CN116520140A - Permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification - Google Patents
Permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification Download PDFInfo
- Publication number
- CN116520140A CN116520140A CN202211564923.0A CN202211564923A CN116520140A CN 116520140 A CN116520140 A CN 116520140A CN 202211564923 A CN202211564923 A CN 202211564923A CN 116520140 A CN116520140 A CN 116520140A
- Authority
- CN
- China
- Prior art keywords
- permanent magnet
- drive motor
- magnet direct
- data
- direct drive
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 31
- 238000000605 extraction Methods 0.000 title claims abstract description 21
- 238000000034 method Methods 0.000 claims abstract description 82
- 238000012549 training Methods 0.000 claims abstract description 45
- 238000011176 pooling Methods 0.000 claims abstract description 41
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 28
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 24
- 238000003745 diagnosis Methods 0.000 claims abstract description 23
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 230000006870 function Effects 0.000 claims description 31
- 238000004364 calculation method Methods 0.000 claims description 21
- 230000008569 process Effects 0.000 claims description 20
- 238000004458 analytical method Methods 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 17
- 238000010586 diagram Methods 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 238000012706 support-vector machine Methods 0.000 abstract description 50
- 238000013136 deep learning model Methods 0.000 abstract description 2
- 238000013473 artificial intelligence Methods 0.000 description 7
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000012545 processing Methods 0.000 description 7
- 239000013598 vector Substances 0.000 description 5
- 238000002405 diagnostic procedure Methods 0.000 description 4
- 238000001228 spectrum Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000032823 cell division Effects 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000001351 cycling effect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/34—Testing dynamo-electric machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Control Of Electric Motors In General (AREA)
Abstract
The invention discloses a permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification, which comprises the following steps: collecting data; preprocessing data; pooling through CNN for several times; determining a value of the SVM; initializing sample weights; training the weak classifier for multiple times to obtain a final classifier; and (5) fault diagnosis results. The invention is based on a Support Vector Machine (SVM) and a Convolutional Neural Network (CNN), and is combined with an integrated learning method, so that the characteristics of various faults of the corresponding permanent magnet direct drive motor can be extracted in a self-adaptive manner, and complex preprocessing of data is not needed. Therefore, fault features are not required to be manually extracted from a large amount of sample data, and the fault features are directly extracted through a deep learning model algorithm.
Description
Technical Field
The invention relates to the technical field of computer science, in particular to a permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification.
Background
At present, the researches on the fault diagnosis method of the permanent magnet direct drive motor can be summarized into three types: analytical model-based methods, signal processing-based methods, artificial intelligence-based methods.
(1) Diagnostic methods based on analytical model processing. According to the method, an analytical model is built for a diagnosis system by using a certain logic language according to the relation among various state parameters of the system, and the input quantity of the system is brought into the system model, so that the output of the system during normal operation is predicted according to the logic relation. Whether a fault occurs or not is judged by comparing whether obvious errors exist between the actual output and the predicted output of the system, and the accurate positioning of the fault type is realized. The main inclusion is: parameter identification under turn-to-turn short circuit fault of permanent magnet direct drive motor is realized by finite element method, and a method for analyzing fault by electric parameter model is provided according to unbalanced state of three-phase current and change of short circuit phase current; and establishing a geometric model for the normal state and the eccentric fault state of the permanent magnet direct-drive motor by using a time-step finite element method, and detecting whether the motor has eccentric faults.
(2) Diagnostic methods based on signal processing. The method is characterized in that signals of a motor are collected and processed, fault characteristics in the signals are analyzed on time domains and frequency domains, fault characteristics in normal motor signals are compared, time-frequency characteristics of the normal motor signals are compared, and fault characteristics are extracted by utilizing differences of the signals in normal and fault modes, so that the motor fault type is judged. The method comprises the steps of extracting and analyzing stator current and noise when the permanent magnet direct-drive motor is operated by adopting a sensor and a data acquisition card, wherein the noise frequency distribution is used for judging turn-to-turn short circuits of the permanent magnet direct-drive motor by extracting that the frequency distribution of the stator current is not changed with the amplitude change under the fault state; the frequency spectrums of the vibration signals and the current signals are obtained by carrying out Fourier transformation on the collected vibration signals and the current signals of each phase in operation, and then the vibration signal spectrums and the current signal spectrums are analyzed by a signal characteristic analysis method; the motor is controlled by a three-phase power supply, and the frequency spectrum of the voltage obtained by carrying out Fourier transformation on each phase voltage is subjected to multi-classification fault detection by taking the voltage harmonic wave of the stator as the characteristics of the rotor eccentric fault, the turn-to-turn short circuit fault and the demagnetizing fault of the permanent magnet direct drive motor.
(3) Diagnostic methods based on artificial intelligence. The method refers to a spool based on experience and logic reasoning of fault diagnosis accumulated for a long time, and utilizes artificial intelligence series to simulate human thinking judgment process so as to complete motor fault detection and diagnosis under the operation condition. At present, faults of permanent magnet direct drive motors are detected by mainly utilizing an artificial intelligence based method, wherein the most common and representative method is an artificial intelligence identification method of SVM. The core idea of the SVM fault diagnosis method is that: regions with higher data density are classified as positive and regions with lower data density are classified as negative. The intelligent recognition of fault diagnosis is realized by converting the multi-classification problem into a plurality of classification problems to be processed. The usual transformation methods are: one-to-many algorithm and one-to-one algorithm:
first, the "one-to-many" algorithm converts the A-class problem into A-class problems. A SVM bi-classifiers are constructed, and each SVM bi-classifier separates one class of samples from other classes. For example, the θ -th classifier is used to separate the θ -th classifier from the samples of other classes, and the classifier regards the i-th class as a positive class, and the other classes are all regarded as negative classes, and the corresponding classification decision function is:
f θ (x)＝sign(<w θ ，x>+b)
the A SVM two classifiers in the algorithm correspond to A classification decision functions, and respectively apply sample vectors x θ In the decision function with A classes, the sample vector x θ The classification function value is the category with the largest classification function value, and the formula is as follows:
the method can keep the trained classifier equal to the sample class, has fewer numbers and relatively high training speed.
Second, the "one-to-one" algorithm converts the A classification problem to an A (A-1)/2 "one-to-one" classification problem. The training sample can be combined and trained to obtain A (A-1)/2 SVM classifiers. In the test process, a voting method is adopted, and a trained SVM two-classifier is used for classifying unmeasured samples. The classifier classifies the results into which class, which class is added with a ticket, and finally calculates the class with the largest number of votes, namely, the class which is determined to be the class output by the test sample. The one-to-one multi-classification algorithm rarely has the situation of wrong classification, and even if a certain SVM classifier has the situation of wrong judgment, the final judgment result is not influenced.
The prior method has the following main defects: (1) The diagnostic method of analytical models is deficient in two ways: on one hand, the method is based on a specific mathematical formula, and various corresponding parameter values are determined, so that a specific analytical model is not easy to accurately establish under a complex use scene. The method has poor anti-interference capability to the outside, so that the accuracy is not up to standard in the process of complex fault detection, and the fault detection of the permanent magnet direct drive motor is error. On the other hand, the fault detection of the permanent magnet direct drive motor is performed by using an analytical model method, and more quantitative data are required. However, in the actual fault detection application process, some faults are not represented by a certain quantitative data, which results in a lack of flexibility and insufficient environment adaptability of the method for analyzing the model in the permanent magnet direct drive motor fault detection process.
(2) The signal processing diagnosis method is indeed simpler and more direct than the analysis model method, but in the signal processing method, some faults correspond to a plurality of different fault types, for example, faults of the permanent magnet direct drive motor comprise turn-to-turn short circuit faults, roller faults, demagnetizing faults and the like, and the fault types can not be accurately judged in some cases only by using one fault characteristic.
(3) In the artificial intelligence diagnosis method, the SVM is an important intelligent identification method for permanent magnet direct drive motor fault detection, is widely applied in solving the problem of permanent magnet direct drive motor fault detection, and is a relatively mature method in the existing artificial intelligence-based method. Under the fault detection scene of the permanent magnet direct-drive motor, the SVM method still has the following defects:
firstly, when the SVM method is used for fault detection of the permanent magnet direct drive motor, more features are extracted manually, so that one-sided and subjectivity of feature extraction can be possibly caused, and the accuracy of fault diagnosis is greatly affected;
secondly, the selection of important parameters in an SVM algorithm is greatly limited by experience of researchers, but the situation is very complex in the scene of fault diagnosis of a permanent magnet direct drive motor, the important parameters are required to be continuously adjusted, and the generalization capability of a model after parameter determination often cannot meet the requirement;
second, there are many types of permanent magnet direct drive motor faults and the size of the sample data collected during the actual application is quite large. The SVM method is complex in processing the multi-classification problem, and can only convert the multi-classification problem into the two-classification problem, so that difficulty is increased in calculating fault diagnosis, and meanwhile, a result is difficult to calculate large-scale sample data in a short time. However, if the calculation difficulty is reduced, the data size is reduced, which causes the data to be over-fitted, so that the accuracy of the fault diagnosis result is reduced.
Disclosure of Invention
Based on the defects existing in the prior art, the invention provides a permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification, which comprises the following specific technical scheme:
the permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification comprises the following steps:
step 1: collecting data;
step 2: preprocessing data;
step 3: pooling through CNN for several times;
step 4: determining a value of the SVM;
step 5: initializing sample weights;
step 6: training the weak classifier for multiple times to obtain a final classifier;
step 7: and (5) fault diagnosis results.
Specifically, the data in step 1 includes a vibration signal, a current signal and a temperature signal;
the vibration signals are collected through a vibration sensor on the permanent magnet direct-drive motor;
the current signal is collected through a current sensor on the permanent magnet direct-drive motor;
the temperature signal is collected through a temperature sensor on the permanent magnet direct-drive motor.
Specifically, the preprocessing in step 2 includes simple trend analysis, comparison analysis and subdivision analysis, and data screening is performed through analysis, so that invalid or data with larger deviation are deleted.
Specifically, the step 3 includes the following substeps:
step 31: inputting data into a convolution layer to obtain a characteristic signal matrix;
step 32: and (3) carrying out a pooling process, wherein the calculation formula is as follows:
wherein,,is the weight value of the connection between the characteristic signal and the pooling layer, down (x) is the sampling function,/->Is a bias term;
the pooling comprises maximum pooling and average pooling, wherein the maximum pooling refers to the pooling of a signal region by selecting the maximum value of the region as the value of the region; the average pooling is to take the average value of the preference area as a pooled value;
step 33: and outputting a result by the full-connection layer, wherein each layer of nodes on the full-connection layer are connected with all nodes of the previous layer, the full-connection layer does not have the capability of extracting the characteristics, all the characteristics extracted in the front are integrated to obtain corresponding output, weights are continuously updated in the full-connection layer, and the final output result is obtained by summing according to the weights and the output results of all the layers, wherein the calculation formula is as follows:
y l ＝f(w l z l-1 +b l )
wherein y is l Is the output of the full connection layer, w l Represent the weight, z l-1 Is the input of the full connection layer, b l Is a bias term, f (x) is a classification function;
step 34: the output result is input to the SVM.
Specifically, step 31 specifically includes:
inputting various signal data into a convolutional neural network to form a matrix with the size w, a convolutional kernel with the size k, a stride with the size s, a filling layer number p, and a characteristic diagram M with the size after convolution:
the convolution includes continuous convolution and discrete convolution, wherein the continuous convolution has the following calculation formula:
where f (t) represents the input, g (t) represents the convolution kernel, and s (t) represents the output;
the discrete convolution is calculated as:
where f (α) represents the input, g (α) represents the convolution kernel, and s (α) represents the output.
Specifically, the step 4 includes the following substeps:
step 41: the self-adaptive extracted characteristic signals after convolutional neural network convolutional pooling are grouped, and the characteristic signals are randomly grouped by using a 6-time 6-fold cross validation method to separate a training data set and a validation data set;
step 42: performing corresponding time domain analysis on various collected signals;
step 43: training the SVM algorithm by using the corresponding classification labels in the training set;
the construction method of the SVM algorithm comprises a direct method, an indirect method and a hierarchical method;
step 44: taking the SVM obtained after training as a sub-classifier;
the SVM controls the performance of the weak classifier by adjusting the kernel function delta and penalty coefficient C.
Specifically, the weights described in step 5 are expressed as:
using weighted D t Learning is carried out on the training set of the (2) to obtain a weak classifier:
h t (x)：X→{-1，+1}
the training set is as follows:
S＝{(x 1 ，y 1 )，(x 2 ，y 2 )，(x 3 ，y 3 )，…(x n ，y n )}
wherein x is i ∈x，y i ∈{-1，+1}；
Calculate h t (x) Classification error W on training set t The calculation formula is as follows:
calculating h by classification error t (x) The weight sigma occupied on the final classifier t The calculation formula is as follows:
specifically, the step 6 specifically includes:
multiple iterations of updating weight distribution D t+1 The calculation formula is:
according to the weak classifier weight sigma t The difference combines the weak classifiers to obtain a final classifier, the calculation formula is:
。
specifically, the step 7 specifically includes:
training to obtain a final model M, and obtaining a decision function of an Adaboost algorithm, wherein the calculation formula is as follows:
and (5) using a final model M to realize fault diagnosis.
The invention can achieve the following beneficial effects:
(1) The invention is based on a Support Vector Machine (SVM) and a Convolutional Neural Network (CNN), and is combined with an integrated learning method, so that the characteristics of various faults of the corresponding permanent magnet direct drive motor can be extracted in a self-adaptive manner, and complex preprocessing of data is not needed. Therefore, fault features are not required to be manually extracted from a large amount of sample data, and the fault features are directly extracted through a deep learning model algorithm.
(2) According to the model training method, after a large number of normal data of different devices are trained, the influence of researchers on detection results by experience selection parameters is overcome, and the convolutional neural network is utilized for adjustment, so that the generalization capability of the model is improved.
(3) The convolution neural network controls the partial fitting capacity of the overall model by utilizing different convolution, pooling and final output characteristic vectors, reduces vector dimension under the condition of over fitting, and improves the output dimension of the convolution layer under the condition of under fitting.
(4) According to the invention, the accuracy of the weak classifier is adjusted by utilizing the weight rule of the multiple-classification Adaboost algorithm, so that the time efficiency of the whole model diagnosis is improved, and the time cost of permanent magnet direct-drive motor fault detection is saved.
Drawings
FIG. 1 is a flow chart of the technical scheme of the invention;
FIG. 2 is a flow chart of step 3 of the present invention;
fig. 3 is a schematic diagram of an operation process of the SVM of the present invention based on the Adaboost algorithm to satisfy the performance of the weak classifier.
Detailed Description
For a clearer understanding of technical features, objects, and effects of the present invention, a specific embodiment of the present invention will be described with reference to the accompanying drawings.
The flow of the invention is shown in figure 1, step 1: and (5) data acquisition. Corresponding various signal data are collected, and three types of signals are mainly contained: vibration signal, current signal, temperature signal. Respectively acquiring vibration signals of the permanent magnet direct drive motor through vibration sensors; collecting a current signal of the permanent magnet direct-drive motor through a current sensor; and acquiring a temperature signal of the permanent magnet direct-drive motor through a temperature sensor. In addition, various signal data useful for fault detection of the permanent magnet direct drive motor can be acquired through other various sensors.
Step 2: and (5) preprocessing data. And (3) preprocessing the data, screening the data by simple data processing modes such as simple trend analysis, comparison analysis, subdivision analysis and the like on various signal data acquired in the step (1), and deleting some invalid data with larger deviation.
Step 3: pooled by CNN through several convolutions. Inputting various signal data acquired in the step 2 into a Convolutional Neural Network (CNN), firstly converting various preprocessed signal sample data into a matrix, wherein matrix elements correspond to various signal data, and then carrying out convolution calculation on the signal matrix and a convolution kernel to obtain a characteristic signal matrix. Wherein the convolution kernel is the most important part of the convolution layer and is also the key of signal matrix extraction features.
Step 4: the value of the SVM is determined. And (3) concentrating the feature vectors extracted in the step (3) into corresponding classification labels to train an SVM algorithm so as to obtain the value of the SVM.
Step 5: sample weights are initialized. And (3) combining the SVM value in the step (4) with the Adaboost algorithm after different weights are given, so as to obtain a trained model.
Step 6: and training the weak classifier for multiple times to obtain a final classifier. Repeatedly adjusting the punishment coefficient C and the kernel function alpha of the training model obtained in the step 5, and performing repeated iterative training on the model trained in the step 4 to obtain a final classifier;
step 7: and (5) fault diagnosis results. And (3) verifying the data in the verification set through the final model trained in the step (6), so as to obtain a detection result of the model, and finding out the best model from the detection result for fault detection of the permanent magnet direct drive motor.
The specific flow of the step 3 is shown in fig. 2.
(1) Inputting the convolution layer to obtain a characteristic signal
Inputting various signal data into a convolutional neural network to form a matrix with the size w, a convolutional kernel with the size k, a stride with the size s, a filling layer number p, and a characteristic diagram M with the size after convolution:
the output matrix generated after the convolution kernel and the matrix of the input signal samples are the characteristic signal matrix (the effect of data dimension reduction can be achieved in the convolution process). The convolution layer comprises a plurality of convolution units, and data features of input samples are adaptively extracted at the same time, and continuous convolution and discrete convolution are included in the operation process. The specific formula selected for the convolution formula is as follows:
first, continuous convolution:
where f (t) represents the input, g (t) represents the convolution kernel, and s (t) represents the output.
Second, discrete convolution:
where f (α) represents the input, g (α) represents the convolution kernel, and s (α) represents the output. In the actual application process of permanent magnet direct drive motor fault diagnosis, input data are often multidimensional, taking a two-dimensional format as an example: let the input signal data sample be the result of an operation with P (i, j), Q (m, n) being the two-dimensional convolution kernel, W (i, j). The procedure of the operation is as follows:
where m and n are the size of the convolution kernel. The process of adding bias to the input signal sample matrix after convolution operation and finally obtaining the characteristic signal matrix through an activation function is as follows:
wherein M is j Is a set of characteristic signals that are to be processed,is the J-th characteristic signal of the L-th layer output,>is an L-layer input,>is a convolution kernel matrix, < >>Is the bias term and f (x) is the activation function. In the process of selecting the activation function, a more commonly used Relu function can be selected, the learning efficiency of the model can be greatly improved, and the expression is as follows:
(2) Information filtering by pooling layer
The most important thing in the pooling process is the selection and construction of the pooling function, wherein different pooling modes are included, and the pooling process is expressed as:
wherein,,is special toThe weight value connected between the sign signal and the pooling layer, down (x) is a sampling function, and includes pooling mode in pooling process, < + >>Is a bias term. There are two common methods of maximum pooling and average pooling in the pooling process. Maximum pooling refers to the pooling of the signal region by selecting the maximum value of the region as the value of the region; the average pooling is to take the average value of the favorite area as a pooled value, and the dimension of the characteristic signal can be obviously reduced after pooling.
(3) Full connection layer output result
After the rolling and pooling are alternately performed, the method enters a full-connection layer, each layer of nodes on the full-connection layer are connected with all nodes of the upper layer, and the method does not have the capability of feature extraction, and integrates all the features extracted in the prior art to obtain corresponding output. The weight is continuously updated in the full connection layer, and the final output result is obtained by summing according to the weight and the output results of all the layers and can be expressed as:
y l ＝f(w l z l-1 +b l )
wherein y is l Is the output of the full connection layer, w l Represent the weight, z l-1 Is the input of the full connection layer, b l Is the bias term and f (x) is the classification function.
(4) And finally, inputting the output result finally obtained by the full connection layer into the SVM.
The specific flow of step 4 to step 7 is shown in fig. 3.
From the perspective of fault diagnosis of the permanent magnet motor, the SVM needs to classify 6 states in fig. 3, but the SVM is essentially a classifier, so that the SVM needs to be correspondingly improved, and the specific flow is as follows:
(1) And grouping the characteristic signals which are adaptively extracted after convolutional neural network convolutional pooling, and randomly grouping the characteristic signals by using a 6-time 6-fold cross validation method to separate a training data set and a validation data set.
(2) And carrying out corresponding analysis on various acquired signals, for example: and carrying out frequency domain analysis on the acquired vibration signals, carrying out time domain analysis on the acquired currents, and carrying out time domain analysis on the acquired temperature signals.
(3) And training the SVM algorithm by using the corresponding classification labels in the training set. The multi-classification SVM model is constructed by the following three methods: first, the direct method: and integrating all single-class parameters meeting the maximization of the SVM algorithm separation hyperplane into a new optimization problem with constraint by directly changing an objective function, and solving the optimization problem to obtain the geometric hyperplane of the multi-classification model. Second, indirect method: and combining a plurality of two-classification sample classes into a multi-classification problem by adopting a mode that one sample class corresponds to one sample class for training or one sample class corresponds to a plurality of sample classes for training to obtain a plurality of classifiers. In the process, two samples are randomly extracted from one sample class corresponding to one sample class for training, and the two samples are combined into a multi-classification problem after circulation; one sample class corresponds to a plurality of sample classes in a way that one sample class is randomly extracted from the training sample set to serve as one sample class, and all the rest sample sets serve as the other sample class. Third, hierarchical: similar to cell division, the entire training set is initially divided into two training sample classes, which are then each further divided, cycling through, and ending when the divided training sample classes are single classes.
(4) And taking the SVM obtained after training as a sub-classifier. The SVM can well control the performance of the weak classifier by selecting two parameters, namely a proper kernel function delta and a punishment coefficient C.
(5) Initializing the weight of the weak classifier, and performing repeated iterative training: assume that the training set is s= { (x) 1 ，y 1 )，(x 2 ，y 2 )，(x 3 ，y 3 )，...(x n ，y n ) X, where x i ∈x，y i E { -1, +1}; the weights of the initialized weak classifiers are:for t=1, 2,3, T, using weighted D t Learning is carried out on the training set of the (2) to obtain a weak classifier:
h t (x)：X→{-1，+1}
calculate h t (x) Classification error W on training set t The following are provided:
calculating h by classification error t (x) The weight sigma occupied on the final classifier t The following are provided:
multiple iterations of updating weight distribution D t+1 The following are provided:
(6) And obtaining a final classifier. According to the weak classifier weight sigma t The differences combine the individual weak classifiers as follows:
(7) Training to obtain a final model M: obtaining a decision function of an Adaboost algorithm:
when the value in the brackets of the decision function is greater than 0, the output value of the function value is 1, so that the decision result is correct; otherwise, the output value is-1, so the judgment result is wrong.
(1) In the step 3, the convolutional neural network CNN is utilized to carry out convolutional pooling correlation processing on the sensing signal data of various permanent magnet direct drive motors detected by the sensors, corresponding characteristic signals are obtained and then input into an SVM algorithm, compared with the traditional simple manual characteristic extraction, the characteristic self-adaptive extraction is carried out through the convolutional neural network by using artificial intelligent equipment such as a computer, so that the characteristic which is not extracted manually can be extracted, and the comprehensiveness of fault detection is improved;
(2) In the step 3, the convolution neural network can control the dimension of the extracted characteristic signals, can reduce the dimension of characteristic signal data when the convolution neural network is in excess of the convolution neural network, can improve the dimension of the characteristic signal data when the convolution neural network is in insufficient of the convolution neural network, and controls the calculation scale of the SVM and the fitting of the characteristic signal data by the convolution neural network, so that the calculation efficiency is improved, the calculation difficulty is reduced, and the time cost for fault detection is reduced. Meanwhile, the dimension of the characteristic signal can be adaptively adjusted through the convolutional neural network, so that the model is more suitable for fault detection of large-scale data, and the generalization capability of the model is improved.
(3) In the step 4, the SVM is subjected to repeated iterative training through an Adaboost algorithm, and the performance requirement of the weak classifier is met through continuous adjustment of a kernel function and a penalty coefficient. And (3) adjusting the kernel function and the penalty coefficient at the beginning stage, keeping the proper penalty coefficient unchanged, enabling the classification precision to be lower than a threshold value, and only adjusting the kernel function in the later process. Compared with the traditional method for solving the multi-classification problem of permanent magnet direct drive motor fault detection by adopting the SVM alone, the method combining the SVM and the Adaboost has the advantage that the accuracy is improved.
(4) In step 4, the sample is initialized by the Adaboost algorithm, so that different classification labels have different weight values, classification conditions are more suitable for multi-classification fault types of the permanent magnet direct drive motor, the fault types can be detected as comprehensively and accurately as possible, and the capability of the SVM in solving fault diagnosis of the permanent magnet direct drive motor is improved.
In step 4, besides using the Adaboost algorithm to use the SVM as a weak classifier to improve the fault detection accuracy of the model, the integrated learning algorithm such as Bagging, stacking and the like can be used to combine with the SVM algorithm, so that the effects of reducing errors and improving the accuracy of the model prediction result can be achieved.
(1) The invention is applied to fault detection of the permanent magnet direct drive motor, and the main idea of the model is to combine the strong feature extraction capability of CNN with the efficient classification capability of SVM, and the accuracy and efficiency of fault detection of the permanent magnet motor are improved by utilizing the characteristics of an Adaboost integrated learning optimization model.
(2) The invention extracts characteristic signals by detecting various signals of the permanent magnet direct drive motor through sensor equipment, converting the various signals into a signal matrix, and obtaining corresponding characteristic signals through CNN convolution pooling. Also in this process, CNN can adjust the dimensionality of the feature signal by convolutionally pooled functions.
(3) The invention carries out random grouping on the characteristic signals through a cross verification method to obtain a training set and a verification set, and then trains the characteristic signals of the training set by using a multi-classification SVM to obtain the SVM meeting the weak classifier.
(4) The Adaboost algorithm is mainly applied to the method by initializing sample data, taking the SVM as a weak classifier, carrying out repeated iterative training to obtain a final classifier, and finally forming a fault diagnosis decision function.
The foregoing has shown and described the basic principles and main features of the present invention and the advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims.
Claims (9)
1. The permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification is characterized by comprising the following steps of:
step 1: collecting data;
step 2: preprocessing data;
step 3: pooling through CNN for several times;
step 4: determining a value of the SVM;
step 5: initializing sample weights;
step 6: training the weak classifier for multiple times to obtain a final classifier;
step 7: and (5) fault diagnosis results.
2. The method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 1, wherein the data in the step 1 comprises a vibration signal, a current signal and a temperature signal;
the vibration signals are collected through a vibration sensor on the permanent magnet direct-drive motor;
the current signal is collected through a current sensor on the permanent magnet direct-drive motor;
the temperature signal is collected through a temperature sensor on the permanent magnet direct-drive motor.
3. The method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 1, wherein the preprocessing in the step 2 comprises simple trend analysis, contrast analysis and subdivision analysis, and screening data is carried out through analysis, and invalid or data with larger deviation is deleted.
4. The method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 1, wherein the step 3 comprises the following substeps:
step 31: inputting data into a convolution layer to obtain a characteristic signal matrix;
step 32: and (3) carrying out a pooling process, wherein the calculation formula is as follows:
wherein,,is a characteristic signal and a pooling layerThe weight value of the connection between, down (x) is the sampling function, +.>Is a bias term;
the pooling comprises maximum pooling and average pooling, wherein the maximum pooling refers to the pooling of a signal region by selecting the maximum value of the region as the value of the region; the average pooling is to take the average value of the preference area as a pooled value;
step 33: and outputting a result by the full-connection layer, wherein each layer of nodes on the full-connection layer are connected with all nodes of the previous layer, the full-connection layer does not have the capability of extracting the characteristics, all the characteristics extracted in the front are integrated to obtain corresponding output, weights are continuously updated in the full-connection layer, and the final output result is obtained by summing according to the weights and the output results of all the layers, wherein the calculation formula is as follows:
y l ＝f(w l z l-1 +b l )
wherein y is l Is the output of the full connection layer, w l Represent the weight, z l-1 Is the input of the full connection layer, b l Is a bias term, f (x) is a classification function;
step 34: the output result is input to the SVM.
5. The method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification as claimed in claim 4, wherein the step 31 is specifically:
inputting various signal data into a convolutional neural network to form a matrix with the size w, a convolutional kernel with the size k, a stride with the size s, a filling layer number p, and a characteristic diagram M with the size after convolution:
the convolution includes continuous convolution and discrete convolution, wherein the continuous convolution has the following calculation formula:
where f (t) represents the input, g (t) represents the convolution kernel, and s (t) represents the output;
the discrete convolution is calculated as:
where f (α) represents the input, g (α) represents the convolution kernel, and s (α) represents the output.
6. The method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 1, wherein the step 4 comprises the following substeps:
step 41: the self-adaptive extracted characteristic signals after convolutional neural network convolutional pooling are grouped, and the characteristic signals are randomly grouped by using a 6-time 6-fold cross validation method to separate a training data set and a validation data set;
step 42: performing corresponding time domain analysis on various collected signals;
step 43: training the SVM algorithm by using the corresponding classification labels in the training set;
the construction method of the SVM algorithm comprises a direct method, an indirect method and a hierarchical method;
step 44: taking the SVM obtained after training as a sub-classifier;
the SVM controls the performance of the weak classifier by adjusting the kernel function delta and penalty coefficient C.
7. The method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 1, wherein the weight in the step 5 is expressed as:
using weighted D t Learning is carried out on the training set of the (2) to obtain a weak classifier:
h t (x):X→{-1,+1}
the training set is as follows:
S＝{(x 1 ，y 1 )，(x 2 ，y 2 )，(x 3 ，y 3 )，...(x n ，y n )}
wherein x is i ∈x，y i ∈{-1，+1}；
Calculate h t (x) Classification error W on training set t The calculation formula is as follows:
calculating h by classification error t (x) The weight sigma occupied on the final classifier t The calculation formula is as follows:
8. the method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 7, wherein the step 6 is specifically:
multiple iterations of updating weight distribution D t+1 The calculation formula is:
according to the weak classifier weight sigma t The difference combines the weak classifiers to obtain a final classifier, the calculation formula is:
。
9. the method for detecting faults of a permanent magnet direct drive motor by combining feature extraction and small sample classification according to claim 8, wherein the step 7 is specifically:
training to obtain a final model M, and obtaining a decision function of an Adaboost algorithm, wherein the calculation formula is as follows:
and (5) using a final model M to realize fault diagnosis.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211564923.0A CN116520140A (en) | 2022-12-07 | 2022-12-07 | Permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211564923.0A CN116520140A (en) | 2022-12-07 | 2022-12-07 | Permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116520140A true CN116520140A (en) | 2023-08-01 |
Family
ID=87398185
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211564923.0A Pending CN116520140A (en) | 2022-12-07 | 2022-12-07 | Permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116520140A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117233603A (en) * | 2023-11-16 | 2023-12-15 | 中国航空工业集团公司沈阳空气动力研究所 | Permanent magnet synchronous motor monitoring device operated in vacuum environment and fault diagnosis method |
| CN117435981A (en) * | 2023-12-22 | 2024-01-23 | 四川泓宝润业工程技术有限公司 | Method and device for diagnosing operation faults of machine pump equipment, storage medium and electronic equipment |
| CN117972623A (en) * | 2023-12-20 | 2024-05-03 | 东方电气集团科学技术研究院有限公司 | Intelligent fault diagnosis method for power generation equipment based on ensemble learning |
-
2022
- 2022-12-07 CN CN202211564923.0A patent/CN116520140A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117233603A (en) * | 2023-11-16 | 2023-12-15 | 中国航空工业集团公司沈阳空气动力研究所 | Permanent magnet synchronous motor monitoring device operated in vacuum environment and fault diagnosis method |
| CN117972623A (en) * | 2023-12-20 | 2024-05-03 | 东方电气集团科学技术研究院有限公司 | Intelligent fault diagnosis method for power generation equipment based on ensemble learning |
| CN117435981A (en) * | 2023-12-22 | 2024-01-23 | 四川泓宝润业工程技术有限公司 | Method and device for diagnosing operation faults of machine pump equipment, storage medium and electronic equipment |
| CN117435981B (en) * | 2023-12-22 | 2024-03-01 | 四川泓宝润业工程技术有限公司 | Method and device for diagnosing operation faults of machine pump equipment, storage medium and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109949317B (en) | Semi-supervised image example segmentation method based on gradual confrontation learning | |
| CN113191215B (en) | Rolling bearing fault diagnosis method integrating attention mechanism and twin network structure | |
| CN116520140A (en) | Permanent magnet direct drive motor fault detection method integrating feature extraction and small sample classification | |
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| CN109582003B (en) | Bearing fault diagnosis method based on pseudo label semi-supervised kernel local Fisher discriminant analysis | |
| CN103728551B (en) | A kind of analog-circuit fault diagnosis method based on cascade integrated classifier | |
| CN113673346B (en) | Motor vibration data processing and state identification method based on multiscale SE-Resnet | |
| CN106845421A (en) | Face characteristic recognition methods and system based on multi-region feature and metric learning | |
| CN113408605A (en) | Hyperspectral image semi-supervised classification method based on small sample learning | |
| CN114564982B (en) | Automatic identification method for radar signal modulation type | |
| CN114841257B (en) | Small sample target detection method based on self-supervision comparison constraint | |
| CN109993236A (en) | Few sample language of the Manchus matching process based on one-shot Siamese convolutional neural networks | |
| CN114048568A (en) | Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework | |
| CN110880369A (en) | Gas marker detection method based on radial basis function neural network and application | |
| CN113920375B (en) | Fusion characteristic typical load identification method and system based on combination of Faster R-CNN and SVM | |
| CN115435892A (en) | Intelligent fault diagnosis method for complex electromechanical equipment | |
| CN117315380B (en) | Deep learning-based pneumonia CT image classification method and system | |
| CN111310719B (en) | Unknown radiation source individual identification and detection method | |
| CN112766283A (en) | Two-phase flow pattern identification method based on multi-scale convolution network | |
| CN112504682A (en) | Chassis engine fault diagnosis method and system based on particle swarm optimization algorithm | |
| CN115859142A (en) | Small sample rolling bearing fault diagnosis method based on convolution transformer generation countermeasure network | |
| CN111639697A (en) | Hyperspectral image classification method based on non-repeated sampling and prototype network | |
| CN111144462A (en) | Unknown individual identification method and device for radar signals | |
| CN110610203A (en) | Electric energy quality disturbance classification method based on DWT and extreme learning machine | |
| CN113111774B (en) | Radar signal modulation mode identification method based on active incremental fine adjustment |
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 |