CN108470570B - Abnormal sound detection method for motor - Google Patents
Abnormal sound detection method for motor Download PDFInfo
- Publication number
- CN108470570B CN108470570B CN201810062638.6A CN201810062638A CN108470570B CN 108470570 B CN108470570 B CN 108470570B CN 201810062638 A CN201810062638 A CN 201810062638A CN 108470570 B CN108470570 B CN 108470570B
- Authority
- CN
- China
- Prior art keywords
- motor
- audio
- sample
- wavelet packet
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000002159 abnormal effect Effects 0.000 title claims abstract description 50
- 238000001514 detection method Methods 0.000 title claims abstract description 19
- 239000013598 vector Substances 0.000 claims abstract description 48
- 230000005236 sound signal Effects 0.000 claims abstract description 25
- 238000000513 principal component analysis Methods 0.000 claims abstract description 16
- 238000005070 sampling Methods 0.000 claims abstract description 10
- 239000011159 matrix material Substances 0.000 claims description 39
- 238000012706 support-vector machine Methods 0.000 claims description 31
- 238000010606 normalization Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 10
- 238000012549 training Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 8
- 238000000354 decomposition reaction Methods 0.000 claims description 7
- 238000012935 Averaging Methods 0.000 claims description 6
- 238000003745 diagnosis Methods 0.000 claims description 5
- 230000010355 oscillation Effects 0.000 claims description 3
- 101100380328 Dictyostelium discoideum asns gene Proteins 0.000 claims 1
- 241000623377 Terminalia elliptica Species 0.000 claims 1
- 101150062095 asnA gene Proteins 0.000 claims 1
- 238000004519 manufacturing process Methods 0.000 abstract description 8
- 230000036541 health Effects 0.000 abstract description 2
- 238000000034 method Methods 0.000 description 14
- 230000008859 change Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 210000005069 ears Anatomy 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 210000004556 brain Anatomy 0.000 description 2
- 230000002950 deficient Effects 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000001052 transient effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000037433 frameshift Effects 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000005406 washing Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
- G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
Landscapes
- Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention discloses a motor abnormal sound detection method, which comprises the steps of setting sampling frequency and sampling duration t, carrying out audio signal acquisition, carrying out frame division windowing on audio signals, setting the frame length L of each frame and the overlapping length M of two adjacent frames, and dividing the audio signals into N frames of signals; then extracting audio features by using 6 layers of wavelet packets; and performing principal component analysis on the audio features extracted from the wavelet packet to obtain feature vectors and the like. The motor abnormal sound detection device can assist workers in identifying abnormal sound of the motor, improve detection efficiency and ensure the delivery quality of products, so that the overall production efficiency of an enterprise is improved, the manufacturing cost of the enterprise is reduced, and the physical health of the workers is protected; the problem that the audio signal of the motor is unstable can be effectively solved, abnormal sound faults of the unstable motor can be effectively detected, and the identification accuracy rate is high.
Description
Technical Field
The invention relates to the field of motor fault detection, in particular to a method for detecting abnormal sound of a motor.
Background
China is the main production place of various small household electric motors for washing machines, household air conditioners, refrigerators, electric fans and the like, and the annual output of the household air conditioner motor is more than billions.
On a small motor production line, a manual listening method is generally adopted before products are off-line to distinguish good products from defective products, namely, workers listen to the sound generated when the motor runs in a sound insulation room sequentially by ears, and whether the motor has faults or not is judged according to personal experience of the workers.
Because of the subjective judgment of people, the automatic device is difficult to replace for a long time. Moreover, the personal experience of the evaluators cannot establish a uniform evaluation standard, and different evaluators may generate different conclusions. In the process of mass production, the working procedure consumes a large amount of labor cost, the repeated and monotonous listening work easily causes personnel fatigue and is easy to misjudge, and if individual defective products are mixed into the whole batch of finished products, serious economic loss is brought to a factory, and even the reputation of the products is seriously influenced.
CN201510266743.8 (publication No. CN104992714A) discloses a method for detecting abnormal sound of a motor, comprising the following steps: 1. carrying out audio acquisition when the motor is in an idle state; 2. converting the acquired time domain audio signal of the motor into a frequency domain waveform through Fourier transform; 3. if the waveform exists outside the highest value of the normal frequency domain range of the motor, the motor is considered to have abnormal sound; if no waveform exists outside the highest value of the normal frequency domain range of the motor, the motor is indicated to have no abnormal sound. The detection method of the abnormal sound of the motor has the following disadvantages: 1. the audio signal belongs to a quasi-steady-state signal, namely, the audio signal is stable for a short time; the Fourier transform is a common means for processing steady-state signals, and can only extract signal characteristics of the steady-state signals; however, in the abnormal sound fault of the motor, a large number of fault samples of transient unstable signals exist; for these non-stationary signals, the fourier transform is not as powerful. 2. Abnormal sound is judged by judging whether a waveform exists beyond a specified maximum value, the method has no self-adaption capability and low universality: since technicians are required to reset the threshold values for different models of motor products; and when the audio signal is collected through the sensor, the distance between the sensor and the sound source also influences the threshold setting. 3. No matter the maximum value is set or the oscillogram is compared, judgment is required by professionals, and full-automatic identification cannot be achieved.
Disclosure of Invention
The invention aims to provide a motor abnormal sound detection method which can identify fault samples of unsteady signals, has high universality and does not need personnel to participate in the fault identification process.
The abnormal sound detection method of the motor comprises the following steps:
step 1, setting sampling frequency and sampling duration t, and collecting audio signals when a motor is in a no-load state;
step 2, performing frame division and windowing processing on the audio signal, setting the frame length L of each frame and the overlapping length (frame shift) M of two adjacent frames, and dividing the audio signal into N frames of signals; the frame length L is preferably that the signal in each frame can be regarded as a steady-state signal, so that the influence of unsteady state and time variation of the whole audio signal is avoided;
step 3, extracting audio features by using 6 layers of wavelet packets;
step 4, performing principal component analysis on the audio features extracted from the wavelet packet to obtain feature vectors;
step 5, selecting N qualified motor audio samples, repeating the steps 2-4, respectively extracting the feature vector of each qualified motor audio sample, and training a support vector machine by using the feature vectors of the qualified motor audio samples to obtain the support vector machine of the qualified samples;
and 6, when motor abnormal sound diagnosis is carried out, repeating the steps 1-4 to obtain the feature vector of the motor audio sample, and inputting the feature vector into a support vector machine to judge whether the sample is qualified.
Further, in step 2, the frame length L is 1S, and the overlap length M is 0.5S.
Further, the step of extracting the audio features in the step 3 comprises the following steps:
step 3-1, performing 6-layer wavelet packet decomposition on each frame of signal, and calculating the energy E of each node of the 6 th-layer wavelet packet6,j(j ═ 0,1,2,3, …,63), layer 6 wavelet packet has a value of 26A wavelet packet node;
step 3-2, energy E for each node6,jCarrying out normalization processing, wherein the normalization calculation formula is as follows:
j represents the jth wavelet packet node of layer 6;
step 3-3, obtaining a normalized matrix { lambdaj，kIn which λ isj，kRepresenting the energy of the jth wavelet packet node of the kth frame; normalizing the mean value of each row of the matrix α, { α }j-forming a column vector of the image data,
column vector { alphajA first set of characteristics of the audio as one sample duration,
calculating alphajWhen the signal is abnormal, j is more than or equal to 3, because in the abnormal sound signal of the motor, the frequency band represented by the wavelet packet nodes (6,0) (6,1) (6,2) belongs to the ultra-low frequency band range, and the human ear is not sensitive to the signal of the frequency band;
step 3-4, solving a normalized matrix { lambdaj，kPeak-to-peak value of β, { β } for each row ofjForm a column vector. Selecting the peak-peak value beta of j more than or equal to 3jAs a second set of features for audio of one sample duration, βj＝max{λj,k|k＝1,2,3,...,59}-min{λj,k1,2,3, 59, and j 3,4,5, 63. Because the frequency band represented by the wavelet packet nodes (6,0) (6,1) (6,2) in the abnormal sound signal of the motor belongs to the ultra-low frequency band range, the human ear is not sensitive to the signal of the frequency band.
Further, in the step 3-1, db5 wavelet basis functions are selected for 6-layer wavelet decomposition; spatially converting each frame signal f (t)
Projection is carried out to obtain: f. ofj,n(t)＝∑k∈Zdj,n(t)uj,n,k(t) wherein,
is the wavelet packet coefficient, { uj，n，k(t)}kE.g. Z as space
J represents a size index and is a frequency domain parameter, k represents a position index and is a time parameter, n represents the oscillation frequency, and Z represents all positive integers;
arranging the nodes of the wavelet packet on the layer 6 from low to high according to the frequency, and solving the energy of each node of the wavelet packet on the layer 6, wherein the calculation formula is as follows:
wherein d isj，k(j ═ 0,1, 2.., 63, k ═ 1,2, 3.., n) denotes S6，jThe wavelet packet coefficient of (a).
Further, the step 4 of performing principal component analysis on the audio features comprises the following steps:
step 4-1, taking M groups of samples, and taking a column vector { alpha ] of each samplejAs a column of matrices to construct a matrix X, X ═ αj，nI j 3,4, 5.., 63, n1, 2, 3.. 60}, n representing the nth sample;
with a column vector of each sample { betajOne column of the matrix is used to construct the matrix Y, Y ═ βj，n3,4, 5.. 63, n ═ 1,2, 3.. 60}, where n denotes the nth sample; the M groups of samples comprise qualified samples and fault samples;
step 4-2, zero-averaging each row of X and Y, namely subtracting the average value of the row from each row element to obtain a zero-averaging matrix
And
step 4-3, solving covariance matrix
And
step 4-4, solving covariance matrix CαCharacteristic values and characteristic vectors of CβThe eigenvalues and eigenvectors of (a);
step 4-5, adding CαCharacteristic vector of (1) according to CαThe eigenvalues of (A) are arranged in rows from large to small, and the first 16 rows are taken to form a matrix PαMixing C withβCharacteristic vector of (1) according to CβThe eigenvalues of (A) are arranged in rows from large to small, and the first 16 rows are taken to form a matrix Pβ，
Calculating wavelet packet characteristics alphajFeature vector lambda after principal component analysisα，λα＝Pα*[α3，α4，α5，…，α63 T；
Computing wavelet packet characteristics betajFeature vector mu after principal component analysisβ，μβ＝Pβ*[β3，β4，β5，...，β63]T。
Further, in step 5, the feature vector λ of each sample is determinedαArranged in rows, each row representing a sample, forming a sample matrix MαUsing a libsvm-mat-2.91-1 open source toolkit, setting parameters of an SVM to be s 2-n 0.053-c 2-g 0.0018, training a support vector machine, and obtaining a support vector machine Model SVM-Model 1;
feature vector mu of each sampleβArranged in rows, each row representing a sample, to form a label matrix MβThe SVM parameter is set to "-s 2-n 0.071-c 2-g 0.0029" using the "libsvm-mat-2.91-1" open source kit. And training the support vector machine to obtain a support vector machine Model SVM-Model 2.
Further, in step 6, when the abnormal sound of the motor is diagnosed, the audio frequency of the motor under the no-load condition is obtained, and the feature vector lambda of the audio frequency of the motor is extractedαAnd a feature vector muβRespectively judging by using a support vector machine Model SVM-Model and a support vector machine Model SVM-Model2, and judging the audio frequency of the motor as a qualified sample when the two support vector machine models are identified to be qualified; the audio of the motor is determined to be a faulty sample whenever one of the support vector machine models identifies a fault.
The invention has the advantages that:
1. the motor abnormal sound identification can be assisted by workers, the detection efficiency is improved, the delivery quality of products is ensured, the overall production efficiency of an enterprise is improved, the manufacturing cost of the enterprise is reduced, and the physical health of the workers is protected;
2. the problem of unsteady state of the motor audio signal can be effectively solved, abnormal sound faults of the unsteady state motor can be effectively detected, and the identification accuracy rate is high;
3. as used herein
The normalization processing is carried out, so that the problem that the proportion of the energy change of the abnormal sound node in the total energy is overlooked due to too small proportion can be effectively solved, and the problem that the detection threshold value is influenced due to the change of the distance between the sensor and the sound source is also avoided.
4. The abnormal sound faults are different on different wavelet packet nodes because the abnormal sound faults are not different on all the wavelet packet nodes, and the abnormal sound faults of different motors are different on different wavelet packet nodes. The feature vectors are automatically converted into the feature space with the maximum difference between qualified signals and abnormal sound fault signals by using principal component analysis, and manual screening is not needed.
5. The support vector machine is used for automatically judging whether abnormal sound faults exist in the motor audio, and manual intervention is not needed in the using process. Three parameters of a penalty function, a kernel function and an abnormal sample proportion of the support vector machine are set, different motor characteristics are used for training different motor types, and the support vector machine model suitable for the corresponding motor can be obtained. And the three parameters have universality for different motors and do not need to be changed. With the machine learning method, it is also not necessary to set different thresholds for different motors. The support vector machine sets a discriminant function aiming at the motor of the model according to the training sample. The method is characterized in that a discriminant function is established on the basis of normal sound samples which are easily obtained on a production line. Moreover, as the number of samples increases and the coverage of the feature distribution is enlarged, the accuracy of the discrimination also increases.
Drawings
Fig. 1 is a schematic diagram of the framing windowing in step 2.
Fig. 2 is a schematic diagram of a 6-layer wavelet packet decomposition.
FIG. 3 is ajSchematic diagram of the calculation method of (1).
FIG. 4 shows normalized energy λ of 12 failure samples and 12 qualified samples of a plastic-packaged motor for YYYW 23-6-7051 air conditionerj(j 4,24,30) experimental comparison results with time.
FIG. 5 is betajSchematic diagram of the calculation method of (1).
FIG. 6 is a diagram of a plastic-encapsulated motor for YYYW 23-6-7051 air conditioner, each of which shows normalized energy λ of one failure sample and 12 qualified samplesj(j 27,36,39) experimental comparisons over time.
Fig. 7 is a flow chart of the present invention.
Fig. 8 is a flowchart at the time of motor abnormal sound diagnosis.
Detailed Description
The abnormal sound detection method of the motor comprises the following steps:
step 1: and (3) collecting audio signals when the motor is in an idle state, wherein the sensor is an electret free field type microphone, the sampling frequency is 48k, and the collection time is 30 s. The reason for the sampling frequency selection of 48k is in the human ear auditory frequency range of 20 to 20 k. According to the Nyquist sampling theorem, the sampling frequency of 48k just covers the auditory range of human ears, and the auditory system of the human ears can be effectively simulated.
Step 2: the audio signal is subjected to frame windowing, i.e. a window with a frame length L of 1s is used to cut the signal, and the overlap length M of two adjacent frames is 0.5s, as shown in fig. 1. A 30s audio signal can be decomposed into 59 frames of audio signal according to this method. The purpose of frame windowing is to avoid the effects of unsteadiness and time-varying of the entire audio signal. The signal can be considered as a steady-state signal within each frame, which facilitates feature extraction. This is also a mechanism of imitating artificial listening, because the human brain judges abnormal sound fault also judges the audio signal heard at the present moment, that is, the brain only judges whether there is fault for the audio frequency in the transient impression.
And step 3: and carrying out wavelet packet decomposition on each frame of signal. And (4) selecting a db5 wavelet basis function to carry out 6-layer wavelet packet decomposition. Spatially converting each frame signal f (t)
Projection is carried out to obtain: f. ofj，n(t)＝∑k∈Zdj，n(t)uj，n，k(t), wherein:
is the wavelet packet coefficient, { uj，n，k(t)}k∈ZIs a space
The parameter j represents a scale index (frequency domain parameter), the parameter k represents a position index (time parameter), the parameter n represents the number of oscillations, and Z represents all positive integers.
And arranging the wavelet packet coefficients of the layer 6 from low to high according to the frequency. As shown in FIG. 2, Si，jRepresenting the jth node of the ith layer. And calculating the energy of each node of the wavelet packet of the layer 6, wherein the calculation formula is as follows:
wherein d isj，k(j ═ 0,1, 2.., 63, k ═ 1,2, 3.., n) denotes S6，jThe wavelet packet coefficient of (a).
4. Then for each E6，j(j ═ 1,2, 3.., 63.) the node energies are "normalized". The calculation formula is as follows:
wherein E is selected6，j-1The normalization basis is determined from a number of experimental results.
According to a large number of experimental results, the abnormal sound faults of the motor are mostly concentrated on a frequency band behind 1.5 kHz. But 80% of the energy of the motor audio is concentrated below 1.5 kHz. If the normalization method of the existing document is adopted, namely the sum of the energies of all nodes or the maximum value of the energy nodes is selected as the normalization denominator, the proportion of the energy change of the key nodes in the total energy is overlooked, and the judgment result is influenced. The 'normalization' operation can also effectively avoid the influence of the distance between the audio sensor and the sound source on the abnormal sound detection result. Since the acquired signal is weaker the further the sensor is from the source. But the ratio of the energy of each frequency band of the sound to the total energy does not change along with the distance of collection. By using the method, the influence of the change of the acquisition distance can be effectively avoided.
5. Processing each frame signal according to the above to obtain a matrix { lambdaj，kJ represents a layer 6 wavelet packet node number (j is 1,2, 3.., 63), and k represents a frame number (k is 1,2, 3.., 59). The average value { alpha ] of each row of the matrix is obtainedjAs a set of characteristics of the 30-second audio, a formula is given below. Only selecting lambda of j ≧ 3jCalculating mean characteristics
This is because, in the abnormal motor sound signal, the frequency band represented by the wavelet packet nodes (6,0) (6,1) (6,2) belongs to the ultra-low frequency band range, and the human ear is not sensitive to the signal of the frequency band.
αjThe calculation method of (2) is shown in fig. 3. FIG. 3 shows the normalized energy λ of a sample xj(j ═ y) changes over time. α y represents the average of λ j in all frames.
Fig. 4 shows the test data in the laboratory. The plastic package motor for the YYW23-6-7051 air conditioner is used in the experiment, and the results of the experiment of the normalized energy change with time under the 6 th layer wavelet packet node of 12 fault samples and 12 qualified samples (j is 4,24 and 30) are compared in the figure. As can be seen from fig. 4, the mean value of the failed samples is significantly different from the mean value of the qualified samples, using αjNormalized energy λ of descriptionjThe average value characteristic of the method can effectively distinguish the abnormal sound samples from the normal qualified samples.
6. Matrix { lambda ] is obtainedj，kPeak-Peak value of each line { beta }jAs another set of features for 30 second audio, only λ after j ≧ 3 is chosenjCalculating mean feature betaj，
βj＝maX{λj，k|k＝1，2，3，...，59}-min{λj，k|k＝1，2，3，...，59}，j＝3，4，5，...，63；
βjThe calculation method of (2) is shown in fig. 5. FIG. 5 shows the normalized energy λ of a sample xj(j ═ y) as a function of time, βjDenotes λjThe maximum value minus the minimum value in all frames.
As shown in FIG. 6, experiments were conducted using a plastic-sealed motor for an air conditioner of YYW23-6-7051, each of which shows normalized energy λ of one failure sample and 12 qualified samplesj(j-31, 33, 37) experimental comparison results with time. β represents the fluctuation range of the signal. As can be seen from fig. 6, the fluctuation range of the fault signal at nodes 31, 33 and 37 is larger than that of the qualified signal, so that the abnormal sound fault can be effectively distinguished by using the description of beta as a characteristic.
7. And 3-6 steps are the steps of extracting audio features by using the wavelet packet. The abnormal sound faults are different on different wavelet packet nodes because the abnormal sound faults are not different on all the wavelet packet nodes, and the abnormal sound faults of different motors are different on different wavelet packet nodes. So for { alphajJ ═ 3,4, 5.., 63) and { beta }jAnd (j) wavelet packet characteristics are subjected to principal component analysis, qualified signals and abnormal sound fault signals can be automatically converted into characteristic spaces with the largest difference, and manual screening is not needed.
Principal component analysis is a technique in statistics that simplifies the data set. It is a linear transformation. This transformation transforms the data into a new coordinate system such that any data projection is maximally different above the first coordinate (called the first principal component), next to the second coordinate (the second principal component), and so on. Principal component analysis can reduce the dimensionality of the data set while preserving the features in the data set that contribute most to variance.
The method specifically comprises the following steps of analyzing the main components of the wavelet packet energy characteristics:
step 7-1, taking M groups of samples, and taking column vector { alpha ] of each samplejAs a column of matrices to construct a matrix X, X ═ αj，nI j 3,4, 5.., 63, n1, 2, 3.. 60}, n representing the nth sample;
with a column vector of each sample { betajOne column of the matrix is used to construct the matrix Y, Y ═ βj，nI j 3,4,5, 3, 63, n1, 2,3, 60, where n denotes the nth oneA sample; the M groups of samples comprise qualified samples and fault samples;
step 7-2, zero-averaging each row of X and Y, namely subtracting the average value of the row from each row element to obtain a zero-averaging matrix
And
step 7-3, solving covariance matrix
And
step 7-4, solving covariance matrix CαCharacteristic values and characteristic vectors of CβThe eigenvalues and eigenvectors of (a);
step 7-5, adding CαCharacteristic vector of (1) according to CαThe eigenvalues of (A) are arranged in rows from small to large, and the first 16 rows are taken to form a matrix PαMixing C withβCharacteristic vector of (1) according to CβThe characteristic values of (A) are arranged in rows from small to large, and the first 16 rows are taken to form a matrix Pβ，
Calculating wavelet packet characteristics alphajFeature vector lambda after principal component analysisα，λα＝Pα*[α3，α4，α5，…，α63 T；
Computing wavelet packet characteristics betajFeature vector mu after principal component analysisβ，μβ＝Pβ*[β3，β4，β5，...，β63]T. Characterizing wavelet packets by alphajAnd betajAnd calculating according to the following formula to obtain the feature vector after principal component analysis.
8. Selecting 60 qualified motor samples, and repeating the steps 1-7 to obtain the lambda of each sampleαAnd muβFeature(s). Arranging the lambda alpha characteristics of each sample according to rows, wherein each row represents one sample, and forming a matrix M of 60 x 16α. The open source toolkit of libsvm-mat-2.91-1 is used to set the parameters of the SVM to "-s 2-n 0.053-c 2-g 0.0018". And training the support vector machine to obtain a support vector machine Model SVM-Model 1.
In the same way, mu of each sample was measuredβThe features are arranged in rows, each row representing a sample, forming a 60 x 16 matrix Mβ. The SVM parameter is set to "-s 2-n 0.071-c 2-g 0.0029" using the "libsvm-mat-2.91-1" open source kit. And training the support vector machine to obtain a support vector machine Model SVM-Model 2.
9. When motor abnormal sound diagnosis is carried out, lambda of motor audio is extractedαAnd muβThe characteristics are discriminated by using an SVM-Model1 and an SVM-Model2, respectively. Only if both support vector machine models are identified as qualified, the final sample is determined to be a qualified sample. If only one model is judged to be a fault, the sample is judged to be a fault sample.
Claims (5)
1. The abnormal sound detection method of the motor comprises the following steps:
the abnormal sound detection method of the motor comprises the following steps:
step 1, setting sampling frequency and sampling duration t, and collecting audio signals when a motor is in a no-load state;
step 2, performing frame division and windowing processing on the audio signal, setting the frame length L of each frame and the overlapping length M of two adjacent frames, and dividing the audio signal into N frames of signals;
step 3, extracting audio features by using 6 layers of wavelet packets;
step 4, performing principal component analysis on the audio features extracted from the wavelet packet to obtain feature vectors;
step 5, selecting N qualified motor audio samples, repeating the steps 2-4, respectively extracting the feature vector of each qualified motor audio sample, and training a support vector machine by using the feature vectors of the qualified motor audio samples to obtain the support vector machine of the qualified samples;
step 6, when motor abnormal sound diagnosis is carried out, repeating the steps 1-4 to obtain a feature vector of a motor audio sample, and inputting the feature vector into a support vector machine to judge whether the sample is qualified or not;
the step 3 of extracting the audio features comprises the following steps:
step 3-1, performing 6-layer wavelet packet decomposition on each frame of signal, and calculating the energy E of each node of the 6 th-layer wavelet packet j6,（j=0,1,2,3, …,63) layer 6 wavelet packet having
A wavelet packet node;
step 3-2, energy E for each node j6,Carrying out normalization processing, wherein the normalization calculation formula is as follows:
，ja jth wavelet packet node representing layer 6;
step 3-3, carrying out the normalization processing on each frame signal to obtain a normalization matrix
Wherein, in the step (A),
is shown askFirst of framejEnergy of the wavelet packet node; average value of each row of normalized matrix
,
A column vector is formed and,
column vector
A first set of features of the audio that are one sample duration;
step 3-4, obtaining a normalized matrix
Peak-to-peak value of each row of
,
Forming column vectors, selecting the peak-peak value with j being more than or equal to 3
As a second set of features for audio of one sample duration,
；
in the step 3-1, db5 wavelet basis functions are selected for 6-layer wavelet decomposition; each frame signalf(t)In space
Projection is carried out to obtain:
wherein, in the step (A),
is the coefficient of the wavelet packet and is,
is a space
J represents a size index and is a frequency domain parameter, k represents a position index and is a time parameter, n represents the oscillation frequency, and Z represents all positive integers;
arranging the nodes of the wavelet packet on the layer 6 from low to high according to the frequency, and solving the energy of each node of the wavelet packet on the layer 6, wherein the calculation formula is as follows:
wherein, in the step (A),
to represent
The wavelet packet coefficient of (a).
2. The abnormal noise detection method of the motor according to claim 1, wherein: in step 2, the frame length L =1S, and the overlap length M = 0.5S.
3. The abnormal noise detection method of the motor according to claim 1, wherein: step 4, the principal component analysis of the audio features comprises the following steps:
step 4-1, taking M groups of samples, and using the column vector of each sample
The matrix X is constructed as one column of a matrix,
denotes the firstnA sample is obtained;
with a column vector of each sample
The matrix Y is constructed as one column of the matrix,
wherein
Is shown asnA sample is obtained; the M groups of samples comprise qualified samples and fault samples;
step 4-2, zero-averaging each row of X and Y, namely subtracting the average value of the row from each row element to obtain zero-averaging moment
And
；
step 4-3, solving covariance matrix
；
Step 4-4, solving covariance matrix
The feature value and the feature vector of (c),
the eigenvalues and eigenvectors of (a);
step 4-5, mixing
Is given by the feature vector
The eigenvalues of (A) are arranged in rows from large to small, and the first 16 rows are taken to form a matrix
Will be
Is given by the feature vector
The characteristic values of (1) are arranged in a row from large to small6 rows form a matrix
，
Computing wavelet packet characteristics
Feature vector after principal component analysis
，
；
Computing wavelet packet characteristics
Feature vector after principal component analysis
，
。
4. The abnormal noise detection method of the motor according to claim 3, wherein: in step 5, the feature vector of each sample is calculated
Arranged in rows, each row representing a sample, forming a sample matrix
Using a libsvm-mat-2.91-1 open source toolkit, setting parameters of an SVM to be s 2-n 0.053-c 2-g 0.0018, training a support vector machine, and obtaining a support vector machine Model SVM-Model 1; feature vector of each sample
Arranged in rows, each row representing a sample, to form a label matrix
And (3) using a libsvm-mat-2.91-1 open source toolkit, setting parameters of the SVM to be s 2-n 0.071-c 2-g 0.0029, training the support vector machine, and obtaining a support vector machine Model SVM-Model 2.
5. The abnormal noise detection method of the motor according to claim 4, wherein: step 6, when motor abnormal sound diagnosis is carried out, the audio frequency of the motor under the no-load condition is obtained, and the feature vector of the motor audio frequency is extracted
And feature vectors
Using support vector machine models, respectivelySVM-ModellSupport vector machine modelSVM-Model2, judging, and when the two support vector machine models are identified to be qualified, judging the audio frequency of the motor to be a qualified sample; the audio of the motor is determined to be a faulty sample whenever one of the support vector machine models identifies a fault.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810062638.6A CN108470570B (en) | 2018-01-23 | 2018-01-23 | Abnormal sound detection method for motor |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810062638.6A CN108470570B (en) | 2018-01-23 | 2018-01-23 | Abnormal sound detection method for motor |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108470570A CN108470570A (en) | 2018-08-31 |
| CN108470570B true CN108470570B (en) | 2021-06-18 |
Family
ID=63266058
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810062638.6A Active CN108470570B (en) | 2018-01-23 | 2018-01-23 | Abnormal sound detection method for motor |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108470570B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112146142B (en) * | 2019-06-27 | 2022-01-25 | 宁波方太厨具有限公司 | Method for recognizing sound fault of range hood |
| CN111860241A (en) * | 2020-07-07 | 2020-10-30 | 国网山东省电力公司威海供电公司 | Power equipment discharge fault identification method based on wavelet packet analysis |
| CN114414038A (en) * | 2022-01-26 | 2022-04-29 | 歌尔科技有限公司 | Method, device, equipment and medium for detecting abnormal rotation sound of motor module |
| CN114974301A (en) * | 2022-04-15 | 2022-08-30 | 北京瑞森新谱科技股份有限公司 | Abnormal sound detection method, computer readable storage medium and electronic device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102944418A (en) * | 2012-12-11 | 2013-02-27 | 东南大学 | Wind turbine generator group blade fault diagnosis method |
| CN104992714A (en) * | 2015-05-22 | 2015-10-21 | 株洲联诚集团有限责任公司 | Motor abnormal sound detection method |
| CN105841797A (en) * | 2016-03-15 | 2016-08-10 | 中南大学 | Window motor abnormal noise detection method and apparatus based on MFCC and SVM |
| CN107393555A (en) * | 2017-07-14 | 2017-11-24 | 西安交通大学 | A kind of detecting system and detection method of low signal-to-noise ratio abnormal sound signal |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FR2994495B1 (en) * | 2012-08-10 | 2015-08-21 | Thales Sa | METHOD AND SYSTEM FOR DETECTING SOUND EVENTS IN A GIVEN ENVIRONMENT |
-
2018
- 2018-01-23 CN CN201810062638.6A patent/CN108470570B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102944418A (en) * | 2012-12-11 | 2013-02-27 | 东南大学 | Wind turbine generator group blade fault diagnosis method |
| CN104992714A (en) * | 2015-05-22 | 2015-10-21 | 株洲联诚集团有限责任公司 | Motor abnormal sound detection method |
| CN105841797A (en) * | 2016-03-15 | 2016-08-10 | 中南大学 | Window motor abnormal noise detection method and apparatus based on MFCC and SVM |
| CN107393555A (en) * | 2017-07-14 | 2017-11-24 | 西安交通大学 | A kind of detecting system and detection method of low signal-to-noise ratio abnormal sound signal |
Non-Patent Citations (2)
| Title |
|---|
| Bearing fault detection of induction motor using wavelet and support vector machines(SVMs);P.Konar etc;《Applied Soft Computing》;20110325;第4203-4211页 * |
| 基于机器学习方法的电机异音检测研究;刘力源;《中国优秀硕士学位论文全文数据库 信息科技Ⅱ辑》;20150315(第03期);第1-56页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108470570A (en) | 2018-08-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108470570B (en) | Abnormal sound detection method for motor | |
| CN111222290B (en) | Multi-parameter feature fusion-based method for predicting residual service life of large-scale equipment | |
| CN111354366B (en) | Abnormal sound detection method and abnormal sound detection device | |
| CN110355608B (en) | Cutter wear amount prediction method based on self-attention mechanism and deep learning | |
| CN109858104B (en) | Rolling bearing health assessment and fault diagnosis method and monitoring system | |
| CN108490349B (en) | Motor abnormal sound detection method based on Mel frequency cepstrum coefficient | |
| TWI587294B (en) | Detection method of abnormal sound of apparatus and detection device | |
| CN114619292B (en) | Milling cutter wear monitoring method based on fusion of wavelet denoising and attention mechanism with GRU network | |
| CN113255848B (en) | Water turbine cavitation sound signal identification method based on big data learning | |
| CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
| CN108447503B (en) | Motor abnormal sound detection method based on Hilbert-Huang transformation | |
| CN111898443B (en) | Flow monitoring method for wire feeding mechanism of FDM type 3D printer | |
| CN110561191A (en) | Numerical control machine tool cutter abrasion data processing method based on PCA and self-encoder | |
| CN111508517A (en) | Intelligent micro-motor product control method based on noise characteristics | |
| CN115576284A (en) | Clothing workshop intelligent management method and system | |
| CN106094749A (en) | Based on the nonlinear fault detection method and the application that improve nuclear entropy component analysis | |
| WO2020056812A1 (en) | Environmental parameter weight determining method and system for evaluating indoor environmental quality | |
| CN112664410B (en) | Big data-based modeling method for unit online monitoring system | |
| CN116627116B (en) | Process industry fault positioning method and system and electronic equipment | |
| CN208147865U (en) | A kind of sound recognition system applied to robot used for intelligent substation patrol | |
| CN111352365A (en) | Dustproof ventilation type electric power and electrical equipment cabinet and control method | |
| CN110727669A (en) | Device and method for cleaning sensor data of power system | |
| CN115841478A (en) | Quality detection system applied to vehicle-mounted air conditioner hose production management and control | |
| CN114358160A (en) | Data anomaly detection method in power system | |
| CN111881502A (en) | Bridge state discrimination method based on fuzzy clustering analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |