CN115015752B

CN115015752B - Motor fault diagnosis method based on sparse decomposition and neighborhood swarm algorithm

Info

Publication number: CN115015752B
Application number: CN202210618232.8A
Authority: CN
Inventors: 戴峻峰; 常波
Original assignee: Huaiyin Institute of Technology
Current assignee: Huaiyin Institute of Technology
Priority date: 2022-06-01
Filing date: 2022-06-01
Publication date: 2024-08-20
Anticipated expiration: 2042-06-01
Also published as: CN115015752A

Abstract

The invention discloses a motor fault diagnosis method based on sparse decomposition and a neighborhood swarm algorithm, which utilizes IM-HHT to extract direct current motor characteristics and signal analysis. Carrying out sparse decomposition denoising treatment on a motor sampling signal, denoising the motor signal before treatment by utilizing an orthogonal matching pursuit algorithm, and selecting a plurality of optimal atomic linear combinations from an overcomplete dictionary; decomposing a given signal into a plurality of IMFs by using empirical mode decomposition, removing false IMFs by using correlation parameters, and finally performing Hilbert transformation to obtain a Hilbert spectrum of an original signal; selecting features based on a focused Euclidean distance judging method; the square neighborhood is used for selecting a manual bee colony algorithm to finish sequencing the motor signal characteristic factors; and classifying and identifying motor faults through a radial basis classifier. Compared with the prior art, the method can realize autonomous fault diagnosis of the motor aiming at direct current motor bearing faults, stator winding faults and rotor faults.

Description

Motor fault diagnosis method based on sparse decomposition and neighborhood swarm algorithm

Technical Field

The invention belongs to the technical field of motor fault diagnosis, and particularly relates to a motor fault diagnosis method based on sparse decomposition and a neighborhood swarm algorithm.

Background

The motor fault problem commonly existing in manufacturing production equipment is solved by the following general methods in precision machining enterprises at present: on the one hand, planned overhauling is carried out, namely, the motor is overhauled and cleaned regularly, the operation standard of the production equipment is standardized as much as possible, the cultivation of operators and maintenance personnel is enhanced, the operators are familiar with the equipment structure, the performance of the motor used by the equipment is known, basic maintenance measures are understood, the equipment maintenance and the maintenance are enhanced, the normal operation of the production equipment is ensured, the service life is prolonged, but in the actual implementation process, the operators are more frequently replaced, the number of the motors is too many, the installation position is inconvenient, and the like, so that the mode is not maintained in place. Because it cannot be determined whether the motor has a fault, the existing fault hidden trouble of some motors may be missed, and the motors without hidden trouble may cause artificial fault hidden trouble due to accidents during disassembly and reassembly, besides, the scheduled maintenance consumes a great deal of manpower, material resources, financial resources and time, various resources are not reasonably utilized, the normal production time of enterprises is reduced, the production efficiency is reduced, and in fact, great economic loss is caused for the enterprises. The traditional fault diagnosis method has some defects, such as excessive dependence on priori knowledge, certain constraint on generalization capability and the like, and can not meet the current motor fault monitoring requirements.

Disclosure of Invention

The invention aims to: aiming at the problems in the prior art, the invention provides a motor fault diagnosis method based on sparse decomposition and a neighborhood swarm algorithm, which ensures that a system has higher precision and fewer characteristics and improves the precision and efficiency of fault diagnosis.

The technical scheme is as follows: the invention provides a motor fault diagnosis method based on sparse decomposition and a neighborhood swarm algorithm, which comprises the following steps:

step 1: extracting the characteristics and signal analysis of the direct current motor by using sparse decomposition IM-HHT; denoising a motor sampling signal through sparse decomposition, denoising the motor signal before processing by using an orthogonal matching pursuit algorithm, selecting a plurality of optimal atoms from an overcomplete dictionary DCT, linearly combining the selected atoms, decomposing an EMD (empirical mode decomposition), decomposing a given signal into a plurality of intrinsic mode functions IMFs by using an empirical mode decomposition EMD, removing false IMFs by using a correlation parameter, and finally performing Hilbert transform to obtain a Hilbert spectrum of an original signal;

Step 2: feature selection is carried out based on a aggregation type Euclidean distance judging method;

Step 3: selecting a artificial bee colony SS-ABC algorithm based on the square neighborhood to sort the motor signal characteristic factors; the square neighborhood selection artificial bee colony SS-ABC algorithm is probability selection by replacing the original ABC algorithm with a square neighborhood selection method, and a search strategy is modified and constructed, so that improvement of the square neighborhood selection artificial bee colony SS-ABC algorithm is realized;

step 4: and classifying and identifying motor faults through a radial basis function RBF classifier.

Further, in the step 1, the specific operation of denoising before processing the motor signal by using the orthogonal matching pursuit algorithm is as follows:

(1) Characterizing signals to be decomposed and models

Wherein F _i is a signal to be decomposed, f= { F ₁,f₂,...f_n } is a signal matrix to be decomposed, m is the number of selected atoms, d= { D ₁,d₂,...d_n } is an overcomplete dictionary, D _i is a selected atom, and x _i is a sparse coefficient corresponding to the selected atom. The selected m atoms are linearly combined to sparsely represent the signal;

The sparsest representation of x _i is found under the condition that equation (1) is satisfied, i.e., the solution with the least non-zero value is sought, the model is as follows:

Wherein, |·| ₀ is the l ₀ norm, i.e. the number of non-zero elements in X _i, x= { X _i } is a sparse coefficient matrix, and ε is an error;

(2) Sparse decomposition

Selecting the maximum value of the inner product absolute value between the N-dimensional signal matrix F and the atom D ₁ from the overcomplete atom library D, wherein D ₁ is the optimal atom, namely, the method meets the following conditions:

|<F，d₁>|＝sup|<F,d_i>| (3)

Wherein, | < F, d ₁ > | represents the absolute value of the inner product of the signal matrix F and the atom d ₁;

the signal after decomposing the signal matrix F a plurality of times can be expressed as:

As the number of decompositions increases, the signal value of the residual signal will be approximately zero, and the signal matrix F to be decomposed can be expressed as:

further, the selection of the overcomplete dictionary DCT in step 1 is as follows:

The DCT dictionary is obtained according to discrete cosine transform, the real numbers are used for replacing complex numbers to analyze on the real number domain through symmetrical signal expansion, and one-dimensional DCT transform can be calculated by the following formula:

wherein d (k) is the kth DCT atom, k is the frequency factor, f (N) is a given signal sequence, and N represents the length of the input signal; the matrix form is represented as follows:

D＝C_Nf (7)

C _N is DCT coefficient matrix, and then the over-complete dictionary can be obtained by fine sampling the DCT coefficient matrix in the frequency domain.

Further, in the step 1, an EMD is decomposed by using an empirical mode, a given signal is decomposed into a plurality of IMFs, then a correlation parameter is utilized to remove a false IMF, and finally Hilbert transformation is performed, so that the specific operation of obtaining the Hilbert spectrum of the original signal is as follows:

1) Decomposing a given signal into a plurality of intrinsic mode functions IMFs by using an empirical mode decomposition EMD;

let x (t) denote the original function of the input and be decomposed into n IMFs by EMD, the EMD model is defined as follows:

2) Defining correlation parameters and removing false IMF

Let IMF component generated by EMD decomposition be x, motor fault signal be y, define the correlation parameter as r:

Wherein:

3) Completing Hilbert transformation and obtaining a corresponding Hilbert spectrum;

Performing Hilbert transformation on the IMF of each motor state signal to obtain a Hilbert spectrum of the motor state signal; HT model is defined as follows: :

Where Pv is a warning value for avoiding an abnormality when τ=t and τ= ±infinity, the hilbert spectrum is defined as follows:

Where H _l (t) represents IMF, H _l (t) is obtainable by hilbert transform, a _l (t) is instantaneous amplitude, and θ _l (t) is instantaneous phase angle.

Further, the specific operation of finding the feature selection based on the aggregate type euclidean distance determination method in step 2 includes:

1) Calculating variances of all samples of the mth feature Average value of

Wherein,

2) Calculating variance of mth feature of class C sampleAverage value of

Wherein,

3) Calculating the weighted variance of cluster center g _c at the mth feature

Wherein,

4) Calculating an inter-class distance d _b ^m and an intra-class distance d _w ^m of the mth feature:

Wherein,

5) In the mth feature, the variance factors v _b ^m of d _b ^m and v _w ^m of d _w ^m are calculated:

Wherein,

6) Calculating a compensation coefficient eta ^m of the m-th feature:

7) Calculating a distance discrimination factor lambda ^m of the mth feature, and obtaining normalized lambda ^m′:

further, the specific step of selecting the artificial bee colony SS-ABC algorithm to sort the motor signal characteristic factors based on the square neighborhood in the step 3 is as follows:

1) Employment of bee search phases

Let X _i＝{X_i,1,X_i,2,...,X_i,D be the ith solution in the group, N be the total number of groups, D be the dimension, the employment bees at each solution X _i (i=1, 2.,; N) surrounding search, trying to find the optimal solution:

wherein x _k is randomly extracted from the whole population, jr is a random integer within [1, D ], and phi _i,jr is randomly generated at [ -1,1 ];

the new solution v _i generates the following:

Where j=1, 2, D; a better solution is obtained using a greedy selection method, as:

2) Following the bee search phase

2.1 Using square neighborhood selection method to replace original probability selection method, and modifying and constructing search strategy based on square neighborhood:

The concept of square neighborhood is utilized, the solution of the group is set to be square topology, and whether the data points are in the neighborhood is judged by the following formula:

Wherein x _i and y _i are two square neighborhood points, sl is the square neighborhood side length, and m is the neighborhood boundary coefficient;

2.2 Modifying the search strategy): modifying and constructing a search strategy based on the square neighborhood side length;

For each solution x _i, selecting the best solution in the square neighborhood of x _i as x _ib, and searching in the vicinity of x _ib by using a corresponding search strategy; during the search, x _i is replaced by x _ib, i.e., the following bees will not search for the neighborhood of x _i, but only the neighborhood of x _ib; the definition is as follows:

Where x _ib is the best solution selected from the square neighborhood of x _i, the weight factor phi _i,jr epsilon-1, +1,

For the ith follower bee only, x _ib will select from the square neighborhood of x _i, the specific modified search method is as follows:

Where j=1, 2, D;

for v _ib and x _ib, a better solution will be determined by:

3) Search stage of exploring bees

The new solution competes with its previous generation solution, when v _ib is better than x _ib, meaning that the search is successful in the x _ib neighborhood; when v _ib is worse than x _ib, then the neighborhood search fails; the failure is characterized by a parameter three _i, specifically the following formula:

If triali is greater than the limit parameter, x _ib is discarded and a new solution for replacement is created by:

x_i,j＝lw_j+rd(0,1)*(u_j-lw_j) (29)

Where j=1, 2, the combination of the first and second components, D, rd (0, 1) is a random number between [0,1], u _j and lw _j.

Further, the specific operation of classifying and identifying the motor fault by the radial basis function RBF classifier in the step 4 is as follows: randomly selecting 70% of data samples from the motor sampling signals as training samples, using the rest 30% of data as test samples,

1) Determining each initialization parameter: input vector X, actual output Y and expected output O, connection weight W (hidden layer-output layer), central parameter C _j (hidden layer), width vector D;

2) Algorithm iteration

Wherein W _k,j (t) is the weight of the kth iteration between the kth output neuron and the jth hidden layer neuron, C _i,j (t) is the central component of the ith hidden layer neuron on the jth input neuron in the t iteration, D _i,j (t) is the width corresponding to the center, eta is a learning factor, E is an evaluation function, and the definition is as follows:

Where o _l,k is the expected output of the kth output neuron at the ith input sample, and y _l,k is the output of the kth output neuron at the ith input sample.

The beneficial effects are that:

The invention researches a motor fault diagnosis method based on sparse decomposition (IM-HHT) and a neighborhood swarm algorithm (SS-ABC). Firstly, extracting direct current motor characteristics and signal analysis by using IM-HHT, carrying out sparse decomposition denoising treatment on motor sampling signals, carrying out denoising before treatment on the motor signals by using an orthogonal matching pursuit algorithm, selecting a plurality of optimal atoms from an overcomplete dictionary, linearly combining the selected atoms, decomposing EMD (empirical mode decomposition) by using an Huang proposed empirical mode, decomposing a given signal into a plurality of intrinsic mode functions, then removing false IMF by using correlation parameters, and carrying out Hilbert transformation to obtain Hilbert spectrum of an original signal; then, selecting the characteristics based on a clustering type Euclidean distance judging method; next, a square neighborhood selection method is utilized to replace probability selection of an original ABC algorithm, a search strategy is modified and constructed, improvement of a square neighborhood-based artificial bee colony (SS-ABC) algorithm is achieved, and sequencing of motor signal characteristic factors is completed; finally, the motor faults are classified and identified through a Radial Basis Function (RBF) classifier. The invention realizes the autonomous fault diagnosis of the motor, thereby realizing the fine diagnosis and on-line monitoring of the motor fault. The early fault diagnosis and the health state prediction of the motor are realized, so that a manager is guaranteed to discover the cause of the motor fault in time, the fault is prevented from further deteriorating, the reasonable arrangement of the maintenance time of the fault motor is guaranteed, the maintenance fund is saved, the economic loss is reduced, and the method has very important significance for improving the production efficiency and the economic benefit of enterprises.

In addition, the invention provides reference examples for practical application of motor fault diagnosis, in particular to application of a fault diagnosis method of sparse decomposition (IM-HHT) and a neighborhood swarm algorithm (SS-ABC), and compared with other traditional diagnosis methods, the method can obtain higher motor fault classification precision.

Drawings

FIG. 1 shows experimental design and implementation processes of a motor fault diagnosis method based on IM-HHT and SS-ABC;

FIG. 2 is a schematic diagram of motor dimension parameters according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a motor winding lead in accordance with an embodiment of the present invention;

FIG. 4 is an overall flow of the IM-HHT algorithm of the present invention;

FIG. 5 is a sparse denoising process of the present invention;

FIG. 6 is a flowchart of the EMD algorithm of the present invention;

FIG. 7 is a schematic diagram of IMF generated by empirical mode decomposition of BLDC Hall signals in accordance with the present invention;

FIG. 8 is a characteristic distribution diagram of the IM-HHT of the present invention;

FIG. 9 is a flow chart of a feature selection implementation of the cluster-based Euclidean distance determination method of the present invention;

FIG. 10 is a flowchart of an algorithm for selecting artificial bee colony (SS-ABC) based on square neighborhood in the invention;

Fig. 11 is a selective supervision radial basis network topology according to the present invention.

Detailed Description

The invention is further described below with reference to the accompanying drawings. The following examples are only for more clearly illustrating the technical aspects of the present invention, and are not intended to limit the scope of the present invention.

Aiming at the problem of motor faults commonly existing in manufacturing production equipment, the invention provides a motor fault diagnosis method based on sparse decomposition (IM-HHT) and a neighborhood swarm algorithm (SS-ABC). The method comprises the steps of carrying out sparse decomposition denoising treatment on a motor sampling signal, carrying out denoising treatment on the motor signal before processing by utilizing an orthogonal matching pursuit algorithm, selecting a plurality of optimal atoms from an overcomplete dictionary, carrying out linear combination on the selected atoms, decomposing EMD (empirical mode decomposition) by utilizing an empirical mode proposed by Huang, decomposing a given signal into a plurality of Intrinsic Mode Functions (IMFs), removing false IMFs by utilizing correlation parameters, and finally carrying out Hilbert transformation to obtain the Hilbert spectrum of an original signal. Then, selecting the characteristics based on a clustering type Euclidean distance judging method; next, a square neighborhood selection method is utilized to replace probability selection of an original ABC algorithm, a search strategy is modified and constructed, improvement of a square neighborhood-based artificial bee colony (SS-ABC) algorithm is achieved, and sequencing of motor signal characteristic factors is completed; finally, the motor faults are classified and identified through a Radial Basis Function (RBF) classifier.

The embodiment of the invention realizes the autonomous fault diagnosis of the motor aiming at the bearing fault, the stator winding fault and the rotor fault of the direct current motor. The bearing fault, stator winding fault, rotor fault data and normal state data of a direct current motor are used as the basis.

The invention discloses a motor fault diagnosis method based on sparse decomposition (IM-HHT) and a neighborhood swarm algorithm (SS-ABC), wherein the experimental design and the specific implementation process are shown in figure 1, and the motor fault diagnosis method comprises the following steps:

Step (1), sampling a BLDC (180W/3000 RPM/DC 24V) motor signal, analyzing by Matlab software, establishing a fault classification system, and setting the BLDC (180W/3000 RPM/DC 24V) to have the following three faults: bearing failure, stator winding failure, and rotor failure.

In the specific implementation, the servo motor generates torque opposite to that of the brushless direct current motor and is used as a load, the brushless direct current motor completes driving operation, then a data acquisition system (NI PXIe-1073) is used for acquiring Hall signals of the brushless direct current motor, the sampling frequency is 800Hz, the measuring time is 1000s, and then measuring data of the Hall signals of the brushless direct current motor can be obtained. The used motor size parameters and motor winding wiring diagrams are shown in fig. 2 and 3.

And (2) extracting the characteristics and signal analysis of the direct current motor by using the IM-HHT. The method comprises the steps of carrying out sparse decomposition denoising treatment on a motor sampling signal, carrying out denoising treatment on the motor signal before processing by using an orthogonal matching pursuit algorithm, selecting a plurality of optimal atoms from an overcomplete dictionary, carrying out linear combination on the selected atoms, then decomposing EMD (empirical mode decomposition) by using an empirical mode proposed by Huang, decomposing a given signal into a plurality of Intrinsic Mode Functions (IMFs), removing false IMFs by using correlation parameters, and finally carrying out Hilbert transformation to obtain the Hilbert spectrum of an original signal.

In step (2):

In the process of extracting the characteristics of four different types of brushless direct current motor Hall signals by utilizing an improved HHT algorithm (IM-HHT), firstly, after motor sampling signals are subjected to sparse decomposition and denoising treatment, EMD is decomposed by utilizing an empirical mode proposed by Huang, a given signal is decomposed into a plurality of Intrinsic Mode Functions (IMFs), then the false IMFs are removed by utilizing correlation parameters, and Hilbert transformation is performed to obtain the Hilbert spectrum of an original signal.

The overall flow of the IM-HHT algorithm is shown in FIG. 4.

Step (2) a: sparse decomposition denoising

(1) Characterizing signals to be decomposed and models

The signal to be decomposed is represented by the formula (1):

wherein F _i is a signal to be decomposed, f= { F ₁,f₂,...f_n } is a signal matrix to be decomposed, m is the number of selected atoms, d= { D ₁,d₂,...d_n } is an overcomplete dictionary, D _i is a selected atom, and x _i is a sparse coefficient corresponding to the selected atom. The signals can be sparsely represented by linearly combining the selected m atoms.

The core problem of sparse representation of signals is described as finding the sparsest representation value of x _i under the condition that equation (1) is satisfied, i.e., a solution with the least non-zero value is sought, model as in equation (2):

Wherein, |·| ₀ is the l ₀ norm, i.e. the number of non-zero elements in X _i, x= { X _i } is the sparse coefficient matrix, and ε is the error.

(2) Sparse decomposition

Selecting the maximum value of the inner product absolute value between the N-dimensional signal matrix F and the atom D ₁ from the overcomplete atom library D, wherein D ₁ is the optimal atom, namely, the formula (₃) is satisfied:

|<F，d₁>|＝sup|<F,d_i〉| (3)

Where, | < F, d ₁ > represents the absolute value of the inner product of the signal matrix F and the atom d ₁, d ₁ is the optimal atom to be matched for the first time.

At this time, F may be expressed as:

F＝<F，d₁>d₁+R¹F (4)

Where < F, d ₁>d₁ denotes the projection of the signal matrix F on d ₁, R ¹ F denotes the remainder of the signal matrix F after the first decomposition.

Thus, the kth decomposed signal can be expressed as:

R^kF＝<R^kF,d_k)d_k+R^k+1F (5)

the sparse denoising process is shown in fig. 5.

Note that: in the loop iteration process, orthogonality is satisfied between the selected atoms and the residual signal, and the selected atoms are removed and are not used any more.

(3) Selection of DCT dictionary

In the implementation, the DCT dictionary is obtained according to discrete cosine transform, and the real number is used for replacing the complex number by the symmetrical signal expansion to analyze on the real number domain. The one-dimensional DCT transform can be calculated by:

where J (k) is the kth DCT atom, k is the frequency factor, f (N) is a given signal sequence, and N represents the input signal length. The matrix form is represented as follows:

D＝C_Nf (10)

Step (2) b: empirical Mode Decomposition (EMD)

A given signal is decomposed into several Intrinsic Mode Functions (IMFs), which are components that meet certain conditions, by Empirical Mode Decomposition (EMD) proposed by Huang.

Let x (t) denote the original function of the input and be decomposed into n IMFs by EMD, the EMD model is defined as equation (11):

The EMD algorithm execution flow is as in fig. 6.

Step (2) c: removing false IMF using correlation parameters

In practice, the correlation parameter formula is defined as follows:

Wherein,

Correlation parameter meaning:

when r is more than 0, representing positive correlation of two variables, and when r is less than 0, the two variables are negative correlation;

when |r|=1, the two variables are completely linear correlation, namely, a functional relation;

when r=0, the radio correlation relationship between the two variables is represented;

When 0 < |r| < 1, it means that there is some degree of linear correlation between the two variables. And the closer the |r| is to 1, the closer the linear relationship between the two variables is; the closer the |r| is to 0, the weaker the linear correlation of the two variables.

Generally, the method can be divided into three stages: r < 0.4 is a low degree linear correlation; the significance correlation is that the r is more than or equal to 0.4 and less than 0.7; and the r is more than or equal to 0.7 and less than 1 is highly linear correlation. In practice, the threshold is set to r=0.5, and if r _i > 0.5, the IMF component is retained; if r _i is less than or equal to 0.5, the IMF component is determined to be spurious and removed.

Step (2) d: hilbert conversion is carried out to obtain corresponding Hilbert spectrum

Empirical Mode Decomposition (EMD), using correlation parameters, removing the spurious IMF, and introducing an Intrinsic Mode Function (IMF) into the Hilbert Transform (HT), thereby obtaining the instantaneous amplitude and instantaneous frequency of the signal. I.e. carrying out Hilbert transformation on the IMF of each motor state signal to obtain the Hilbert spectrum of the motor state signal. HT model is defined as formula (13):

In specific implementation, BLDC (180W/3000 RPM/DC 24V) motor signals are sampled, matlab software is utilized for analysis, a fault classification system is established, and the BLDC (180W/3000 RPM/DC 24V) is set to have the following three faults: bearing failure, stator winding failure, and rotor failure. Fault classification systems were investigated for their ability to classify three different types of faults. After the IM-HHT feature extraction and analysis of the measurement sampling data, the extracted features capable of reflecting the motor state are normalized, and the feature value of the motor type is ensured to be between 0 and 1, so that the gradient explosion problem in the classifier is avoided.

In specific implementation, a servo motor generates torque opposite to that of a brushless direct current motor and is used as a load, the brushless direct current motor completes driving operation, a data acquisition system (NI PXIe-1073) is used for acquiring Hall signals of the brushless direct current motor, the sampling frequency is 800Hz, the measuring time is 1000s, and then measuring data of the Hall signals of the brushless direct current motor can be obtained for researching the fault state and performance of the direct current motor. Four BLDC were tested in total, one motor being normal and the other three being faulty, the rated voltage of the brushless dc motor used was 24V, the rated rotational speed of the motor was configured to 3000RPM, and the parameters of the brushless dc motor are listed in table 1.

TABLE 1 BLDC parameters

Type(s)	Rated current	Rated for	Rated rotational speed	Rated output power	Efficiency rating
						57BL115S18	22A	12Kg-cm	3000RPM	180W	83％

The four types of motor hall signals are processed through EMD decomposition in an improved HHT algorithm (IM-HHT), the signals are decomposed into a first layer to a sixth layer (IMF 1 to IMF 6), and instantaneous amplitude and instantaneous frequency of each layer are obtained through Hilbert-Huang transform. Further, in the time domain, a maximum value (Tmax), a minimum value (Tmin), a mean value (Tmean), a mean square error (Tme), and a standard deviation (Tstd) are captured. In the frequency domain, the maximum value (Fmax), the average value (Fmean), the mean square error (Fme), and the standard value deviation (Fstd) are obtained. Each IMF extracts 10 features and normalizes them so that their feature values are distributed between 0 and 1. A total of 60 features were obtained as shown in table 2.

Table 2 HHT extracted motor signal characteristic parameters

In practice, the IMF waveform generated by empirical mode decomposition of the BLDC hall signal is shown in fig. 7. The transformation first extracts a signal high frequency, the subsequent layers of IMF are low frequency waveforms, and fig. 8 is a characteristic distribution diagram of HHT.

And (3) selecting the characteristics based on the aggregation type Euclidean distance judging method.

The feature selection stage is used for selecting few key features based on the feature selection of the clustering type Euclidean distance judging method, and in implementation, the clustering type Euclidean distance judging method is used for calculating the separability of feature categories, wherein high calculation factors represent important features. The characteristic distance discrimination factor lambda ^m is based on the same categoryEuclidean distance between features of (a) and different categoriesEuclidean distance between them. The Euclidean distance of a feature is defined by the center of the class featureAnd center of sample featureAnd (5) calculating to obtain the product. Wherein c, m, i are class number, feature number and sample number,Is a characteristic of the sample. The compensation factor eta ^m passes through the distance varianceAndThe calculation is obtained and the calculation process is as follows:

Step (3) a: calculating variances of all samples of the mth feature Average value of

Wherein,

Step (3) b: calculating variance of mth feature of class C sampleAverage value of

Wherein,

Step (3) c: calculating the weighted variance of cluster center g _c at the mth feature

Wherein,

Step (3) d: calculating the m-th feature inter-class distance d _b ^m and intra-class distance d _w ^m

Wherein,

Step (3) e: in the mth feature, the variance factor vbm of d _b ^m and the variance factor vwm of d _w ^m are calculated

Wherein,

Step (3) g: calculating the compensation coefficient eta ^m and the distance discrimination factor lambda ^m of the mth characteristic and normalizing lambda ^m′

The feature selection flow chart of the Euclidean distance judging method based on the clustering is shown in fig. 9.

And (4) replacing probability selection of an original ABC algorithm by using a square neighborhood selection method, modifying and constructing a search strategy, realizing improvement of a square neighborhood-based artificial bee colony (SS-ABC) algorithm, and completing sequencing of motor signal characteristic factors.

After the feature selection, the selected feature factors are arranged in a descending order by utilizing an improved Artificial Bee Colony (ABC) algorithm, and the feature of some Hall signals can reduce the recognition rate of the classifier or the recognition result does not influence, so that the feature factors can be deleted through the feature selection, and the calculation time is saved. In combination with the Artificial Bee Colony (ABC) algorithm, computational efficiency or recognition rate may be improved. The invention sets the identification result as the fitness value, inputs the identification result as the feature level, outputs the identification result as the new feature level, sorts the features, sorts the optimized features through the ABC algorithm, and can improve the performance identity of the features.

An Artificial Bee Colony (ABC) algorithm is a powerful tool problem for solving an optimization problem, belongs to branches of artificial intelligence, and simulates the behavior of the bee colony in the development and discarding process of searching food sources by the ABC algorithm. The location of the food sources indicates a possible way to solve the problem, in order to find the best solution, the artificial bees interact and exchange information, focus on the desired large solution interval, and leave the desired smaller solution area steadily by using collective knowledge, thereby collectively improving the solution step by step, repeating the search through the algorithm until the predetermined stop condition is met.

In the ABC algorithm, bees are divided into three distinct populations: employ bees, follow bees, and spy bees. The employment bees master the original honey source information and send the information to the surrounding bees through the swing; the following bees choose to follow and collect honey according to own judgment, the scout bees judge whether to discard old honey source information to search for a new replacement, the scout bees search for surrounding environments, each time find a new honey source, namely store the new honey source in a memory, if the new honey source is better than the previous honey source, the bees can memorize the new situation and forget the previous situation. These three phases are repeatedly performed until an optimal solution is obtained.

Step (4) a: employment of bee search phases

Let X _i＝{X_i,1,X_i,2,...,X_i,D be the i-th solution in the group, N be the total number of groups, D be the dimension, then hire bees to search around each solution X _i (i=1, 2..once., N), try to find the optimal solution, see formula (21)

Where x _k is randomly drawn from the whole population, jr is a random integer within [1, D ], φ _i,jr is randomly generated at [ -1,1], whereby a new solution v _i is generated as in (22):

Where j=1, 2, D;

next, a better solution is obtained using a greedy selection method, as shown in equation (23).

Step (4) b: following the bee search phase

At this stage, the employing bees complete the neighborhood search for all solutions, and the following bees obtain search information from the employing bees. Unlike employment of bees, at this stage, following bees only pick a portion of better solutions for further searching. In the invention, an improved ABC algorithm method based on neighborhood selection is provided, namely, square neighborhood selection artificial bee colony (SS-ABC). The main improvement is that the square neighborhood selection method is used for replacing the original probability selection method, and the search strategy is modified and constructed based on the square neighborhood.

(1) Square neighborhood selection

And setting the group solution as square topology by utilizing the concept of square neighborhood. And judging whether the data point is in the neighborhood or not according to the formula (24):

wherein x _i and y _i are two square neighborhood points, sl is the square neighborhood side length, and m is the neighborhood boundary coefficient.

When x _i and y _i meet the formula (24), namely the data points in the neighborhood are determined, the algorithm does not need to carry out multiplication operation and evolution operation, and the clustering efficiency can be improved.

(2) Modifying search strategies

Next, the build search strategy is modified based on the square neighborhood side length. A following bee search scheme selection mechanism is proposed, for each solution x _i, selecting the best solution in the square neighborhood of x _i as x _ib, and then searching around x _ib using the corresponding search strategy. In contrast to the traditional ABC algorithm, this approach does not require calculation of the probability of selection for each solution.

During the search, x _i is replaced by x _ib, i.e., the following bees will not search for the neighborhood of x _i, but only the neighborhood of x _ib. Definition as the formula (25)

where j=1, 2,..d.

For v _ib and x _ib, a better solution will be determined by equation (28):

step (4) c: search stage of exploring bees

The new solution competes with its previous generation solution. When v _i is better than x _i, this means that the search was successful in the x _i neighborhood. When v _i is worse than x _i, then it indicates that the neighborhood search failed. The failure count is characterized by a parameter, three _i, when the search is successful, three _i =0, and when the search is failed, three _i is incremented by one, see equation (29):

If the three _i is greater than the limit parameter, the corresponding solution x _i is discarded and a new solution for replacement is created by equation (30):

x_i,j＝lw_j+rd(0,1)*(u_j-lw_j) (30)

The flow of the artificial bee colony (SS-ABC) algorithm for square neighborhood based selection is shown in figure 10.

And (5) classifying and identifying motor faults by using a Radial Basis Function (RBF) classifier.

Building Radial Basis (RBF) classifier

And classifying and identifying motor faults through a Radial Basis Function (RBF) classifier. From the motor sample signal, 70% of the data samples were randomly selected as training samples, and the remaining 30% of the data were used as test samples. The radial basis (Radial Basis Function, RBF) network is typically a three-layer structure, with a selective supervision radial basis network topology as shown in fig. 11.

The nodes of the input layer and the output layer of the radial basis neural network are both linear functions, and the hidden layer is a radial basis function and has local approximation capability. In the iteration process of the RBF algorithm, a gradient descent method is adopted for self-adaptively adjusting the center, the width and the weight, and the method specifically comprises the following steps:

Step (5) a: determining an input vector X

X＝{x₁,x₂,…x_n}^T (31)

Where n is the number of input neurons.

Step (5) b: determining the actual output Y and the desired output O

Y＝{y₁,y₂,…y_q}^T,O＝{o₁,o₂,…o_q}^T (32)

Wherein q is the number of output neurons.

Step (5) c: initializing connection weights W (hidden layer-output layer)

W_k＝{w_k,1,w_k,2,…w_k,p}^T (33)

Wherein p is the number of hidden layer neurons.

Step (5) d: initializing central parameter C _j (hidden layer)

C_j＝{c_j,1,c_j,2,…c_j,n}^T (34)

Step (5) e: initializing width vector D

D_i＝{d_i,1,d_i,2,…d_i,n}^T

Wherein d is an adjustment coefficient

Step (5) f: iterative algorithm until the error condition is met or the maximum iterative times are reached

The foregoing embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the present invention and to implement the same, not to limit the scope of the present invention. All equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of the present invention.

Claims

1. A motor fault diagnosis method based on sparse decomposition and neighborhood swarm algorithm is characterized by comprising the following steps:

the specific operation of denoising before processing the motor signal is as follows:

(1) Characterizing signals to be decomposed and models

Wherein F _i is a signal to be decomposed, F= { F ₁,f₂,...f_n } is a signal matrix to be decomposed, m is the number of selected atoms, D= { D ₁,d₂,...d_n } is an overcomplete dictionary, D _i is a selected atom, x _i is a sparse coefficient corresponding to the selected atom, and the signals can be sparsely represented by linearly combining the selected m atoms;

(2) Sparse decomposition

|<F,d₁>|＝sup|<F,d_i>| (3)

the choice of the overcomplete dictionary DCT is as follows:

D＝C_Nf (7)

C _N is a DCT coefficient matrix, and then the DCT coefficient matrix is finely sampled in a frequency domain to obtain an overcomplete dictionary;

Decomposing EMD by using an empirical mode, decomposing a given signal into a plurality of intrinsic mode functions IMFs, removing false IMFs by using correlation parameters, and finally performing Hilbert transformation to obtain a Hilbert spectrum of an original signal, wherein the specific operation comprises the following steps:

2) Defining correlation parameters and removing false IMF

Wherein: ；

Where H _l (t) represents IMF, H _l (t) is obtainable by hilbert transform, a _l (t) is instantaneous amplitude, θ _l (t) is instantaneous phase angle;

2. The motor fault diagnosis method based on sparse decomposition and neighborhood swarm algorithm according to claim 1, wherein the specific operation of finding the clustering-based euclidean distance determination method for feature selection in step 2 comprises the following steps:

1) Calculating variances of all samples of the mth feature Average value of

Wherein,

2) Calculating variance of mth feature of class C sampleAverage value of

Wherein,

3) Calculating the weighted variance of cluster center g _c at the mth feature

Wherein,

6) Calculating a compensation coefficient eta ^m of the m-th feature:

7) Calculating a distance discrimination factor lambda ^m of the mth feature, and obtaining normalized lambda ^m':

。

3. The motor fault diagnosis method based on sparse decomposition and neighborhood swarm algorithm according to claim 1, wherein the specific steps of selecting artificial swarm SS-ABC algorithm based on square neighborhood in the step 3 to order the motor signal characteristic factors are as follows:

1) Employment of bee search phases

the new solution v _i generates the following:

2) Following the bee search phase

where j=1, 2, D;

for v _ib and x _ib, a better solution will be determined by:

3) Search stage of exploring bees

if the three _i is greater than the limit parameter, x _ib is discarded and a new solution for replacement is created by:

x_i,j＝lw_j+rd(0,1)*(u_j-lw_j) (29)

4. The motor fault diagnosis method based on the sparse decomposition and neighborhood swarm algorithm according to claim 1, wherein the specific operation of classifying and identifying the motor fault by the radial basis function RBF classifier in the step 4 is as follows: randomly selecting 70% of data samples from the motor sampling signals as training samples, using the rest 30% of data as test samples,

1) Determining each initialization parameter: input vector X, actual output Y and expected output O, connection weight W of hidden layer-output layer, central parameter C _j of hidden layer, width vector D;

2) Algorithm iteration