A fault diagnosis method for ultrasonic motors based on deep learning

Abstract

The present invention discloses a method for ultrasonic motor fault diagnosis based on deep learning, the method comprising two steps of offline modeling and online application, wherein the offline modeling comprises the following steps: step 1, collecting ultrasonic motor performance monitoring data and environmental monitoring data under normal and various fault modes; step 2, combining and arranging performance monitoring data channels. In the present invention, no artificial fault feature extraction is required, and an end-to-end approach is adopted to obtain a deep learning fault diagnosis model, which can accurately estimate the normal / fault state of the ultrasonic motor, extract the correlation characteristics between the various monitoring parameters using a convolutional neural network, and extract the correlation characteristics between the monitoring parameters and time using a recurrent neural network, and have accurate ultrasonic motor fault diagnosis capabilities, with a fault diagnosis accuracy of 100% on existing verification data.

Description

Technical Field

[0001] The present invention relates to the technical field of life prediction methods for aerospace electromechanical products, and in particular to an ultrasonic motor fault diagnosis method based on deep learning. Background Art

[0002] Ultrasonic motors do not require coils and speed change mechanisms. They have the advantages of light weight, small size, low electromagnetic radiation, fast speed response, and high unit energy density. They are widely used in aerospace applications such as space smart drive devices and extraterrestrial planetary ground rover drive devices, and are highly valued by the aerospace community of various countries. With years of development and improvement, the service life and reliability of ultrasonic motors have been continuously improved, and more and more applications are being used in long-term working scenarios. Summary of the invention

[0003] The purpose of the present invention is to propose an ultrasonic motor fault diagnosis method based on deep learning in order to solve the problem of accurate ultrasonic motor fault diagnosis.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for ultrasonic motor fault diagnosis based on deep learning, the method comprising two steps: offline modeling and online application, wherein the offline modeling comprises the following steps:

[0006] Step 1: Collecting ultrasonic motor performance monitoring data and environmental monitoring data under normal and various fault modes;

[0007] Step 2: Combining and arranging the performance monitoring data channels;

[0008] Step 3: Establish a deep learning model based on convolutional networks, recurrent networks and fully connected networks;

[0009] Step 4: Conduct deep learning model training;

[0010] Step 5: Get the trained deep learning model as input for online applications;

[0011] The online application includes the following steps:

[0012] Step 6: Collect ultrasonic motor performance monitoring data and environmental monitoring data in real time;

[0013] Step 7: Combining and arranging the performance monitoring data channels;

[0014] Step 8: using the deep learning model obtained in the offline stage to estimate the state of the ultrasonic motor;

[0015] Step 9: If the equipment is fault-free, continue monitoring. If the equipment is faulty, alarm and output the fault type.

[0016] Preferably, the ultrasonic motor failure modes in step one include four types: cracking of piezoelectric ceramic plates, wear of friction plates, loosening of rubber layers, and broken teeth of elastomers. The sampling samples should cover different environmental conditions.

[0017] Preferably, the ultrasonic motor performance monitoring data and environmental monitoring data in step 1 and step 6 have a sampling frequency of 500 kHz and a single sampling time of 5 s.

[0018] Preferably, the ultrasonic motor performance monitoring data in step 1 and step 6 are driving voltage, driving current, isolated pole feedback voltage, driving frequency, internal temperature, and ultrasonic motor speed.

[0019] Preferably, the ultrasonic motor environment monitoring data in step 1 and step 6 are: ambient temperature and vibration root mean square value.

[0020] Preferably, the performance monitoring data channel combination arrangement method in step 2 and step 7 is: each performance monitoring data is combined in pairs as a channel, and the channels are stacked.

[0021] Preferably, in the training process in step 4, multivariate classification cross entropy is used as the loss function, and RMSProp is used as the optimization algorithm.

[0022] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The present invention does not require manual fault feature extraction, and adopts an end-to-end approach to obtain a deep learning fault diagnosis model, which can accurately estimate the normal / fault state of the ultrasonic motor.

[0025] 2. The present invention utilizes convolutional neural networks to extract the correlation characteristics between various monitoring parameters, and adopts a combination of two types of performance parameter channels to overcome the problem of classical convolutional neural networks focusing on local correlation.

[0026] 3. The present invention utilizes a recurrent neural network to extract the correlation characteristics between monitoring parameters and time.

[0027] 4. The present invention has the ability to accurately diagnose ultrasonic motor faults, and the fault diagnosis accuracy rate is 100% based on existing verification data. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart of a method provided according to an embodiment of the present invention is shown;

[0029] Figure 2It shows the establishment of a deep learning model based on a convolutional network, a recurrent network and a fully connected network according to an embodiment of the present invention;

[0030] Figure 3 It shows the temperature test of the ultrasonic motor provided according to the embodiment of the present invention;

[0031] Figure 4 It shows the change in accuracy of the deep learning model training process provided by an embodiment of the present invention;

[0032] Figure 5 The ultrasonic motor fault diagnosis result provided by the embodiment of the present invention is shown. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] See also Figure 1-5 , the present invention provides a technical solution:

[0035] A method for ultrasonic motor fault diagnosis based on deep learning, the fault diagnosis method includes two steps: offline modeling and online application, wherein the offline modeling includes the following steps:

[0036] Step 1: Collect the performance monitoring data and environmental monitoring data of the ultrasonic motor under normal and various fault modes; the ultrasonic motor failure modes include cracking of piezoelectric ceramics, wear of friction plates, loosening of rubber layers, and broken teeth of elastomers. The sampling samples should cover different environmental conditions;

[0037] Step 2: Combining and arranging the performance monitoring data channels; combining each performance monitoring data in pairs as a channel;

[0038] Step 3: Establish a deep learning model based on convolutional networks, recurrent networks and fully connected networks;

[0039] The deep learning model is shown in Table 2 below:

[0040] Table 2 Deep learning models

[0041]

[0042] Among them, the output is a one-hot encoding method, and the state with the highest probability is taken as the output. The expression is as follows:

[0043]

[0044] Among them, y1 corresponds to the normal state, y2 corresponds to the cracking of the piezoelectric ceramic plate, y3 corresponds to the wear of the friction plate, y4 corresponds to the loosening of the rubber layer, and y5 corresponds to the broken teeth of the elastic body.

[0045] Step 4: Conduct deep learning model training:

[0046] The multivariate classification cross entropy is used as the loss function during training:

[0047]

[0048] where y represents the value of the loss function; i is the true category label; p i Category prediction probability; i is the current category; n type is the number of categories, which is 5 in the present invention;

[0049] RMSProp is used as the optimization algorithm. The detailed steps are as follows:

[0050]

[0051] Wherein, ω is the neural network parameter to be optimized, and the initial value is sampled from the standard normal distribution; α is the learning rate, which is 0.001 in the present invention; S dω is a value container for carrying the weighted average of the squared gradient; β is an attenuation factor, which is 0.9 in the present invention;

[0052] Step 5: Get the trained deep learning model as input for online applications;

[0053] The online application consists of the following steps:

[0054] Step 6: Real-time collection of ultrasonic motor performance monitoring data and environmental monitoring data, the data collection method is the same as step 1;

[0055] Step 7: Combining and arranging the performance monitoring data channels;

[0056] Step 8: using the deep learning model obtained in the offline stage to estimate the state of the ultrasonic motor;

[0057] Step 9: If the equipment is fault-free, continue monitoring; if the equipment is faulty, alarm and output the fault type;

[0058] In order to verify whether the proposed ultrasonic motor fault diagnosis method based on deep learning can effectively realize the state discrimination of ultrasonic motor, ultrasonic motor fault injection test data is carried out for verification.

[0059] The ultrasonic motor test system consists of an ultrasonic motor, a coupling, a magnetic torquer, a controller, a data acquisition card, and a host computer. According to the analysis of the ultrasonic motor fault evolution mechanism, four fault modes are determined: piezoelectric ceramic cracking, friction plate wear, rubber layer loosening, and elastic body tooth breakage. The faulty parts are processed. Among them, the piezoelectric ceramic cracking fault is injected by breaking, the ceramic plate wear is injected by load torque acceleration, the rubber layer loosening is injected by less glue when processing the test piece, and the elastic body tooth breakage is injected by tooth surface grinding.

[0060] The ultrasonic motor does not work during the satellite launch phase, but needs to withstand impact loads and random vibration loads. During the satellite on-orbit phase, the ultrasonic motor works intermittently according to the orbit change requirements. At the same time, since the satellite needs to frequently enter and exit the earth's shadow, the ultrasonic motor needs to withstand the influence of temperature alternation loads. In order to truly simulate the working state of the ultrasonic motor when it is in orbit, the present invention simulates the temperature and micro-vibration environment during the on-orbit phase, wherein the temperature cycle includes high temperature during the illumination period and low temperature during the earth's shadow period, the micro-vibration environment remains unchanged, and the ultrasonic motor works normally during the on-orbit phase. The test includes 2 normal motors and 8 abnormal motors, wherein the abnormal motor includes 4 fault modes, which are obtained by fault injection. During the test, the functional performance signals and environmental signals of the ultrasonic motor are collected in real time. The functional performance signals include driving voltage, driving current, driving frequency, lone pole feedback voltage, ultrasonic motor speed, ultrasonic Motor temperature, environmental signals include ambient temperature and vibration signals. In order to obtain more test samples, the motor is frequently turned on and off under different profiles. After the motor speed stabilizes, a data acquisition card is used to collect data every 10 seconds. The sampling time is 5 seconds and the sampling frequency is 500kHz. For motors #1, #3, #5, #7, and #9, the motors are turned on / off 5 times in each environmental combination, and 10 samples are collected each time the motors are turned on. For motors #2, #4, #6, #8, and #10, the motors are turned on / off 5 times in each environmental combination, and 2 samples are collected each time. Among them, 500 groups of test data of motors #1, #3, #5, #7, and #9 are used for model training, and 40 groups of test data of motors #2, #4, #6, #8, and #10 are used for model testing. The samples and numbers obtained are shown in Table 3 below:

[0061] Table 3 Test data classification

[0062]

[0063]

[0064] In order to verify the model training effect and prevent overfitting, the 500 training models are divided into two categories. The first category contains 450 groups of samples for training. The second category contains 50 groups of samples for verifying the training effect. The samples are classified randomly. After the classification is completed, manual verification is performed to ensure that both the training set and the verification set contain test data of all motor states under different environmental conditions. The model accuracy changes during the training process are shown in the attached figure. Figure 4 As shown in the figure, it can be seen that the accuracy trends of the training set and the validation set models are synchronized, indicating that there is no overfitting phenomenon in the model. Finally, the accuracy of the training set and the validation set models are both 100%, which shows that there are indeed significant differences between the monitoring quantities of the working parameters of each fault mode, and the regularity in the data is very strong. In order to further verify the fault diagnosis ability of the model, the model is used to distinguish the test data obtained from other ultrasonic motor tests. The discrimination results are shown in the attached figure. Figure 5 As shown, it can be seen that all samples can be successfully predicted and the fault diagnosis accuracy is 100%, proving that the model has good versatility.

[0065] Specifically, Figure 1 As shown in the figure, the ultrasonic motor failure modes in step one include cracking of piezoelectric ceramics, wear of friction plates, loosening of rubber layers, and broken teeth of elastomers. The sampling samples should cover different environmental conditions. The sampling frequency is 500kHz, and the single sampling time is 5s. The performance monitoring data are: driving voltage, driving current, isolated pole feedback voltage, driving frequency, internal temperature, and ultrasonic motor speed. The environmental monitoring data are: ambient temperature and vibration root mean square value.

[0066] Specifically, Figure 1 As shown, the ultrasonic motor performance monitoring data and environmental monitoring data in step 1 and step 6 have a sampling frequency of 500kHz and a single sampling time of 5s.

[0067] Specifically, Figure 1 As shown, the ultrasonic motor performance monitoring data in step 1 and step 6 are driving voltage, driving current, isolated pole feedback voltage, driving frequency, internal temperature, and ultrasonic motor speed.

[0068] Specifically, as shown in the figure, the ultrasonic motor environment monitoring data in step 1 and step 6 are: ambient temperature and vibration root mean square value.

[0069] Specifically, Figure 1 As shown, the performance monitoring data channel combination arrangement method in step 2 and step 7 is: each performance monitoring data is combined into a channel, and the channels are stacked. The combination arrangement method of step 7 is the same as that of step 2. The data channel combination arrangement method is shown in Table 1 below:

[0070] Table 1 Performance monitoring parameter arrangement

[0071] Channel number Test parameter 1 Test parameter 2 Channel #1 Driving voltage Drive current Channel #2 Driving voltage Isolated voltage Channel #3 Driving voltage Drive frequency Channel #4 Driving voltage Internal temperature rise Channel #5 Driving voltage Motor speed Channel #6 Drive current Isolated voltage Channel #7 Drive current Drive frequency Channel #8 Drive current Internal temperature Channel #9 Drive current Motor speed Channel #10 Isolated voltage Drive frequency Channel #11 Isolated voltage Internal temperature Channel #12 Isolated voltage Motor speed Channel #13 Drive frequency Internal temperature Channel #14 Drive frequency Motor speed Channel #15 Internal temperature Motor speed

[0072] Stack different channels and generate an input dimension of n sample ×n performance ×n channel , where n sample is the number of sampling points for a single sampling of 5s, here it is 2500000; n performance is the number of performance data per channel, here it is 2; n channel is the number of combinations of the six channels in pairs, here it is 15.

[0073] Specifically, as shown in the figure, in the training process in step 4, multivariate classification cross entropy is used as the loss function, and RMSProp is used as the optimization algorithm.

[0074] The above description of the embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for ultrasonic motor fault diagnosis based on deep learning, characterized in that: The fault diagnosis method comprises two steps: offline modeling and online application, wherein the offline modeling comprises the following steps: Step 1: Collecting ultrasonic motor performance monitoring data and environmental monitoring data under normal and various fault modes; Step 2: Combining and arranging the performance monitoring data channels; Step 3: Establish a deep learning model based on convolutional networks, recurrent networks and fully connected networks; Step 4: Conduct deep learning model training; Step 5: Get the trained deep learning model as input for online applications; The online application includes the following steps: Step 6: Collect ultrasonic motor performance monitoring data and environmental monitoring data in real time; Step 7: Combining and arranging the performance monitoring data channels; Step 8: using the deep learning model obtained in the offline stage to estimate the state of the ultrasonic motor; Step 9: If the equipment is fault-free, continue monitoring; if the equipment is faulty, alarm and output the fault type; The ultrasonic motor failure modes in step 1 include cracking of piezoelectric ceramics, wear of friction plates, loosening of rubber layers, and broken teeth of elastomers. The sampling samples should cover different environmental conditions. The ultrasonic motor performance monitoring data and environmental monitoring data in step 1 and step 6 have a sampling frequency of 500kHz and a single sampling time of 5s; The ultrasonic motor performance monitoring data in step 1 and step 6 are driving voltage, driving current, isolated pole feedback voltage, driving frequency, internal temperature, and ultrasonic motor speed; The performance monitoring data channel combination arrangement method in step 2 and step 7 is: each performance monitoring data is combined into a channel in pairs, and the channels are stacked.

2. The ultrasonic motor fault diagnosis method based on deep learning according to claim 1 is characterized in that: The ultrasonic motor environmental monitoring data in step 1 and step 6 are: ambient temperature and vibration root mean square value.

3. The ultrasonic motor fault diagnosis method based on deep learning according to claim 1 is characterized in that: In the training process in step 4, multivariate classification cross entropy is used as the loss function, and RMSProp is used as the optimization algorithm.

Images