aspect of parameter setting, the step length is set to be the same as the size of a convolution kernel and is set to be larger in the application of the conventional convolution neural network, and the setting has the advantage that the calculation speed can be increased, but information among different data sections can be lost.
Form of signal data: the method provided in the application of the existing convolutional neural network aims at vibration signals, is a classical industrial robot signal analysis form, and needs to be expanded on the analysis of other signals. The method of the invention expands the application of the convolutional neural network in current signal analysis. Compared with the vibration signal analysis, the current signal analysis has the following advantages: (1) vibration signal need place the sensor in the robot outward appearance, and under operating condition, the motion variety of arm has brought the difficulty for the sensor is walked the line design, and current signal then reads data from the switch board, is fit for practical application scene more. (2) The vibration signal is easily influenced by environmental vibration, under the actual working condition, a plurality of robots often work together, and other equipment which can cause vibration can also exist in the environment, so that a plurality of environmental errors are introduced into the acquired vibration signal, the current signal is not easily interfered by the outside, the data error is smaller, and the running condition of the robot can be reflected more truly.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (5)
1. A fault diagnosis method for an industrial robot based on current signal analysis is characterized by comprising the following steps:
1) setting sampling frequency according to the cycle duration of single action of the robot to collect current signals of normal operation and fault operation of the experimental robot, and collecting the signals as original data;
2) signal data extraction: extracting data from the original data by adopting a random sliding windowing method, wherein the set window length n is greater than the number of sampling data corresponding to the period duration of a single action of the robot; the same amount of normal operation data and fault operation data is extracted to form a data set,
3) data preprocessing: dividing each one-dimensional data in the data set in the step 2) into n/2 sections, wherein each section comprises n/2 sampling point data, and each data is changed into n/2 x n/2 two-dimensional data; processing all data in the data set into two-dimensional data as input of a convolutional neural network;
4) building and training a convolutional neural network model:
building a convolutional neural network model: the convolutional neural network sequentially comprises a first convolutional layer, a first pooling layer treatment, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer treatment, an expansion layer, a first full-connection layer and a second full-connection layer;
dividing the two-dimensional gray scale image preprocessed in the step 3) into a training set and a testing set, inputting training set data into an established convolutional neural network model for training, setting a loss function as a cross entropy loss function, setting an optimizer as an Adam optimizer, training until a training error reaches a set value, verifying the generalization capability of the convolutional neural network model by using the testing set data, continuously iterating until the network performance is qualified, and storing parameters of the convolutional neural network model;
5) and (3) robot fault diagnosis:
and (3) after the preprocessing operation of the step 2) and the step 3) is carried out on the newly acquired robot current data, inputting the current data into the well trained convolutional neural network model in the step 4), and outputting the fault condition of the robot.
2. The method for diagnosing the fault of the industrial robot based on the current signal analysis as claimed in claim 1, wherein the period duration of the single action of the robot in the step 1) is about 2s, and the signal sampling frequency is set to be 2.5khz, so that each period comprises about 5000 sampling points.
3. The method for diagnosing the fault of the industrial robot based on the current signal analysis as claimed in claim 2, wherein the window length in the step 2) is 10000, 20 groups of data are selected in the normal operation daily window, 80 groups of data are selected in the fault operation daily window, and data collected in the normal operation for 4 days and data collected in the fault operation for 1 day are selected as the data sets.
4. The method for diagnosing the failure of the industrial robot based on the current signal analysis as claimed in claim 3, wherein the step 4) comprises the steps of inputting 1 number of channels into the first convolution layer, outputting 32 number of channels, convolution kernel size of 5 x 5, step size of 1, and filling 2 pixel lengths with 0 at the boundary;
the first pooling layer mode is maximum pooling, the size of a pooling core is 2 x 2, and the step length is 2;
the number of input channels of the second convolution layer is 32, the number of output channels of the second convolution layer is 32, the convolution kernel size is 5 x 5, the step length is 1, and the boundary is filled with 0 for 2 pixel lengths;
the second pooling layer pattern is maximal pooling, the pooled core size is 2 x 2, and the step size is 2;
the number of input channels of the third convolution layer is 32, the number of output channels is 64, the convolution kernel size is 5 x 5, the step size is 1, and the boundary is filled with 0 for 2 pixel lengths;
the third pooling layer mode is maximum pooling, the size of a pooling core is 2 x 2, and the step length is 2;
the expansion layer expands the multi-channel two-dimensional data output by the third pooling layer into a one-dimensional vector;
the number of output channels of the one-dimensional vector output by the expansion layer sequentially passes through the first layer of full-connection layer and the second layer of full-connection layer is 2.
5. The method for fault diagnosis of an industrial robot based on current signal analysis according to any of claims 1 to 4, characterized in that the two-dimensional data of step 3) is mapped to a value range of 0-255, and a visible n/2 x n/2 two-dimensional gray scale map is obtained as an input of the convolutional neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111596134.0A CN114282577A (en) | 2021-12-24 | 2021-12-24 | Industrial robot fault diagnosis method based on current signal analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111596134.0A CN114282577A (en) | 2021-12-24 | 2021-12-24 | Industrial robot fault diagnosis method based on current signal analysis |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114282577A true CN114282577A (en) | 2022-04-05 |
Family
ID=80874670
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111596134.0A Pending CN114282577A (en) | 2021-12-24 | 2021-12-24 | Industrial robot fault diagnosis method based on current signal analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114282577A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117383126A (en) * | 2023-11-29 | 2024-01-12 | 广州赛志系统科技有限公司 | Plate sorting buffer storage position scheduling method, control system and intelligent sorting production line |
| CN117972371A (en) * | 2024-01-29 | 2024-05-03 | 重庆大学 | Industrial robot running state long-sequence current signal identification method |
-
2021
- 2021-12-24 CN CN202111596134.0A patent/CN114282577A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117383126A (en) * | 2023-11-29 | 2024-01-12 | 广州赛志系统科技有限公司 | Plate sorting buffer storage position scheduling method, control system and intelligent sorting production line |
| CN117383126B (en) * | 2023-11-29 | 2024-04-09 | 广州赛志系统科技有限公司 | Plate sorting buffer storage position scheduling method, control system and intelligent sorting production line |
| CN117972371A (en) * | 2024-01-29 | 2024-05-03 | 重庆大学 | Industrial robot running state long-sequence current signal identification method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114282577A (en) | Industrial robot fault diagnosis method based on current signal analysis | |
| CN110595780B (en) | Bearing fault identification method based on vibration gray level image and convolution neural network | |
| CN109946080B (en) | Mechanical equipment health state identification method based on embedded circulation network | |
| CN108241873A (en) | A kind of intelligent failure diagnosis method towards pumping plant main equipment | |
| CN111122161B (en) | Variable working condition bearing fault diagnosis method based on fast kurtogram and deep residual learning | |
| CN109815855A (en) | A kind of electronic equipment automatic test approach and system based on machine learning | |
| CN114216682B (en) | Service life prediction method and device of rolling bearing based on TCN and BLS | |
| CN114549589B (en) | Rotator vibration displacement measurement method and system based on lightweight neural network | |
| CN111738521B (en) | Non-invasive power load monitoring sequence generation method, system, equipment and medium | |
| CN111879508A (en) | Method and device for estimating instantaneous rotating speed of rotating machine based on time-frequency transformation and storage medium | |
| CN113505830A (en) | Rotating machine fault diagnosis method, system, equipment and storage medium | |
| CN116049632A (en) | Wind power main shaft bearing fault diagnosis method, device and application | |
| CN102509076B (en) | Principal-component-analysis-based video image background detection method | |
| CN115328661B (en) | Computing power balance execution method and chip based on voice and image characteristics | |
| CN116541771A (en) | Unbalanced sample bearing fault diagnosis method based on multi-scale feature fusion | |
| CN116663739A (en) | Complex terrain fan output prediction method, system, equipment and storage medium | |
| CN113139590B (en) | Dimension reduction method and device for time series data, computer equipment and storage medium | |
| CN106772032B (en) | Fault feature extraction method for hydroelectric generating set | |
| CN115144211A (en) | Fault detection method and device based on discrete wavelet transform and gated cyclic unit | |
| CN110472741B (en) | Three-domain fuzzy wavelet width learning filtering system and method | |
| Wang et al. | An adaptive model for time-varying speed fault diagnosis under strong noise interference | |
| US11310349B1 (en) | Transforming multivariate time series data into image data to generate image-based predictions | |
| CN117892073B (en) | Irrigation area water metering system and water metering method | |
| KR102194303B1 (en) | System and Method for Preprocessing and Data Set Augmentation for training AIwith 3D Data Processing | |
| CN103829975B (en) | Method for calculating ultrasonic elastography displacement value |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |