CN116361727A

CN116361727A - Audio feature and SRC-Adaboost-based battery power conversion system driving gear fault diagnosis method

Info

Publication number: CN116361727A
Application number: CN202310314068.6A
Authority: CN
Inventors: 游祥龙; 胡晓松; 游肖文; 周时国; 郭怡斐; 李佳承; 赵宇斌; 刘文学
Original assignee: Chongqing University
Current assignee: Chongqing University
Priority date: 2023-03-28
Filing date: 2023-03-28
Publication date: 2023-06-30

Abstract

The invention belongs to the field of fault diagnosis for guaranteeing high-strength transportation means, and particularly relates to a fault diagnosis method for a battery power conversion system driving gear based on audio characteristics and SRC-AdaBoost. The invention adopts axial, horizontal and vertical vibration signals to carry out fault diagnosis. The fault diagnosis method comprises the steps of audio feature extraction, robust dictionary generation, sparse representation, synthesis and pattern recognition. The audio features used in the present invention are MFCC and GTCC, which are sensitive to low frequency components of the vibration signal; then, the audio features form an original dictionary, and before sparse representation, data enhancement and dictionary learning are carried out to improve the accuracy of sparse representation; finally, in order to improve the robustness under the speed fluctuation condition, the sparse vectors are integrated based on AdaBoost, and then fault diagnosis is carried out according to the weight distribution of the final sparse vector.

Description

Audio feature and SRC-Adaboost-based battery power conversion system driving gear fault diagnosis method

Technical Field

The invention belongs to the field of fault diagnosis for guaranteeing high-strength transportation means, and particularly relates to a fault diagnosis method for a battery power conversion system driving gear based on audio characteristics and SRC-AdaBoost.

Background

Transport electrification is a process that occurs worldwide, and for commercial transport, electric heavy trucks are recently replacing natural gas-fueled heavy trucks, becoming the most commonly used heavy trucks. In some situations, such as special line transportation, short-distance transportation of branch lines and port transportation, the electric heavy truck needs to work continuously between two or more points and cannot be stopped for charging for a long time, so that in these high-intensity works, the battery power conversion scheme is widely used, and the advantage is obvious. Since EHTs (electric heavy trucks ) are always much larger than electric passenger cars, power battery systems always contain more than one replaceable battery pack, which is placed on the side of the EHT. Based on this situation, a commonly adopted power conversion scheme is that an empty battery pack is pulled out by a BSS, a charged battery pack is pushed in by the BSS, a motion platform of the BSS (battery swapping system, battery power conversion system) is driven by a driving system, and the driving system is deployed inside the motion platform.

The driving system consists of a servo motor, a transmission system, a driving gear, an intermediate gear and a frame. According to the historical maintenance data, the failure rate of the drive gear and the intermediate gear is higher than that of other components, which is caused by stress concentration and electric erosion. Therefore, in the present study, a fault diagnosis method of a drive gear was developed, and several common faults including single-sided tooth wear (UTW), double-sided tooth wear (BTW), and Tooth Breakage (TB) were injected into the drive gear.

Based on vibration signal processing, most of the research in existence focuses on gear fault diagnosis in planetary gearboxes. In order to extract the weak features caused by the low-speed state, many improvements based on time-frequency analysis methods, such as adjustable factor wavelet transform, mutation pattern decomposition, ensemble empirical pattern decomposition, local average decomposition and the like, have been proposed. However, due to the reciprocating motion of the motion platform, the continuous monitoring data of the driving gear in the BSS is much shorter than the gears in the planetary gearbox, which weakens the effectiveness of the above method, while the stress conditions of the gears in the BSS are different from those in the planetary gearbox. Other gear fault diagnosis methods in low speed state are based on monitoring data obtained from non-vibrating sensors such as rotary encoders, optical encoders and acoustic emissions, however, due to the compact structure of BSS, it is difficult to deploy these sensors.

For signals with strong noise, the previous fault diagnosis method proposes different solutions in the preprocessing stage, the feature extraction stage and the pattern recognition stage. During the preprocessing stage, various filters are designed for signal denoising, such as conventional bandpass filters, or adaptive filters. In the feature extraction stage, features insensitive to noise, such as conjugate geometric package decomposition, minimum entropy deconvolution and wavelet-related feature scale entropy are extracted. In the pattern recognition phase, machine learning methods are always employed, including data enhancement strategies to improve noise immunity, for example, note that residual prototype networks, stacked denoising auto-encoders, and gated recursive unit neural networks are used for fault diagnosis in noisy environments. However, for uncertain noise and mixed noise in vibration signals caused by interference of electromagnetic environment and coupling system, the conventional noise reduction method is difficult to remove all noise due to complex frequency distribution, so that improving noise immunity is a key problem of gear fault diagnosis in BSS.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a battery power conversion system driving gear fault diagnosis method based on audio characteristics and SRC-AdaBoost, so as to improve noise immunity of driving gear fault diagnosis in the battery power conversion system.

In order to achieve the above purpose, the technical scheme adopted by the invention is as follows.

A battery power conversion system driving gear fault diagnosis method based on audio characteristics and SRC-AdaBoost comprises the following steps:

s1 Audio feature extraction

Collecting multidirectional vibration signals of a sample to be diagnosed, and respectively extracting the characteristic of the mel cepstrum coefficient and the characteristic of the Gammatone cepstrum coefficient as audio characteristics for the multidirectional vibration signals;

s2 robust dictionary generation

Constructing a robust dictionary matrix by utilizing the extracted audio features in the step S1, constructing an original dictionary matrix by utilizing the extracted audio features, performing sample amplification on each column vector in the original dictionary matrix by utilizing a data enhancement method, forming an extended dictionary matrix by utilizing the amplified column vectors, and finally learning from the redundant extended dictionary matrix by utilizing a dictionary learning method to generate the robust dictionary matrix;

s3, fault diagnosis of driving gear is realized through integrated sparse expression

Based on the robust dictionary matrix generated in the step S2, calculating sparse vectors corresponding to the Mel cepstrum coefficient features and the Gamma cepstrum coefficient features extracted in the step S1 by using a sparse expression method; then, acquiring weight values corresponding to the sparse vectors by using an ensemble learning method; and finally, acquiring a final sparse vector by adopting a weighted summation mode, and realizing fault diagnosis of the driving gear by utilizing the distribution of non-zero elements in the final sparse vector.

Further, in step S1, the audio feature extraction specifically includes the following steps:

s11, cutting the vibration signals into frames, and performing short-time Fourier transform on each frame of signals to obtain a frequency spectrum y (k) of the frame of signals;

s12, filtering the frequency spectrum y (k) through a Mel triangular filter bank and a Gamma filter bank respectively to obtain a logarithmic spectrum S (m) based on the Mel triangular filter and a logarithmic spectrum T (m) based on the Gamma filter bank;

s13, processing the spectrums S (M) and T (M) through discrete cosine transform to obtain MFCC characteristics and GTCC characteristics of each frame of vibration, and representing the MFCC characteristics in a matrix form, wherein the MFCC characteristics matrix is expressed as M _MFCC The GTCC feature matrix is denoted as M _GTCC The number of rows of the matrix is m ₁ The number of columns is n ₁ 。

Further, the MFCC feature matrix M _MFCC And GTCC feature matrix M _GTCC The composed label sample is split into two training sample data sets, the first training sample data set is used for dictionary learning and the second training sample data set is used to obtain the weights of the AdaBoost algorithm.

Further, in step S2, the robust dictionary generation specifically includes the following steps:

s21, the audio feature matrix M _MFCC And M _GTCC The method is converted into an audio feature vector v, the feature vector v forms a sub-original dictionary, the number of modes is p, the number of label samples of each mode is q, the p sub-original dictionaries form an original dictionary OD, and the matrix size of the original dictionary OD is m ₁ n ₁ ×pq；

S22, carrying out random Cutout data enhancement operation on each label sample in the original dictionary OD, and expanding one label sample into n _E The labels and the enhanced samples to obtain an extended dictionary ED, wherein the matrix size of the extended dictionary ED is m ₁ n ₁ ×pn _E q；

S23, performing dictionary learning by using an OMP algorithm to obtain an optimized sub-dictionary RD based on the sub-dictionary of the extended dictionary ED _s Updating and optimizing the sub-dictionary RD by a K-SVD algorithm _s Obtaining a studyRear sub dictionary RD _s Namely sub-dictionary of robust dictionary RD, the sub-dictionary RD after learning is utilized _s Constructing a robust dictionary RD, wherein the matrix size of the robust dictionary RD is m ₁ n ₁ ×pq。

Further, in step S3, the implementation of the fault diagnosis of the driving gear by integrating sparse representation specifically includes the following steps:

s31, using the robust dictionary RD as a dictionary matrix of the sparse representation, extracting and obtaining sparse vectors

Sparse vector->

The size of (2) is pq×1;

s32 from sparse vectors

Extracting a distribution vector, and calculating a single-mode sparse vector element and beta _i Maximum beta _i The index of (a) is the mode of the test sample, and the result of mode identification is expressed as a mode vector P;

s33, combining a group of sparse vectors calculated based on audio features by using an AdaBoost algorithm to obtain the weight w of the AdaBoost algorithm _u By weight w _u Calculating to obtain final sparse vector

And according to the final sparse vector

And obtaining a final test sample mode, and judging the fault type of the driving gear.

Further, in step S2, the specific steps of constructing the original dictionary matrix are as follows: for single audio features of a single-direction vibration signal, selecting audio features with known fault modes to form column vectors, wherein each mode has the same number of column vectors, and arranging the column vectors in sequence transversely according to the modes to form an original dictionary matrix.

Further, in step S2, the specific step of performing sample amplification on each column vector in the original dictionary matrix is: the method of Cutout data enhancement is used to expand 1 column vector into a number of column vectors of a certain number.

Further, in step S2, the dictionary learning method includes: and performing dictionary learning by using a K-SVD method, and enabling the number of column vectors in the learned dictionary matrix to be the same as the number of column vectors in the original dictionary matrix.

Further, in step S3, the method for calculating the sparse vector includes: and calculating sparse vectors of the audio features under the expression condition of the corresponding robust dictionary matrix by using the specific audio features of the sample to be diagnosed in the specific direction and the corresponding robust dictionary matrix and adopting a batch matching tracking method.

Furthermore, when the batch matching tracking method is adopted to calculate the sparse vector, the least square method can be utilized to iterate and obtain non-zero elements in the sparse vector, and a plurality of non-zero elements with determined quantity can be calculated at the same time in each iteration.

Further, in step S3, the method for obtaining the weight value corresponding to each sparse vector includes: and acquiring weights of sparse vectors corresponding to the audio features in all directions by using training samples with known fault modes and adopting an AdaBoost method.

The vibration signal is collected by three mutually perpendicular vibration sensors, and the vibration signal comprises an axial vibration sensor, a horizontal vibration sensor and a vertical vibration sensor, meanwhile, a Hall speed sensor is deployed above the driving gear to collect the speed signal, and vibration data and speed data are synchronously collected by the data collecting case.

The beneficial effects are that:

(1) The invention provides a fault diagnosis scheme of a driving gear in a BSS based on audio characteristics and SRC-AdaBoost, which can realize fault diagnosis of the driving gear under complex conditions of reciprocating motion, low speed and fluctuation, complex noise interference and the like in the BSS and can improve maintenance efficiency.

(2) Aiming at the low-speed operation working condition of the power conversion system, a characteristic extraction method combining MFCC (Mel-frequency cepstral coefficients, mel frequency cepstrum coefficient) and GTCC (Gammatone cepstral coefficients, gamma-pass cepstrum coefficient) for a driving gear of the power conversion system is provided, so that the significance of extracted fault characteristics under the low-speed condition can be improved.

(3) The robust dictionary generation method based on data enhancement and dictionary learning improves the robustness of SRC (sparse representation-based classification, classification based on sparse expression) and the sparsity of sparse vectors.

(4) Aiming at the complex working conditions of rotation speed fluctuation and noise interference of a power conversion system, a classification algorithm combining sparse expression and AdaBoost is provided, and the accuracy of fault diagnosis is improved.

Drawings

Fig. 1 is a main flow chart of the present invention.

Fig. 2 is a flowchart of audio feature extraction in an embodiment of the invention.

FIG. 3 is a diagram of a vibration data framework in an embodiment of the present invention.

Fig. 4 is a flowchart of robust dictionary generation in an embodiment of the present invention.

Fig. 5 is a flowchart of the feature matrix conversion in the embodiment of the present invention.

Fig. 6 is a data Cutout and expansion flow chart in an embodiment of the invention.

FIG. 7 is a flow chart of pattern recognition based on a single sparse vector in an embodiment of the invention.

FIG. 8 is a diagram of training sample data set assignments in accordance with an embodiment of the present invention.

Fig. 9 is a schematic diagram of fault classification of a driving gear according to an embodiment of the invention.

FIG. 10 is a diagram showing the result of extracting MFCC features in the embodiment of the present invention.

Fig. 11 is a schematic diagram of a result of extracting GTCC features in an embodiment of the present invention.

Fig. 12 is a comparison diagram of audio feature extraction of the Mel-triangle filter bank and the gammatine filter bank in the embodiment of the present invention.

FIG. 13 is a diagram of an original dictionary based on MFCC in an embodiment of the present invention.

Fig. 14 is a schematic diagram of an original dictionary based on GTCC in an embodiment of the present invention.

FIG. 15 is a schematic diagram of an extended dictionary based on MFCC in an embodiment of the present invention.

Fig. 16 is a schematic diagram of a GTCC-based expansion dictionary in an embodiment of the present invention.

Fig. 17 is a schematic diagram of a robust dictionary based on MFCC in an embodiment of the present invention.

Fig. 18 is a schematic diagram of a robust dictionary based on GTCC in an embodiment of the present invention.

Fig. 19 is a diagram showing the result of the MFCC-based SRC for training data set II in the embodiment of the present invention.

Fig. 20 is a diagram showing the result of the SRC based on GTCC for training data set II in the embodiment of the present invention.

FIG. 21 is a schematic representation of the variables in AdaBoost in an embodiment of the invention.

FIG. 22 is a diagram showing the sparse vector calculation results of test samples according to an embodiment of the present invention.

Fig. 23 is a view showing final sparse vectors and fault diagnosis results in an embodiment of the present invention.

Detailed Description

In order that the above-recited objects, features and advantages of the present invention will be more clearly understood, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced in other ways than those described herein, and therefore the present invention is not limited to the specific embodiments disclosed below.

The invention provides a fault diagnosis method for a driving gear of an electric heavy truck power conversion system based on audio features and integrated sparse representation (SRC-Adaboost), wherein the electric heavy truck is an electric heavy truck with a power battery pack for replacement; the power exchanging system is used for moving the battery pack by utilizing a gear rack structure so as to realize replacement; the driving gear is a driving gear in a gear rack driving structure; aiming at the conditions of low rotation speed, rotation speed fluctuation and noise interference of a driving gear, the method firstly extracts audio characteristics by utilizing multidirectional vibration signals; then constructing a robust dictionary matrix by utilizing the extracted audio features; finally, through integrated sparse expression, fault diagnosis of the driving gear is realized.

According to the invention, fault diagnosis is carried out by adopting axial, horizontal and vertical vibration signals, namely, IEPE type acceleration vibration sensors are respectively arranged on a rotating shaft supporting mechanism connected with a driving gear in the horizontal direction, the vertical direction and the axial direction, and the acceleration type vibration signals of the driving gear are obtained through an IEPE type data acquisition system; the drive gear monitoring data includes noise caused by vibration interference of the coupling system and electromagnetic interference. The fault diagnosis scheme provided by the invention comprises the steps of audio feature extraction, robust dictionary generation, sparse expression, integration and pattern recognition. The audio features employed in the present invention are MFCC (mel-frequency cepstral coefficient) and GTCC (Gammatone-frequency cepstral coefficient), which are sensitive to low frequency components of vibration signals. Then, an original dictionary is constructed by using the audio features, and data enhancement and dictionary learning are performed before sparse representation to improve the accuracy of sparse representation. Finally, in order to improve the robustness under the speed fluctuation condition, the sparse vectors are integrated based on AdaBoost, and then fault diagnosis is carried out according to the weight distribution of the final sparse vector.

1. Audio feature extraction

MFCCs and GTCCs are classical audio features that are widely used for voiceprint recognition and automatic speech recognition. They are also used for fault diagnosis due to their advantages in non-stationary signal processing and sensitivity in low frequency component analysis. It has been demonstrated that audio features such as MFCC can reduce noise in vibration signals, extract more useful low frequency fault features for subsequent pattern recognition, and exhibit better performance in bearing and gearbox fault diagnosis problems.

Thus, to track the harmonic frequencies associated with drive gear failure, the present invention employs both MFCC and GTCC audio features. Based on the filter bank design, the MFCC is more sensitive to low frequency components of the vibration signal, while the GTCC is more stable when processing noisy signals. The synergy between MFCC and GTCC provides an effective feature for diagnosis of drive gears in battery-powered drive systems. Both MFCC and GTCC are cepstral coefficients, which are calculated based on the Mel-triangle filter bank and the gammatine filter bank, respectively, and can be regarded as energies of different frequency components in the cepstral. The audio feature extraction flow chart shown in fig. 2 shows the extraction process of MFCCs and GTCCs. In this embodiment, the audio feature extraction includes the following steps.

1.1 Framing and windowing

In order to perform a Short Time Fourier Transform (STFT) on the vibration signal, the signal needs to be cut into frames. In this embodiment, a hamming window (hamming window function) is used to construct the vibration signal, and the coefficients of the hamming window are generated by:

in the formula (1), w (N) is a Hamming window coefficient, N is a Hamming window coefficient serial number, and N is a Hamming window width;

and the length of the window is L _F N of window =n ^th The value is w (n), where n ^th Is the nth coefficient value in the hamming window.

Then, the vibration signal is divided by the overlapped sliding window based on the Hamming window, as shown in FIG. 3, the window length, i.e., the frame length is L _F Frame shift to L _S 。

Here, in order to preserve the characteristic frequency of the STFT, the frame length and frame shift are determined by the rotational speed variation of the drive gear. Assuming that the rotation speed of the driving gear is omega RPM, omega E [ omega ] _min ,ω _max ]The number of teeth of the driving gear is n _G The sampling rate of the vibration data is f _S Then:

1.2 fast Fourier transform

Then, the vibration signal of each frame is converted into a frequency domain through short-time Fourier transform (STFT), so that a frequency spectrum y (k) of a framing signal is obtained, and the calculation formula is as follows:

in the equation (4), x (n) is a vibration signal frame, w (n) is a coefficient of a hamming window, and k is a coefficient number of a frequency spectrum after fourier transform.

1.3 Mel Filter bank based Filtering

Since the human auditory system is more sensitive to low frequency components, to simulate this perceptual feature, the frequency spectrum on the hertz Scale is converted to Mel-Scale:

where M (f) is the Mel frequency and f is the frequency on the Hertz scale.

The spectrum is then filtered by a Mel-triangle filter bank, which is assumed to contain n ₁ The mth Mel filter is denoted as:

in the formula (6), f (m) is the center frequency of the Mel triangular filter, m is the filter serial number in the Mel filter bank, H _m (k) Is an expression for the mth filter.

The logarithmic energy for each filter can then be calculated as:

where S (m) is the logarithmic spectrum and y (k) is the spectrum of the framed signal obtained by equation (4).

1.4 Gamma filter bank based filtering

Unlike the Mel triangle filter bank, the gammatine filter bank is composed of curves, which inspire the cognitive psychology and physiology observations of the human auditory system, and the filter bank is a standard cochlear implant filter model. The mth gammatine filter is obtained according to the following formula:

where a is the amplitude, b defines the bandwidth, f _C Is the center frequency of the filter and phi is the phase. The definition of bandwidth b is:

b＝1.019×24.7×(4.37f _C +1) (9)

then, the logarithmic energy of the gammatine filter is obtained:

where T (m) is the logarithmic spectrum and y (k) is the spectrum of the vibration signal.

1.5 Discrete cosine transform

MFCC and GTCC are obtained by applying discrete cosine transform to the logarithmic spectrum S (m) and the logarithmic spectrum T (m). The signal mfcc (q) _t ) The MFCC of (c) is calculated as follows:

similarly, the GTCC is calculated according to the following formula:

wherein q _t Is the order of MFCC and GTCC eigenvaluesNumber (x).

1.6 Characteristics of MFCC and GTCC

Based on the above procedure, the MFCC characteristic and the GTCC characteristic of each frame of vibration are obtained and expressed in a matrix form. Thereafter, the MFCC feature is denoted as M _MFCC GTCC features are denoted as M _GTCC The number of rows of the matrix is m ₁ ：

Where L is the length of the vibration data sample. The number of columns of the matrix is the same as the number of filters in the filter bank and is denoted as n ₁ . Thus M _MFCC And M _GTCC Is all m in size ₁ ×n ₁ 。

2. Robust dictionary generation

The audio features extracted from the vibration signals in all directions are used to construct an original dictionary. However, in the present embodiment, in order to improve the robustness of the failure diagnosis in consideration of the speed fluctuation of the drive gear and the mixed disturbance caused by the coupling system and the severe electromagnetic environment, a robust dictionary is generated on the basis of data enhancement and dictionary learning. Fig. 4 shows the generation process of the robust dictionary.

2.1 Original dictionary structure

The size of the audio feature matrix is m ₁ ×n ₁ ，m ₁ Is the number of frames, n ₁ The number of columns. The matrix of audio features is converted into feature vectors prior to constructing the original dictionary. It is assumed that the audio feature matrix extracted from the marked data samples is:

then, the audio feature vector v converted from the audio feature matrix M is:

thus, v has a size of m ₁ n ₁ X 1, the conversion process is shown in fig. 5.

The number of patterns in the present invention is p, and assuming that the number of label samples in the original dictionary OD is q for each pattern, the matrix size of the original dictionary OD is m ₁ n ₁ X pq, as shown in the upper part of fig. 4:

OD＝(OD ₁ OD ₂ … OD _p ) (16)

the original dictionary is composed of p sub-original dictionaries, each of which is composed of feature vectors extracted from a label sample of a specific pattern. The sub-original dictionary is denoted as OD _s S=1, 2, …, p, and OD _s Has a size of m ₁ n ₁ ×q。

2.2 Cutout-based data enhancement

Data enhancement strategies are widely used in machine learning, particularly in image recognition, which can significantly improve the performance of image classification. In order to deal with speed fluctuations and complex disturbances, the present invention employs data enhancement techniques to improve the performance of fault diagnostics.

The most advanced data enhancement strategies mainly include Cutout, mixup, cutMix, and the like. Considering that the details of the MFCC and the GTCC are affected by speed fluctuations, the present invention employs a Cutout strategy in order to improve robustness. Cutout is a data enhancement strategy that deletes random elements in feature vectors and repopulates them with zero or random values. At the same time, different Cutout will produce different new samples, and therefore the number of marked samples will also be enlarged.

The Cutout operation applies to each column of the audio feature matrix, i.e., each segment in the audio feature vector v. For each segment, randomly choose n _C The elements, note n _C ＜＜m ₁ Selected n _C The elements are deleted and then the elements are refilled with μ _r,s ,r＝1,2,…,n ₁ S=1, 2, …, p, for each column in the audio feature matrix M, or each segment in the audio feature vector v, element μ _r,s The calculation mode of (2) is as follows:

in the formula (17), r is a column number in the audio feature matrix formula (14), s is a mode number in p modes, i is a row number in the audio feature matrix formula (14), j is a sample number of q samples in each mode, a _ir Is the (r.m) th in the audio feature vector matrix formula (15) ₁ + i) elements.

FIG. 6 shows the element removal and refilling process, in this embodiment, n _C =2. As shown in fig. 6, n is performed on each label sample in the original dictionary OD _E After the sub-random Cutout data enhancement operation, a label sample is expanded to n _E The labels and the enhanced samples, an extended dictionary ED is obtained. At the same time, the number of marked samples of each pattern in the extended dictionary ED is n _E q, ED has a size of m ₁ n ₁ ×pn _E q, as shown in the middle part of fig. 4:

ED＝(ED ₁ ED ₂ … ED _p ) (18)

here, the sub-expansion dictionary is denoted ED _s S=1, 2, …, p, s is the sub-dictionary number, sub-expansion dictionary ED _s Is the s < th > sub-original dictionary OD _s Extended ED of (D) _s Has a size of m ₁ n ₁ ×n _E q。

2.3 Dictionary learning based on K-SVD algorithm

K-means singular value decomposition (K Singular Value Decomposition, K-SVD) is a classical dictionary learning method that provides a suitable dictionary for sparsity reduction in sparse representation, in other words, samples can be better represented by atoms learned using the K-SVD method. In the present invention, we define a dictionary that learns from the extended dictionary mode by mode as a robust dictionary, as shown in the lower half of fig. 4. Note that in the present invention, intra-class dictionary learning is employed, in other words, dictionary learning is performed in a sub-expansion dictionary ED _s S=1, 2, …, p.

The dictionary learning process is an optimization process, and the learning process is as follows:

where Y is the sample set, RD _s Is the optimized sub-dictionary output after dictionary learning, X is coefficient matrix, and X _i Is the column vector in X, i is the column vector number in matrix X, T ₀ Is the number of non-zero elements required in the learning process.

The invention firstly utilizes an orthogonal matching pursuit (Orthogonal Matching Pursuit, OMP) algorithm to obtain an initial sub-expansion dictionary ED _s The corresponding sparse vector set X is then updated and optimized sub-dictionary RD by adopting a K-SVD algorithm _s And a sparse vector set X. Optimizing sub-dictionary RD _s During the atomic updating of (a), the sub-dictionary RD is optimized _s Each atom in (c) is updated in turn and is denoted as d _k At the same time, the sparse vector set X corresponds to d _k Is calculated, note that these elements constitute the kth row vector in the sparse vector set X

Instead of column vectors, the function is:

in the formula (20), d _j Is RD _s The (j) th atom of (a) is,

is the j-th row vector in X.

Wherein E is _k Is formed by optimizing sub-dictionary RD _s The sub-dictionary RD is optimized by an error matrix generated in kth column _s Is updated by an error matrix E _k Using K-means singular value decomposition (K-SVD):

after that, we will ω _k Defined as a set of sequence numbers corresponding to those

Positions that are non-zero elements:

in the formula (22), K is the number of non-zero elements in the sparse vector, namely the iteration times; i is a sequence number.

Then, using matrix Ω _k To solve for an error matrix, at matrix Ω _k Middle, (omega) _k (i) I) th is 1, elsewhere 0, matrix Ω _k The size of (2) is Nx|omega _k | a. The invention relates to a method for producing a fibre-reinforced plastic composite. Then multiply Ω by equation (20) _k ：

For a pair of

By K-SVD decomposition, here, < >>

Is of the order of omega _k The error matrix after multiplication, we get +.>

Thereafter, we replace d with the first column of U _k . Likewise, sub-dictionary RD is optimized _s Atom d of (B) _k K=1, 2, …, K is updated column by column, finally we get a learned sub-dictionary matrix RD _s S=1, 2, …, p, namely the sub-dictionary of the robust dictionary RD, the number of atoms in the robust dictionary RD is the number of iterations K, in the present invention:

K＝q (24)

as shown in the lower part of fig. 4.

Then, the learned sub-dictionary RD is utilized _s Building a robust dictionary RD:

RD＝(RD ₁ RD ₂ … RD _p ) (25)

RD has a size of m ₁ n ₁ ×pq。

SRC-AdaBoost and Fault diagnosis

3.1 Sparse representation based on batch matching pursuit

SRC is a pattern recognition method developed from compressed sensing. In the SRC framework, the test samples y are represented by a pre-built dictionary matrix D, which is composed of training samples grouped by pattern:

in the formula (26), v _o Is a sample of the labels (i.e., atoms in dictionary matrix D), D _s S=1, 2, …, p is a sub-dictionary of the dictionary matrix D. Then, a sparsely expressed sparse vector is obtained, expressed as:

theoretically, sparse vectors

The non-zero elements in (a) are contributions of corresponding atoms in the dictionary matrix D when expressed as test samples y, the larger the elements means the higher the similarity, i.e. the greater the contributions. Thus, the exponents and values of non-zero elements in the sparse vector may be used for pattern recognition.

For each test sample, the audio characteristics are transformed based on the formulas (14), (15) and (5), and the characteristic vector obtained from the test sample is z with the size of the characteristic vector z being m ₁ n ₁ X 1, robust dictionary RD is used as a sparse representation dictionaryThe matrix, therefore, the sparse vector of the sparse representation in this embodiment is expressed as:

then, sparse vector

Is a key issue of sparse representation, in this embodiment we use a batch matching pursuit algorithm (BMP algorithm) to calculate the sparse vector +.>

Details of the algorithm are set forth in table 1 below.

Table 1 batch matching pursuit algorithm

/>

/>

As shown in table 1, since the number of patterns in the present study is p, the sparse vector

The size of (2) is pq×1:

then, from the sparse vector

Extracting one fromDistribution vector:

B _p×1 ＝(β ₁ β ₂ … β _p ) ^T (38)

wherein beta is _i Is a single-mode sparse vector element sum, i.e

The sum of all elements corresponding to pattern i, as shown in fig. 7:

theoretically, maximum beta _i The index of (a) is the pattern of the test sample:

Pattern＝i,s.t.max(β _i ),i＝1,2,…,p (40)

then, in order to facilitate the use of the AdaBoost algorithm, the result of pattern recognition is expressed as a pattern vector P, the number of elements of the pattern vector is P, and assuming that the pattern of the test sample is pattern i, the i-th element in P is 1, and the other elements are 0.

3.2 AdaBoost-based sparse vector set)

For all audio features extracted from the axial, horizontal and vertical vibration signals, sparse vectors are calculated according to table 1. Thereafter, six sparse vectors were obtained for each test sample, as shown in table 2 below.

TABLE 2 variable in sparse representation and AdaBoost

/>

The AdaBoost algorithm is one of the most classical algorithms in ensemble learning, which combines a set of weak classifiers to generate a strong classifier. In this embodiment, to facilitate accuracy of fault diagnosis, a set of sparse vectors calculated based on audio features are combined by using an AdaBoost algorithm, which is defined as an SRC-AdaBoost algorithm.

In this embodiment, the marked samples are split into two training sample data sets, the first data set being used for K-SVD based dictionary learning and the other data set being used to obtain the weights of the AdaBoost algorithm, as shown in FIG. 8.

The steps of the SRC-AdaBoost algorithm in this embodiment are as follows:

step 1: selecting sparse vectors for training data sets

i ₁ =1, 2, …, n, u=1 (MA), 2 (MH), 3 (MV), 4 (GA), 5 (GH), 6 (GV), where i1 is the index of the sparse vector calculated from the corresponding training samples, the number of training samples is n, U is the subscript of the weak classifier, where the number of weak classifiers is u=6. The initial weights of the sparse vectors are:

step 2: determining an identified pattern vector of the training samples according to formulas (37) - (40), the identified pattern vector being noted as

Then the overall error of all weak classifiers is calculated, identifying the pattern vector by comparison>

And theoretical mode vector P _ui1 And weight d of sparse vector _ui1 Calculate uth weak classifier ζ _u Is the overall classification error of (a):

the weights of the sparse vectors of the uth weak classifiers are:

D _u ＝(d _u1 … d _ui1 … d _un ),i ₁ ＝1,2,…,n (44)

step 3: computing a weak classifier w _u Weights in constructing the strong classifier:

step 4: updating weights of training samples of the training set:

/>

step 5: the procedure of steps 1-4 is iterated for all six weak classifiers, then we get six weights w _u ,u＝1,2,3,4,5,6。

The final sparse vector is obtained as follows:

final sparse vector

Is->

And->

The structure of (3) is the same, and can be obtained:

finally, by using formulas (37) to (40), according to

A final pattern of test samples is obtained.

4. Method test

4.1 Test setup and data specification

In order to verify the method provided by the invention, single-side tooth wear (UTW) faults, double-side tooth wear (BTW) faults and broken Tooth (TB) faults are injected into a driving gear of a BSS experimental system. The drive gears for different faults are shown in fig. 9.

For each mode, 70 data samples were obtained, one data sample consisting of axial vibration data, horizontal vibration data, vertical vibration data, and speed data. The sampling rate is sr=20 KS/s, and the acquisition time of each data sample is one second, which is determined by the motion time of the quasi-uniform motion stage.

The 70 data samples are divided into a training sample data set I, a training sample data set II, containing 20 labeling samples, and a test sample data set containing 30 data samples. The training sample data set I is used for dictionary learning and the training sample data set II is used for AdaBoost. Table 3 lists the management of the data samples.

Table 3 detailed description of training sample data set and test sample data set

4.2 Audio feature extraction

For each data sample, MFCCs and GTCCs are extracted. Since the minimum rotational speed and the maximum rotational speed of the driving gear are 170.71RPM and 191.47RPM, respectively, the frame length L can be obtained according to the formulas (2) and (3) _F =351, frameDisplacement is L _S =38. Meanwhile, the number of the filters is n ₁ =20, so that the number of lines of the available audio features is m according to equation (13) ₁ =63, the number of columns of audio features is n ₁ ＝20。

In this embodiment, the MFCC characteristics of the 1 st data sample (vertical direction) of the extracted training data set I are shown in fig. 9, and the GTCC characteristics thereof are shown in fig. 11. For all time frames of MFCCs and GTCCs, the energy of the first frequency channel is much greater than the other channels. However, as shown in fig. 12, these two audio features exhibit different characteristics due to differences in shape and arrangement of filter banks of MFCCs and GTCCs. Since the Mel-triangle filter has a higher resolution than the gammatine filter in the low frequency band, the MFCC provides better resolution than the GTCC in the low frequency band, particularly in the first frequency channel. On the other hand, since the gammatine filter is much wider in the high frequency band than the Mel triangle filter, the GTCC is shown in more detail in the medium and high frequency bands than the MFCC. Thus, MFCCs and GTCCs are complementary features of vibration signals, and the synergy of MFCCs and GTCCs provides perfect performance for fault diagnosis.

4.3 Robust dictionary generation

Before the original dictionary is built, the matrix of audio features is converted into feature vectors, the conversion process is shown in fig. 5. These feature vectors extracted from the training sample dataset I are then used to construct the original dictionary and the vectors are grouped in a pattern of labeled samples. Here, the number of marked samples per pattern is q=20, and the number of patterns is p=4. Since we have vibration data obtained from three directions (axial, transverse, longitudinal) and both MFCCs and GTCCs of these data are extracted, we have six raw dictionaries. The extraction method is shown in the formulas (14) to (16), where the original dictionaries of MFCCs and GTCCs extracted from the vertical vibration data are plotted as shown in fig. 13 and 14. Since the variation ranges of the first frequency channel and the other channels are different, they are drawn in different scales, respectively, as shown in the right-hand portions of fig. 13 and 14, the lower right-hand diagram is an enlarged view of the first frequency channel, and the upper right-hand diagram is an enlarged view of the other channels. The pattern of labeled samples in training dataset I can be distinguished based on MFCCs and GTCCs, however, MFCCs provide more detail in the first channel due to the difference in filter banks, while GTCCs provide better performance in the intermediate and high frequency channels.

Then, the marked samples in the original dictionary are expanded based on the Cutout data enhancement method. Here, for each fragment of each atom in the dictionary, n _C The element of =3 is randomly replaced with the mean μ _r,s One atom is extended to n based on the Cutout data enhancement method _E =5 atoms. Thus, after data enhancement, the column number of the dictionary is expanded to pn _E q=80×5=400, and an extended dictionary is obtained as shown in fig. 15 and 16.

Finally, a robust dictionary is generated by dictionary learning based on the K-SVD algorithm. In the dictionary learning process, the maximum coefficient used in the coefficient calculation of the OMP algorithm is n _Coe =6, iteration number n _KSVDIter =30. Fig. 17 and 18 show the output of dictionary learning, i.e., a robust dictionary obtained by K-SVD algorithm-based dictionary learning.

4.4 SRC-Adaboost and fault diagnosis

In this section, the sparse vector of training dataset II was calculated based on BMP algorithm, see table 2 for details of the algorithm. In this embodiment, the number of support vectors in each batch is set to n _SV =2, iteration number n _Iter =2, thus, there is n in each sparse vector _SV n _Iter =2×2=4 non-zero elements. Since the number of atoms in the robust dictionary is pq=4×20=80, the length of each sparse vector is pq=80, and the position of the non-zero element in the sparse vector is also an index of the atoms in the robust dictionary, and at the same time, the value of the non-zero element represents the contribution of the corresponding atom in the test sample reconstruction.

As shown in table 3, the number of marked samples in training data set II is 80, so we get 80 sparse vectors. Sparse vectors, distribution vectors, and recognition pattern vectors calculated based on MFCCs are shown in fig. 19, and sparse vectors, distribution vectors, and recognition pattern vectors calculated based on GTCCs are shown in fig. 20, with misclassified samples marked with boxes.

Then, based on the identified pattern vector and the theoretical pattern vector, the overall error of the training data set II is calculated using equations (42) and (43), while the weights of the set are also calculated based on the AdaBoost algorithm. Since the number of weak classifiers is u=6, the number of iterations is also 6, which is demonstrated on the horizontal axis of fig. 21.

The number of samples of training data set II is pq=80, and therefore, as shown in the upper part of fig. 21, the number of rows is 80, and each element in the matrix represents one sample weight. The first column is the initial sample weight, and all sample weights are the same value, that is:

the middle part of fig. 21 is the overall error for each iteration, column 1 is the initial overall error, which is determined by the accuracy of the MFCC-based SRC results (fig. 19) and the GTCC-based SRC results (fig. 20).

Finally, the lower half of fig. 21 shows the set weights for the set sparse vectors. Based on the aggregate weights, the final sparse vector is obtained:

finally, the audio features of the test samples are utilized to verify the SRC-AdaBoost based fault diagnosis scheme. The number of sparse vectors is the same as the number of test samples, as shown in column 4 of table 3, and since the number of test samples is 120, the number of sparse vectors is also 120, as shown in fig. 22, most sparse vectors can correctly represent the patterns of test samples, however, the sparse vectors calculated by GTCC or MFCC alone cannot accurately identify the patterns of all test samples.

According to equation (50), the calculated final sparse vector will improve the accuracy of fault diagnosis. The final sparse vector of the test samples is shown in the upper half of fig. 23, the corresponding distribution vector is shown in the middle of fig. 23, and it can be seen from the display result that the final sparse vector calculated by the aggregate is more effective than any one of the sparse vectors in fig. 22. The fault diagnosis result was determined according to the final distribution vector, and as shown in the bottom of fig. 23, only the 55 th test sample was misclassified, and thus, the accuracy of the final fault diagnosis was 99.17%.

The invention provides a fault diagnosis scheme of a driving gear in a BSS based on audio characteristics and SRC-AdaBoost, and provides an effective solution for fault diagnosis under complex conditions such as reciprocating motion, low speed and fluctuation, complex noise interference and the like in the BSS. The invention proves that the audio characteristics such as MFCC, GTCC and the like are potential fault diagnosis characteristics in a low-speed state based on the mechanism of an auditory system. Meanwhile, for monitoring data samples only comprising a few periods, such as driving gear monitoring data in a BSS, the synergistic effect between data enhancement and dictionary learning improves the robustness of SRC, and can be widely applied to other mode recognition applications based on SRC. The combination of SRC and AdaBoost obviously improves the fault diagnosis precision under the conditions of speed fluctuation and complex noise interference. Considering low carbonization and electrification trend of commercial transportation, the fault diagnosis scheme provided by the invention provides an efficient solution for intelligent maintenance of BSS, and finally improves the safety and efficiency of transportation based on EHT, thereby having good application prospect.

The following sets forth some of the key term explanations:

gammatone cepstral coefficients (GTCC), gamma-pass cepstral coefficients.

Mel-frequency cepstral coefficients (MFCC), mel frequency cepstral coefficients.

Battery Swapping System (BSS), battery recharging system.

sparse representation-based classification (SRC), based on sparse representation.

Electric Heavy Trucks (EHT), electric heavy truck.

Adaboost is an iterative algorithm whose core idea is to train different classifiers (weak classifiers) for the same training set, and then to aggregate these weak classifiers to form a stronger final classifier (strong classifier).

Claims

1. The battery power conversion system driving gear fault diagnosis method based on the audio frequency characteristics and the SRC-AdaBoost is characterized by comprising the following steps of:

step S1, collecting multidirectional vibration signals of a sample to be diagnosed, and respectively extracting the characteristics of the mel cepstrum coefficient and the characteristics of the Gammatone cepstrum coefficient as audio characteristics for the multidirectional vibration signals;

step S2, firstly constructing an original dictionary matrix by using the extracted audio features, then carrying out sample amplification on each column vector in the original dictionary matrix by using a data enhancement method, forming an extended dictionary matrix by using the amplified column vectors, and finally learning from the redundant extended dictionary matrix by using a dictionary learning method to generate a robust dictionary matrix;

step S3, based on the robust dictionary matrix generated in the step S2, calculating sparse vectors corresponding to the mel cepstrum coefficient characteristics and the gammatine cepstrum coefficient characteristics extracted in the step S1 by using a sparse expression method; then, acquiring weight values corresponding to the sparse vectors by using an ensemble learning method; and finally, acquiring a final sparse vector by adopting a weighted summation mode, and realizing fault diagnosis of the driving gear by utilizing the distribution of non-zero elements in the final sparse vector.

2. The driving gear failure diagnosis method according to claim 1, wherein in step S2, the specific steps of constructing the original dictionary matrix are: for single audio features of a single-direction vibration signal, selecting audio features with known fault modes to form column vectors, wherein each mode has the same number of column vectors, and arranging the column vectors in sequence transversely according to the modes to form an original dictionary matrix.

3. The method for diagnosing a drive gear failure according to claim 1, wherein in step S2, the specific step of performing sample amplification on each column vector in the original dictionary matrix is as follows: the method of Cutout data enhancement is used to expand 1 column vector into a number of column vectors of a certain number.

4. The drive gear failure diagnosis method according to claim 1, wherein in step S2, the dictionary learning method includes: and performing dictionary learning by using a K-SVD method, and enabling the number of column vectors in the learned dictionary matrix to be the same as the number of column vectors in the original dictionary matrix.

5. The drive gear failure diagnosis method according to claim 1, wherein in step S3, the method of calculating the sparse vector is: and calculating sparse vectors of the audio features under the expression condition of the corresponding robust dictionary matrix by using the specific audio features of the sample to be diagnosed in the specific direction and the corresponding robust dictionary matrix and adopting a batch matching tracking method.

6. The method according to claim 5, wherein when the sparse vector is calculated by using a batch matching pursuit method, non-zero elements in the sparse vector can be obtained by iteration using a least square method, and a certain number of non-zero elements can be calculated simultaneously in each iteration.

7. The drive gear failure diagnosis method according to claim 1, wherein in step S3, the method of acquiring the weight value corresponding to each sparse vector is as follows: and acquiring weights of sparse vectors corresponding to the audio features in all directions by using training samples with known fault modes and adopting an AdaBoost method.