Inter-turn Short Circuit Fault Diagnosis Method of Electro-Mechanical Actuator Based on Ensemble Learning Algorithm

Abstract

The present invention provides a method for diagnosing inter-turn short circuit faults of an electro-mechanical actuator based on an ensemble learning algorithm. A vector control model and a lumped parameter fault model of the electro-mechanical actuator are established through Simulink software to determine fault features for diagnosis. An ensemble learning fault diagnosis framework based on the Stacking model fusion strategy is established, and the model built by the ensemble learning fault diagnosis framework is optimized using K-fold cross-validation. An electro-mechanical actuator test bench is built and the collected output data is processed to extract fault features, and the fault diagnosis is realized using the ensemble learning fault diagnosis framework. The present invention realizes accurate modeling of inter-turn short circuit faults; the mechanical part is considered during the overall modeling of the electro-mechanical actuator, so as to accurately extract fault features. The method of K-fold cross-validation is adopted to improve the prediction accuracy of the model while accelerating the training speed of the model.

Description

Technical Field

[0001] The present invention relates to the field of electromechanics, and in particular to a method for fault diagnosis of an actuator, specifically a method for diagnosing inter-turn short circuit faults of an electromechanical actuator. Background Art

[0002] With the development of the aviation industry, people's requirements for the comprehensive performance of aircraft are getting higher and higher. As a result, more-electric aircraft have emerged on the historical stage. Their significant advantage is that the traditional hydraulic / pneumatic system in aircraft is replaced by an electric power system, which can significantly optimize the aircraft structure, effectively improve controllability and reduce maintenance costs while achieving this goal. To achieve this goal, electromechanical actuators have emerged as the times require. An electromechanical actuator is an electric-driven actuating device that controls a drive motor through an electrical signal of a flight control system and drives a control surface through mechanical transmission devices such as a gear reduction box, a transmission shaft, and a ball screw. The electromechanical actuator driven by a power converter converts electrical energy into mechanical energy to drive the control surface, enabling the power transfer from the second energy system of the aircraft to each actuator of the actuating system to be transmitted in the form of electrical energy through wires, thus replacing the hydraulic pipelines currently spread throughout the aircraft fuselage, reducing weight, and greatly saving the energy consumption of the aircraft. Therefore, its working reliability directly determines the flight quality of the aircraft. Usually, the working conditions of electromechanical actuators are relatively harsh, inevitably leading to various types of faults. Therefore, it is particularly crucial to diagnose faults of airborne electromechanical actuators.

[0003] Currently, the research on inter-turn short circuit faults of electromechanical actuators mainly includes finite element fault modeling, lumped parameter fault modeling, and experimental vibration signal extraction. The finite element method can describe motor faults more accurately, but it has problems such as large computational amount, long simulation time, and difficulty in implementing servo control; the lumped parameter modeling method describes motor faults through physical equations and mathematical equations. Although this method has lower accuracy than the finite element method, it has a small computational amount and is easy to implement servo control, which can meet the requirements of fault diagnosis; the extraction of experimental vibration signals is to extract vibration signals using vibration sensors and extract fault features through signal processing, but this method is not very practical for motors in special working states. At the same time, the current diagnostic methods for inter-turn short circuit fault signals have problems such as insufficient diagnostic accuracy of traditional machine learning methods and slow training speed of deep learning methods. Therefore, it is necessary to develop a method with both high diagnostic accuracy and not taking too long. Summary of the Invention

[0004] To overcome the deficiencies of the prior art, the present invention provides a method for diagnosing inter-turn short circuit faults of an electromechanical actuator based on an ensemble learning algorithm. In order to accurately detect and diagnose faults in the electromechanical actuator, the present invention discloses a method for diagnosing inter-turn short circuit faults of an electromechanical actuator based on an ensemble learning algorithm. As a means for fault detection and diagnosis of an electromechanical actuator, compared with the problems of insufficient accuracy of traditional machine learning methods and slow training speed of deep learning methods, this method adopts a model fusion strategy based on Stacking, absorbs the respective advantages of XGBoost, LightGBM, and CatBoost, and optimizes the model fusion process on the basis of the traditional Stacking strategy. The method of K-fold cross-validation is used to improve the prediction accuracy of the model while accelerating the training speed of the model.

[0005] The technical solutions adopted by the present invention to solve its technical problems include the following steps:

[0006] Step 1: Analyze the composition principle and possible fault types of the electromechanical actuator, and speculate that the possible fault of the electromechanical actuator is an inter-turn short circuit fault of a permanent magnet synchronous motor;

[0007] The inter-turn short circuit fault refers to the phenomenon that the insulation skin of the internal winding of the motor is damaged due to factors such as moisture, high temperature, and overcurrent, and the current is short-circuited between the windings. The external manifestation is that the three-phase current of the motor is unbalanced, and the motor speed and electromagnetic torque fluctuate greatly;

[0008] Step 2: Establish a vector control model and a lumped parameter fault model of the electromechanical actuator through Simulink software, and determine the fault features for diagnosis;

[0009] Step 3: Establish an ensemble learning fault diagnosis framework based on the Stacking model fusion strategy, and optimize the model built by the ensemble learning fault diagnosis framework using K-fold cross-validation;

[0010] Step 4: Build an electromechanical actuator test bench and process the collected output data, extract fault features, and use the ensemble learning fault diagnosis framework to achieve fault diagnosis.

[0011] In step 2, a lumped parameter fault model of the electromechanical actuator is established through Simulink software, and the fault features for diagnosis are obtained by simulation. The fault features are the three-phase current, or the harmonic signal of the current, or the vibration signal of the motor;

[0012] Step 2.1: Calculate the inductance matrix of the faulty motor through Simulink software. Set the motor fault to occur in one of the three phases, phase A. Calculate the self-inductance and mutual inductance between phases under the fault, and form an inductance matrix. The calculation method is as follows:

[0013] Step 2.1.1: Given that the self-inductance of a normal motor phase is L and the mutual inductance is M. When a fault occurs, phase A is divided into a normal winding a and a short-circuited winding f. The calculation relationship between the inductances of the normal winding and the short-circuited winding is as follows:

[0014] L′ aa +2M af +L ff =L aa

[0015] In the formula, L aa represents the self-inductance of phase A, L’ aa represents the self-inductance of the normal winding a, M af represents the mutual inductance between the normal winding a and the short-circuited winding f, and L ff represents the self-inductance of the short-circuited winding f;

[0016] Step 2.1.2: During the actual operation of the motor, there will be inductance leakage. Therefore, the calculation relationship for the supplementary inductance is as follows:

[0017]

[0018] Step 2.1.3: The relationship between the inductances of the normal winding and the short-circuited winding also depends on the number of turns of the short-circuited coil. The calculation formula is as follows:

[0019]

[0020] In the formula, n a is the number of turns of the normal winding a, n b is the number of turns of the short-circuited winding f. From Steps 2.1.1 to 2.1.3, the three inductance parameters L’ aa 、M af 、L ff of the short-circuited and un-short-circuited windings of phase A can be calculated;

[0021] Step 2.1.4: The mutual inductances M’ ab between the normal winding a and windings B and C, and the mutual inductances M bf between the short-circuited winding f and windings B and C satisfy the calculation formula:

[0022]

[0023] M bf =σM

[0024]

[0025] In the formula, σ is the short-circuit ratio of phase A winding;

[0026] Step 2.2: Based on the parameters of the motor itself and the inductance parameters calculated in Step 2.1, establish a fault motor model:

[0027] Step 2.2.1: Calculate the current values of each phase from the input voltage, self-resistance, inductance, and flux linkage parameters of the motor. The calculation formula is as follows:

[0028]

[0029]

[0030] In the formula, R is the phase resistance of the motor, U a 、U b 、U c are the input three-phase voltages, i a 、i b 、i c 、i f are the output currents, ψ ma 、ψ mb 、ψ mc 、ψ mf are the parameters related to the flux linkage and electrical angle;

[0031] Step 2.2.2: Perform coordinate transformation on the current values of each phase, and calculate the electromagnetic torque T e output by the motor from the electromagnetic torque equation of the motor in the d-q coordinate system. The calculation formula is

[0032]

[0033] In the formula, p n is the number of pole pairs of the motor, i q is the q-axis current calculated after coordinate transformation of each phase current, and ψ f is the permanent magnet flux linkage;

[0034] Step 2.3: Establish a cylindrical gear reducer model. According to the number of teeth Z1 of the front gear, the number of teeth Z2 of the rear gear, and the transmission efficiency η1, calculate the transmission ratio At the same time, obtain the relationship between the input torque and the output torque T2 = i * T1, and the relationship between the input speed and the output speed T1 is the input torque of the reducer, T2 is the output torque of the reducer, ω1 is the input speed of the reducer, and ω2 is the output speed of the reducer;

[0035] Step 2.4: Establish a planetary roller screw drive model. According to the number of screw threads n s of the screw, the pitch p, the lead s, the output torque T2 of the reducer, and the transmission efficiency η2, calculate the output force of the screw as

[0036] In Step 3, establish an integrated learning fault diagnosis framework based on the Stacking model fusion strategy, and optimize the model using K-fold cross-validation:

[0037] Step 3.1: Preprocess the collected data to generate an available dataset with labels. Then, perform feature extraction or feature dimensionality reduction as needed. Finally, divide the available dataset into a training set and a test set according to a certain ratio, where the ratio of the training set to the test set is 7:3 or 8:2.

[0038] Step 3.2: Create three primary learners, namely XGBoost, LightGBM, and CatBoost respectively. Adopt the K-fold cross-validation method, randomly shuffle and evenly divide the training set into K parts, namely train1, train2, train3, train4, …, train K , and then conduct K rounds of iterative training. When performing the i-th round of training, use the training set data other than train i to train the three primary learners, and use the three trained primary learners to predict train i respectively to obtain the prediction results Y i . After K rounds of iteration, combine the Y1, Y2, Y3, Y4, …, Y i , …, Y K obtained in each round of iteration as the features of the new training set. The test set undergoes the same K-fold cross-validation as the training set to obtain the features of the new test set.

[0039] Step 3.3: Create an SVM main learner, train the SVM learner based on the new training set in Step 3.2, and use the trained SVM model to predict the new test set. Take the prediction results as the final diagnosis result of the framework.

[0040] The preprocessing in Step 3.1 adopts one-hot encoding preprocessing.

[0041] The feature extraction or feature dimensionality reduction in Step 3.1 is as follows: Feature extraction includes performing Fourier transform and wavelet packet transform signal processing on the data to obtain frequency-domain data; Feature dimensionality reduction is to divide the samples. Each sample is a matrix of a×b. Take the mean and variance of every c rows. a is an integer multiple of c, and the data will be dimensionally reduced. In this way, both the features of the data are extracted and the amount of data is reduced, accelerating the training speed. Feature extraction and feature dimensionality reduction are often carried out simultaneously.

[0042] In Step 4, build an electromechanical actuator test bench and collect the output fault features, and use the integrated learning fault diagnosis framework to achieve fault diagnosis:

[0043] Step 4.1: Build an experimental platform for collecting fault data. This experimental platform consists of an electromechanical actuator (EMA), a control unit, a loading device, a signal acquisition device, and a fault simulation device.

[0044] Step 4.2: Based on the LabVIEW software, establish a data acquisition module, a motion control module, a data display module, and a data storage module for measuring EMA;

[0045] Step 4.3: Through the experimental environment built in Steps 4.1 and 4.2, collect the three-phase current data of EMA under different degrees of turn-to-turn short circuit faults and different displacement commands under a constant load. To increase the amount of data, conduct multiple experiments for different working conditions;

[0046] Step 4.4: Process the current data collected in Step 4.3 and use the integrated learning model established in Step 3 for diagnosis:

[0047] Step 4.4.1: Since the data collected in the experiment is in tdms file format, convert the tdms file to an xlsx file through excel, read the xlsx file using the pandas library, and then convert the data of the read xlsx file into DataFrame type data that can be recognized by the machine learning model;

[0048] Step 4.4.2: Create labels for the DataFrame current data transformed in Step 4.4.1 through one-hot encoding. The steps to create labels are as follows: Encode N labels through an N-bit register. When creating the nth label, set the nth bit of the register to 1 and the other bits to 0;

[0049] Step 4.4.3: Divide the data encoded in Step 4.4.2 into samples. Since the current changes periodically during the uniform motion stage of EMA, take H rows of data in the encoded data as a sample, and calculate the mean, average absolute variance, kurtosis, and skewness of each of the H rows of data once to obtain 4 rows of statistical data. Then, merge the 4 rows of statistical data of all samples row by row to generate a matrix P, and then flatten the matrix P row by row to achieve feature dimensionality reduction; Divide the data after dimensionality reduction into a training set and a test set in a ratio of 7:3 for the next training of the learner;

[0050] Step 4.4.4: Adopt the method of K-fold cross-validation. First, divide the training set evenly into K parts and conduct K rounds of iteration. When conducting the i-th round of iteration, where 1 ≤ i ≤ K, use the other K - 1 parts except the i-th part as the sub-training set to train three primary learners, and use the trained primary learners to predict the i-th part of the data that did not participate in the training in the current round of iteration to obtain the probability matrix Y of each sample in this part of the data belonging to each health state i ; After K rounds of iteration, the probability matrices Y1, Y2, …, Y i 、…、Y KMerge by rows as the new training set; when generating features for the test set, use the three trained primary learners in each iteration to predict the entire test set. After K iterations, K sets of feature data are obtained. Perform mean processing on the K data sets, and use the result as the final new test set features; train the main learner SVM called with the new training set, and then use the main learner SVM to test the new test set. The test result is the final diagnosis result;

[0051] Step 4.4.5: The classification effect of the main learner SVM in Step 4.4.4 highly depends on the parameters gamma and the penalty term C, and hyperparameter tuning is required. Hyperparameter tuning uses grid search, that is, first set the value ranges of the hyperparameters gamma and C, and then evenly divide the new training set in Step 4.4.4 into K parts. In each iteration, use K - 1 of them for hyperparameter tuning, and the other 1 part except K - 1 is used to verify the effect of the hyperparameters. After K iterations, obtain the hyperparameter combination with the maximum accuracy; use this hyperparameter combination to create an SVM model and perform model training based on the new training set, and finally pass the prediction accuracy of the test set samples.

[0052] The beneficial effects of the present invention are as follows: Based on Simulink, the inductance calculation and overall modeling of the motor part of the electro-mechanical actuator under inter-turn short-circuit faults are carried out. When calculating the inductance matrix, the changes in the self-inductance and mutual-inductance parameters of each phase when the motor has an inter-turn short-circuit fault are fully considered, realizing the accurate modeling of the inter-turn short-circuit fault; when performing the overall modeling of the electro-mechanical actuator, the mechanical part is considered, including the modeling of the reduction gearbox and the planetary roller screw, so as to accurately extract the fault features. The fault diagnosis method based on the ensemble learning algorithm disclosed in the present invention, as a means for the fault detection and diagnosis of electro-mechanical actuators, compared with the problems of insufficient accuracy of traditional machine learning methods and slow training speed of deep learning methods, this method adopts a model fusion strategy based on Stacking, absorbs the respective advantages of XGBoost, LightGBM, and CatBoost, and optimizes the model fusion process on the basis of the traditional Stacking strategy. The method of K-fold cross-validation is used to improve the model prediction accuracy while accelerating the model training speed. Brief Description of the Drawings

[0053] Figure 1 It is a structural diagram of an electro-mechanical actuator.

[0054] Figure 2 It is a diagram of the inter-turn short-circuit fault of the motor winding.

[0055] Figure 3 It is a flow chart of the inter-turn short-circuit fault diagnosis of the electro-mechanical actuator.

[0056] Figure 4 It is a flow chart of the ensemble learning fault diagnosis.

[0057] Figure 5 It is a flowchart of five-fold cross-validation.

[0058] Figure 6 It is a flowchart for data feature extraction.

[0059] Figure 7 It is a comparison chart of the effects of the ensemble learning algorithm and other algorithms. Detailed implementation manners

[0060] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0061] Among the existing fault modes of the electromechanical actuator, the inter-turn short circuit fault is an early fault with a high occurrence frequency. It is not easy to be detected and diagnosed, and it is extremely easy to evolve into a serious fault, and it will also cause the occurrence of other fault types. Therefore, it is extremely important to accurately diagnose it. The present invention proposes an accurate inductance calculation method and a modeling method for the electromechanical actuator according to the inter-turn short circuit fault mechanism, which can fully consider the changes in the self-inductance and mutual inductance of each phase winding when the inter-turn short circuit fault occurs, and effectively inject faults and extract features for the inter-turn short circuit fault of the electromechanical actuator, so as to achieve rapid fault detection and diagnosis.

[0062] The electromechanical actuator is a complex non-linear system. Based on the actual working process of the disclosed inter-turn short circuit fault injection method and fault detection and diagnosis method and in conjunction with the accompanying drawings, Figure 1 the specific implementation process of the inter-turn short circuit fault injection method and fault detection and diagnosis method for the electromechanical actuator shown is described. The fault injection principle of the present invention during actual use is as Figure 2 shown, and the entire fault diagnosis process is as Figure 3 shown. The specific steps are as follows:

[0063] Step 1: Analyze the composition principle of the electromechanical actuator and the possible fault types. Considering that the inter-turn short circuit fault of the electromechanical actuator is not easy to detect in the early stage, and long-term fault operation will cause the fault to deepen and lead to the occurrence of other faults, the fault mode of the electromechanical actuator is determined to be the inter-turn short circuit fault; the inter-turn short circuit fault refers to the phenomenon that the insulation skin of the internal winding of the motor is damaged due to factors such as high temperature and overcurrent, and the current is short-circuited between the windings. The external manifestation is that the three-phase current of the motor is unbalanced, and the motor speed and electromagnetic torque fluctuate greatly.

[0064] Step 2: Establish a vector control model and a lumped parameter fault model of the electromechanical actuator through Simulink software, and determine the fault features for diagnosis as the three-phase current. The specific steps are as follows:

[0065] Step 2.1: Establish a lumped parameter fault model for the motor part of the electro-mechanical actuator. The core is to calculate the motor inductance matrix that changes due to the fault. Set the motor fault in phase a, the short-circuit ratio is σ, and calculate the self-inductance of each phase and the mutual inductance between phases under the fault according to the motor self-inductance L and mutual inductance M, and form the inductance matrix. The calculation method is as follows:

[0066] Step 2.1.1: Calculate the self-inductance and mutual inductance of phases a and f with the largest parameter changes due to the fault. Let Then the self-inductance of phase a is The self-inductance of phase f is The mutual inductance between phase a and phase f is

[0067] Step 2.1.2: Calculate the mutual inductance that causes the change in the normal phase winding due to the fault. The mutual inductance between phase a and phases b and c is calculated as M(1 - σ), and the mutual inductance between phase f and phases b and c is calculated as Mσ;

[0068] Step 2.1.3: Calculate the self-inductance of the normal phase winding, which is the same as the winding self-inductance in the normal state of the motor. The self-inductance of phases b and c is still L;

[0069] Step 2.2: Establish a fault motor model according to the phase resistance R of the motor and the inductance matrix calculated in Step 2.1. The specific steps are as follows:

[0070] Step 2.2.1: Calculate the three-phase current values according to the input three-phase voltage of the motor and its own resistance, inductance, and flux linkage parameters. The calculation formula is

[0071]

[0072] In the formula, e abcf is a parameter related to the motor flux linkage and electrical angle;

[0073] Step 2.2.2: Perform a coordinate transformation on the three-phase current values of the motor, and calculate the electromagnetic torque output by the motor from the electromagnetic torque equation of the motor in the d-q coordinate system. The calculation formula is:

[0074]

[0075] In the formula, p n is the number of pole pairs of the motor, i q is the q-axis current calculated after the coordinate transformation of each phase current, ψ f is the permanent magnet flux linkage;

[0076] Step 2.3: Establish a typical cylindrical gear reducer model. According to the number of teeth Z1 and Z2 of the front and rear gears and the transmission efficiency η, calculate the transmission ratio At the same time, obtain the relationship between the input torque and output torque, input speed and output speed T2 = i * T1,

[0077] Step 2.4: Establish a planetary roller screw drive model. According to the number of screw thread heads n s , pitch p, lead s, input torque T, and transmission efficiency η of the screw, calculate the output force of the screw as

[0078] Step 3: Establish an integrated learning fault diagnosis framework based on the Stacking model fusion strategy. The complete integrated learning fault diagnosis process for the electro-mechanical actuator is as Figure 4 shown, and the K-fold cross-validation is used to optimize the model. The specific steps are as follows:

[0079] Step 3.1: Preprocess the collected data to generate an available dataset with labels. Then, perform feature extraction or feature dimensionality reduction as needed. Finally, divide the dataset into a training set and a test set according to a certain ratio;

[0080] Step 3.2: Adopt the method of K-fold cross-validation. Randomly shuffle and evenly divide the training set into 5 parts, namely train1, train2, train3, train4, and train5. Then, perform 5 rounds of iterative training. When performing the i-th round of training, use the training set data other than train i to learn and train each primary learner, and use each trained primary learner to predict train i respectively, to obtain the prediction result Y i ; After 5 rounds of iteration, combine the Y1, Y2, Y3, Y4, and Y5 obtained in each round of iteration as the features of the new training set. The optimized process block diagram for the five-fold cross-validation in the diagnosis process based on the Stacking fusion strategy is as Figure 5 shown;

[0081] Step 3.3: Create three primary learners, namely XGBoost, LightGBM, and CatBoost, respectively. Train them based on the training set, and use the trained models to predict the training set and the test set respectively, and take the results as the new training set and the new test set;

[0082] Step 3.4: Create an SVM master learner, train it based on the new training set, and use the trained SVM model to predict the new test set, and take the result as the final diagnosis result of the framework.

[0083] Step 4: Build an electro-mechanical actuator test bench and collect the output fault features. Use the integrated learning fault diagnosis framework to realize fault diagnosis. The feature processing method for the experimental data is as Figure 6 shown. The specific steps are as follows:

[0084] Step 4.1: Build an experimental platform for collecting fault data, which mainly consists of an EMA, a control unit, a loading device, a signal acquisition device, and a fault simulation device;

[0085] Step 4.2: Based on the LabVIEW software, establish a data acquisition module, a motion control module, a data display module, and a data storage module for measuring the EMA;

[0086] Step 4.3: Through the experimental software and hardware platform, collect the three-phase current data of the EMA under different degrees of turn-to-turn short circuit faults at a constant load and different displacement commands. To increase the amount of data, conduct multiple experiments under different working conditions;

[0087] Step 4.4: Process the collected current data and use the integrated learning model established in Step 3 for diagnosis:

[0088] Step 4.4.1: Since the data collected in the experiment is in tdms format, convert it to an xlsx file through excel and read it using the pandas library, and then convert it into a DataFrame type data that can be recognized by the machine learning model;

[0089] Step 4.4.2: Create labels for the converted current data through one-hot encoding. The method is to encode N labels through an N-bit register. When creating the nth label, just set the nth bit of the register to 1 and the other bits to 0;

[0090] Step 4.4.3: Divide the original data into samples. Since the current changes periodically during the uniform operation stage of the EMA, every 200 rows of the original data are used as a sample. In this way, each health state contains 500 samples, and the dimension of each sample is 200×3. For each sample, calculate the mean, average absolute variance, kurtosis, and skewness every 20 rows to generate a 40×3 matrix, and then flatten the 40×3 matrix to transform it into a 1×120 matrix, reducing the dimension of the sample from 200×3 to 1×120;

[0091] Step 4.4.4: Using the method of 5-fold cross-validation, first divide the training set into 5 equal parts. In each round of iteration, use 4 of them as the sub-training set to train three primary learners, and use the trained primary learners to predict the data that did not participate in the training in the current round of iteration, obtaining the probability of each sample belonging to each health state in this part of the data. Since there are a total of 4 health states, each primary learner will generate a 1×4 class probability matrix for a single sample, and the 3 primary learners together will generate a 1×12 class probability matrix as the new feature of this sample. Through 5 iterations, the feature generation of the entire training set can be completed, and the dimension of the new training set sample feature is 1×12. When generating features for the test set, use the three primary learners trained in each round of iteration to predict the entire test set, obtaining a new test set feature dimension of 1×12. After 5 rounds of iteration, 5 pieces of feature data are obtained. Perform mean processing on these 5 data sets and use the result as the final new test set feature;

[0092] Step 4.4.5: Step 4.4.4 uses a method that combines 5-fold cross-validation and grid search. First, set the value ranges of the hyperparameters gamma and C to [0.001, 0.01, 0.1, 1, 10, 100]. Then divide the new training set into 5 equal parts. In each round of iteration, use 4 of them for hyperparameter optimization, and the other 1 part is used to verify the effect of the hyperparameters. After 5 iterations, the best-performing hyperparameter combination is gamma = 0.001 and C = 100. Create an SVM model using the above hyperparameters and train the model based on the new training set to finally obtain the prediction accuracy of the test set samples;

[0093] Step 4.4.6: Respectively use traditional machine learning algorithms such as Gaussian Naive Bayes (GNB), Logistic Regression (LR), and Support Vector Machine (SVM), three Boosting algorithms of XGBoost, LightGBM, and CatBoost, deep learning algorithms such as neural networks, and the ensemble learning algorithm proposed in the present invention for effect comparison to verify the fault diagnosis effect of the ensemble learning algorithm, as Figure 7 shown.

[0094] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and purposes of the present invention.

Claims

1. A method for diagnosing inter-turn short circuit faults of an electromechanical actuator based on an ensemble learning algorithm, characterized in that It includes the following steps: Step 1: Analyze the composition principle and possible fault types of the electro-mechanical actuator, and speculate that the possible fault of the electro-mechanical actuator is the inter-turn short circuit fault of the permanent magnet synchronous motor; The inter-turn short circuit fault refers to the phenomenon that the insulation skin of the internal winding of the motor is damaged due to factors such as moisture, high temperature, and overcurrent, and the current is short-circuited between the windings. The external manifestation is that the three-phase current of the motor is unbalanced, and the motor speed and electromagnetic torque fluctuate greatly; Step 2: Establish a vector control model and a lumped parameter fault model of the electro-mechanical actuator through Simulink software, and determine the fault characteristics for diagnosis; Step 3: Establish an integrated learning fault diagnosis framework based on the Stacking model fusion strategy, and optimize the model built by the integrated learning fault diagnosis framework using K-fold cross-validation; Step 3.1: Preprocess the collected data to generate an available data set with labels, then perform feature extraction or feature dimension reduction according to needs, and finally divide the data set into a training set and a test set according to a certain ratio; Step 3.2: Adopt the method of K-fold cross-validation, randomly shuffle and evenly divide the training set into 5 parts, namely train1, train2, train3, train4, train5, and then perform 5 rounds of iterative training; when performing the i-th round of training, use the training set data other than traini to learn and train each primary learner, and use each trained primary learner to predict traini respectively to obtain the prediction result Yi; after 5 rounds of iteration, combine the Y1, Y2, Y3, Y4, Y5 obtained in each round of iteration as the features of the new training set; Step 3.3: Create a total of three primary learners, namely XGBoost, LightGBM, and CatBoost, train them based on the training set, and use the trained models to predict the training set and the test set respectively, and use the results as the new training set and the new test set; Step 3.4: Create an SVM main learner, train it based on the new training set, and use the trained SVM model to predict the new test set, and use the result as the final diagnosis result of the framework; Step 4: Build an electro-mechanical actuator test bench and process the collected output data, extract fault characteristics, and use the integrated learning fault diagnosis framework to realize fault diagnosis.

2. The method for diagnosing the inter-turn short circuit fault of the electro-mechanical actuator based on the integrated learning algorithm according to claim 1, characterized in that: In step 2, a lumped parameter fault model of the electro-mechanical actuator is established through Simulink software, and the fault characteristics for diagnosis are obtained by simulation. The fault characteristics are three-phase current, or harmonic signal of current, or vibration signal of the motor; Step 2.1: Calculate the inductance matrix of the faulty motor through Simulink software, set the motor fault to occur in one of the three phases, phase A, calculate the self-inductance and mutual inductance between phases under the fault, and form an inductance matrix. The calculation method is: Step 2.1.1: Given that the self-inductance of a normal motor phase is L and the mutual inductance is M. When a fault occurs, phase A is divided into a normal winding a and a short-circuited winding f. The calculation relationship of the inductance between the normal winding and the short-circuited winding is as follows: L′ aa + 2M af + L ff = L aa where, L aa represents the self-inductance of phase A, L’ aa represents the self-inductance of the normal winding a, M af represents the mutual inductance between the normal winding a and the short-circuited winding f, L ff represents the self-inductance of the short-circuited winding f; Step 2.1.2: Inductance leakage will occur during the actual operation of the motor. Therefore, the calculation relationship of the supplementary inductance is as follows: Step 2.1.3: The relationship between the inductances of the normal winding and the short-circuited winding also depends on the number of turns of the short-circuited coil. The calculation formula is as follows: Where n a is the number of turns of the normal winding a, and n b is the number of turns of the short - circuited winding f. From steps 2.1.1 - 2.1.3, the three inductance parameters L’ aa , M af , and L ff of the short - circuited and non - short - circuited windings of phase A can be calculated; Step 2.1.4: Mutual inductance M’ between the normal winding a and the B and C phase windings ab , and mutual inductance M between the short-circuited winding f and the B and C phase windings bf satisfy the calculation formula: M′ ab = (1 - σ)M M bf = σM In the formula, σ is the short-circuit ratio of phase A winding; Step 2.2: Based on the motor's own parameters and the inductance parameters calculated in Step 2.1, establish a faulty motor model: Step 2.2.1: Calculate the current values of each phase from the input voltage, self-resistance, inductance, and magnetic flux parameters of the motor. The calculation formula is as follows: Wherein, R is the phase resistance of the motor, U a 、U b 、U c are the input three-phase voltages, i a 、i b 、i c 、i f are the output currents, ψ ma 、ψ mb 、ψ mc 、ψ mf are parameters related to the magnetic flux linkage and the electrical angle; Step 2.2.2: Perform coordinate transformation on the phase current values, and calculate the electromagnetic torque T output by the motor according to the electromagnetic torque equation of the motor in the d-q coordinate system e , and the calculation formula is where p n is the number of pole pairs of the motor, i q is the q-axis current calculated after coordinate transformation of each phase current, and ψ f is the permanent magnet flux linkage; Step 2.3: Establish a cylindrical gear reducer model, and calculate the transmission ratio according to the number of teeth Z1 of the front gear, the number of teeth Z2 of the rear gear, and the transmission efficiency η1 Meanwhile, obtain the relationship between the input torque and the output torque T2 = i * T1, and the relationship between the input speed and the output speed T1 is the input torque of the reducer, T2 is the output torque of the reducer, ω1 is the input speed of the reducer, and ω2 is the output speed of the reducer; Step 2.4: Establish a planetary roller screw drive model. According to the number of screw threads n of the screw s , pitch p, lead s, the output torque T2 of the reducer, and transmission efficiency η2, calculate the output force of the screw as 3. The method for diagnosing the inter-turn short-circuit fault of an electromechanical actuator based on an integrated learning algorithm according to claim 1, characterized in that: The preprocessing in step 3.1 uses one-hot encoding preprocessing.

4. The method for diagnosing the inter-turn short-circuit fault of an electromechanical actuator based on an integrated learning algorithm according to claim 1, characterized in that: The feature extraction or feature dimension reduction in step 3.1 is as follows: Feature extraction includes performing Fourier transform and wavelet packet transform signal processing on the data to obtain frequency-domain data; Feature dimension reduction is to divide the samples. Each sample is a matrix of a×b. Take the mean and variance of every c rows. a is an integer multiple of c, and the data will be dimension-reduced. Feature extraction and feature dimension reduction are often carried out simultaneously.

5. The method for diagnosing the inter-turn short-circuit fault of an electromechanical actuator based on an integrated learning algorithm according to claim 1, characterized in that: In step 4, build an electromechanical actuator test bench and collect the output fault features, and use the integrated learning fault diagnosis framework to implement fault diagnosis: Step 4.1: Build an experimental platform for collecting fault data. This experimental platform consists of an electromechanical actuator EMA, a control unit, a loading device, a signal acquisition device, and a fault simulation device; Step 4.2: Based on the LabVIEW software, establish a data acquisition module, a motion control module, a data display module, and a data storage module for measuring EMA; Step 4.3: Through the experimental environment built in steps 4.1 and 4.2, collect the three-phase current data of EMA under different degrees of inter-turn short-circuit faults at a constant load and different displacement commands. To increase the data volume, conduct multiple experiments under different working conditions; Step 4.4: Process the current data collected in step 4.3 and use the integrated learning model established in step 3 for diagnosis: Step 4.4.1: Since the data collected in the experiment is in tdms file format, convert the tdms file to an xlsx file through excel, read the xlsx file using the pandas function library, and then convert the data in the read xlsx file into DataFrame type data that can be recognized by the machine learning model; Step 4.4.2: Create labels for the DataFrame current data transformed in Step 4.4.1 through one-hot encoding. The steps for creating labels are as follows: Encode N labels through an N-bit register. When creating the nth label, set the nth bit of the register to 1 and the other bits to 0; Step 4.4.3: Divide the data encoded in Step 4.4.2 into samples. Since the current varies periodically during the EMA uniform operation stage, consider H rows of data in the encoded data as one sample. Then, calculate the mean, mean absolute deviation, kurtosis, and skewness for each of the H rows of data separately to obtain 4 rows of statistical data. Next, merge the 4 rows of statistical data for all samples row by row to generate matrix P, and then flatten matrix P row by row to achieve feature dimensionality reduction. Divide the dimensionality-reduced data into a training set and a test set in a 7:3 ratio for the next training of the learner; Step 4.4.4: Adopt the method of K-fold cross-validation. First, evenly divide the training set into K parts and perform K rounds of iteration. When performing the i-th round of iteration, where 1 ≤ i ≤ K, use the other K - 1 parts except the i-th part as the sub-training set to train three primary learners, and use the trained primary learners to predict the i-th data that did not participate in the training in the current round of iteration to obtain the probability matrix Y of each sample in this data belonging to each health state. i ; After K rounds of iteration, the probability matrices Y1, Y2, …, Y i , …, Y K are merged by rows as the new training set; when generating features for the test set, use the three trained primary learners in each round of iteration to predict the entire test set. After K rounds of iteration, K sets of feature data are obtained. Perform mean processing on the K sets of data sets and use the result as the final new test set feature; use the new training set to train the called main learner SVM, and then use the main learner SVM to test the new test set. The test result is the final diagnosis result. Step 4.4.5: The classification effect of the main learner SVM in Step 4.4.4 highly depends on the parameters gamma and the penalty term C, and hyperparameter optimization is required. Hyperparameter optimization uses grid search, that is, first set the value ranges of the hyperparameters gamma and C. Then, divide the new training set in Step 4.4.4 into K equal parts. In each round of iteration, use K - 1 of these parts for hyperparameter optimization, and use the other 1 part (excluding K - 1 parts) to verify the effect of the hyperparameters. After K iterations, obtain the hyperparameter combination with the maximum accuracy; Use this hyperparameter combination to create an SVM model and train the model based on the new training set, and finally predict the accuracy through the test set samples.

Images