Motor fault monitoring method based on key audio recognition

Abstract

The application discloses a motor fault monitoring method based on key audio recognition, comprising: acquiring a motor audio sample; acquiring a signal frequency domain feature based on the motor audio sample; positioning a fault audio section and a normal audio section in the signal frequency domain feature; acquiring a fault feature sample and a normal feature sample based on the fault audio section and the normal audio section; training an SVM support vector machine by using the fault feature sample and the normal feature sample to acquire a fault monitoring model, and the fault monitoring model is used for motor fault monitoring. The application uses a sound sensor to collect motor operation sound in real time, uses normal audio and fault audio to establish a training sample, extracts an audio feature, and trains a machine learning algorithm model, so that automatic real-time diagnosis of motor faults is realized, and motor maintenance efficiency is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault monitoring, and particularly relates to a motor fault monitoring method based on key audio recognition. BACKGROUND

[0002] In long-term uninterrupted operation, motors are prone to faults. For example, an industrial site auxiliary cooling motor is required to operate continuously for 24 hours according to the nature of the equipment. If maintenance is not performed in time after a fault occurs, the equipment will fail, which may cause serious consequences.

[0003] For motor fault monitoring, the traditional method is manual periodic inspection, and the automatic method usually uses motor current monitoring or motor vibration signal monitoring.

[0004] The monitoring method based on current or vibration signals needs to select points in advance and arrange corresponding sensors, and a fault diagnosis algorithm is established by combining the electrical or mechanical model analysis of the motor. The implementation process is relatively complex, and when it is implemented in a motor that has been installed and operated, there are problems such as difficulty in installing and arranging sensors.

[0005] The manual inspection method based on sound signals needs manual on-site viewing, which is labor-intensive and inefficient, and cannot timely and effectively find faults.

[0006] The existing automatic diagnosis method based on sound signals has a relatively conventional feature extraction method, which usually applies a mel-frequency cepstral coefficient or other voiceprint features to cover all time-domain or frequency-domain features. In actual application, especially in an industrial site with a complex sound environment, sudden noise interference that is not considered in model training, such as occasional human voice, bird calls, thunder, and short-time work equipment sound in the field, will have a great impact on sample features. These methods require comprehensive collection of fault audio to ensure better recognition of faults. Or in the model training process, artificial selection is required to determine the typical fault sound frequency domain to further ensure accuracy. SUMMARY

[0007] To solve the above technical problems, the application provides a motor fault monitoring method based on key audio recognition, which can use an automatic method to automatically measure the sound signals emitted by the motor using a sound collection device, and process and analyze the signals in real time to improve the efficiency of motor fault monitoring and reduce the labor intensity.

[0008] To achieve the above purpose, the application provides a motor fault monitoring method based on key audio recognition, which comprises:

[0009] obtaining a motor audio sample;

[0010] obtaining a signal frequency domain feature based on the motor audio sample;

[0011] locating a fault audio segment and a normal audio segment in the signal frequency domain feature;

[0012] based on the fault audio segment and the normal audio segment, obtaining a fault feature sample and a normal feature sample;

[0013] training an SVM support vector machine using the fault feature sample and the normal feature sample, to obtain a fault monitoring model, the fault monitoring model being used for motor fault monitoring.

[0014] Optionally, based on the motor audio sample, obtaining a signal frequency domain feature comprises:

[0015] normalizing the motor audio sample to obtain a normalized audio sample;

[0016] performing signal filtering processing on the normalized audio sample to obtain a filtered audio sample;

[0017] based on the filtered audio sample, obtaining the signal frequency domain feature.

[0018] Optionally, based on the filtered audio sample, the method for obtaining the signal frequency domain feature is:

[0019] performing fast Fourier transform on the filtered audio sample to obtain a frequency complex spectrum;

[0020] based on the frequency complex spectrum, obtaining a sample frequency domain feature;

[0021] based on the sample frequency domain feature, obtaining the signal frequency domain feature.

[0022] Optionally, the method for performing fast Fourier transform on the filtered audio sample to obtain X(k) is:

[0023]

[0024] wherein, x i is the output signal after filtering in the previous step, k is the frequency component number, i is the sampling sequence number, N is the sample sequence length, X(k) is the kth frequency component complex spectrum, and j is the imaginary unit.

[0025] Optionally, locating a fault audio segment and a normal audio segment in the signal frequency domain feature comprises:

[0026] obtaining the minimum horizontal distance and the peak height of adjacent peaks in the signal frequency domain feature;

[0027] Automatic peak searching is performed based on the minimum horizontal distance of the adjacent peaks and the peak searching height, and the fault audio segment and the normal audio segment are obtained.

[0028] Optionally, the method for obtaining the peak searching height in the signal frequency domain feature comprises the following steps:

[0029]

[0030] wherein, h F is the peak searching height, k h is a constant coefficient, F(k) is the signal frequency domain feature, N m is N / 2, that is, half of the sample sequence length, and k is the frequency component number.

[0031] Optionally, based on the fault audio segment and the normal audio segment, the fault feature sample and the normal feature sample are obtained by the following steps:

[0032] Based on the fault audio segment and the normal audio segment, a peak value list is determined.

[0033] Based on the peak value list, a peak center is determined.

[0034] Based on the peak center, an interval width is obtained.

[0035] The interval width is weighted by using a double exponential weighting method, and a sample frequency domain feature statistical index is obtained.

[0036] Based on the sample frequency domain feature statistical index, the fault feature sample and the normal feature sample are obtained.

[0037] Optionally, the method for obtaining the sample frequency domain feature statistical index by weighting the interval width by using the double exponential weighting method comprises the following steps:

[0038]

[0039] wherein, T(p i ) is the sample frequency domain feature statistical index, W is the statistical interval width, k m is a main frequency amplification constant coefficient, p i is the frequency position of the i-th wave peak, F(p i -W+l) is the amplitude of the p i -W+l-th frequency component, F(p i +l) is the amplitude of the p i +l-th frequency component.

[0040] Compared with the prior art, the present application has the following advantages and technical effects:

[0041] 1.The motor fault monitoring method of the present application uses an automatic method to monitor the motor fault using a sound signal, which can effectively improve the efficiency of motor maintenance. The application of the sound monitoring method makes the equipment installation and layout easier, does not damage the original working structure of the motor, and as a non-contact measurement, the monitoring system is not easily disturbed by the high voltage of the primary side.

[0042] 2.The present application locates the main frequency band of the fault audio through an automatic method, which is more accurate and representative than the traditional voiceprint recognition such as MFCC (Mel Frequency Cepstral Coefficient) feature in extracting fault features, and the feature quantity parameters used for training the model are less and the calculation amount is smaller.

[0043] 3.The present application determines the feature extraction interval by analyzing the typical fault audio frequency domain signal, has strong anti-interference ability to the occasional noise located in other frequency domains in the field, and the identification result is more accurate.

[0044] 4.The feature extraction method of the present application considers the influence of slight shift of the fault audio frequency domain, and the extraction method amplifies the energy of the frequency domain center and considers the effect of the energy near it.

[0045] 5.The method of the present application is simple and easy to implement, does not need prior electrical or mechanical model, and only needs a small amount of data samples.

[0046] 6.The present application uses a machine learning method to automatically establish a fault diagnosis model, and according to the different monitoring objects required, the process parameters can be adjusted in advance during the model establishment process, so that the monitoring accuracy is more optimal, and no other artificial intervention is needed in the actual application. DETAILED DESCRIPTION

[0047] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and their

[0048] Figure 1 is a motor fault monitoring method flow chart based on key audio recognition of the present application;

[0049] Figure 2 is a peak recognition result graph of the present application;

[0050] Figure 3 is a weighted result schematic diagram of different schemes of the present application;

[0051] Figure 4 is a normal audio and fault audio feature comparison schematic diagram of the present application;

[0052] Figure 5 is a test result schematic diagram of the present application. DETAILED DESCRIPTION

[0053] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0055] The present embodiment proposes a motor fault monitoring method based on key audio recognition, as shown in Figure 1 The present embodiment proposes a motor fault monitoring method based on key audio recognition, as shown in

[0056] Obtain motor audio samples;

[0057] Based on the motor audio samples, obtain signal frequency domain features;

[0058] Locate the fault audio segment and the normal audio segment in the signal frequency domain features;

[0059] Based on the fault audio segment and the normal audio segment, obtain fault feature samples and normal feature samples;

[0060] Train the SVM support vector machine using the fault feature samples and the normal feature samples to obtain a fault monitoring model, which is used for motor fault monitoring.

[0061] Specifically, when a motor occurs some faults, it often emits specific fault sound signals. For example, an industrial site auxiliary cooling motor will emit obvious high-frequency fault sound when it occurs mechanical faults such as bearing damage, and experienced workers can often use this point to judge whether the motor state is normal by listening to the sound on site. This embodiment uses this principle to propose a motor fault monitoring method based on key audio recognition, uses a sound sensor to collect motor running sound in real time, uses normal audio and fault audio to establish training samples, extracts audio features, and trains a machine learning algorithm model to realize automatic real-time diagnosis of motor faults and improve motor maintenance efficiency.

[0062] Further, based on the motor audio samples, the signal frequency domain features include:

[0063] Normalizing the motor audio samples to obtain normalized audio samples;

[0064] Signal filtering processing is performed on the normalized audio samples to obtain filtered audio samples;

[0065] Based on the filtered audio samples, obtain signal frequency domain features.

[0066] Specifically, the motor sound signal sample is collected:

[0067] The sound signal collection device is used to record the motor to be measured, which should be installed near the motor to be measured and can effectively collect the sound signal emitted by the motor running. In order to effectively collect the high-frequency fault sound of the motor, the collection device should have a high sampling frequency. In this example, the sampling frequency f s = 16 kHz.

[0068] The original signal obtained by sampling is marked with a sampling sequence number and recorded as:

[0069] x′ i = x′1, x′2, x′3,..., x′ N , i = 1, 2, 3,..., N

[0070] In the formula, i is the sampling sequence number, N is the sample length of the sampling sequence sample, and the sample length should be consistent in the implementation process. In this example, each sample is 1 s long, and under the condition of 16 kHz sampling frequency, N = 16000.

[0071] Normalization preprocessing of sound signal:

[0072] During the sound monitoring process, the distance and direction of the collection device are different, which will cause the decibel value of the collected sound signal to be different. In order to remove this influence, the signal should be normalized. The maximum positive and negative values are taken as the standard, and the scale is taken to the interval of -1~1.

[0073] The formula is:

[0074]

[0075] In the formula, i is the sampling sequence number, x′ i is the original sound signal sample, and x″ i is the output sequence of this step.

[0076] Filtering of signal:

[0077] In the actual application of the monitoring system, there are many environmental noise interferences in the field, such as cooling fan noise, and through the previous research, it is found that the specific fault sound is mainly in the form of high frequency. Therefore, the signal is filtered.

[0078] Because the motor fault sound is usually in the form of high frequency, the pre-emphasis filtering is performed in this scheme, and the formula is:

[0079] x i = x″ i - c1x″ i-1 , i = 1, 2, 3,..., N

[0080] wherein i is a sample sequence number, x i is an output sequence of the pre-processed normalized signal, x i is an output sequence of the pre-emphasis of this step. c1 is a pre-emphasis filter coefficient, which should be in the range of 0

[0081] Further, based on the filtered audio samples, a method for obtaining a signal frequency domain feature is as follows:

[0082] performing fast Fourier transform on the filtered audio samples to obtain a frequency complex spectrum;

[0083] based on the frequency complex spectrum, obtaining a sample frequency domain feature;

[0084] based on the sample frequency domain feature, obtaining a signal frequency domain feature.

[0085] Specifically, the signal frequency domain feature is calculated as follows:

[0086] First, performing fast Fourier transform (FFT) on the signal:

[0087]

[0088] wherein x i is the output signal after filtering in the previous step, k is a frequency component number, i is a sample sequence number, N is a sample sequence length, X(k) is a kth frequency component complex spectrum, and j is an imaginary unit.

[0089] Due to the symmetry of Fourier transform, the first half of X(k) is taken as the sample frequency domain feature, and the amplitude of each frequency component of the signal sample is calculated as follows:

[0090] F(k) = |X(k)|, k = 0, 1, 2, 3,..., N m

[0091] wherein N m is an integer of N / 2.

[0092] Further, positioning the fault audio segment and the normal audio segment in the signal frequency domain feature includes:

[0093] obtaining the minimum horizontal distance and the peak height of adjacent peaks in the signal frequency domain feature;

[0094] performing automatic peak searching based on the minimum horizontal distance and the peak height of adjacent peaks to obtain the fault audio segment and the normal audio segment.

[0095] Specifically, the fault audio segment is positioned as follows:

[0096] Because the fault sound and the long-term stable existing environmental noise in the field mainly reflect in several fixed frequency bands, and the field occasional noise interference basically presents other frequency bands.

[0097] In order to effectively locate the frequency domain in which the fault sound mainly reflects several frequency bands, first, collect typical fault sound frequency samples, and ensure that the motor is actually applied in the field, and there is a continuous background noise. The sample is obtained by recording the fault motor, and the signal frequency domain feature F(k) is obtained by processing the sample according to the preceding steps 1-4.

[0098] The peak value list is obtained by automatically searching the peak of F(k).

[0099] In order to effectively identify the higher frequency energy, the minimum horizontal distance between two adjacent peaks is not less than d F , and the peak value height is not less than h F , in this example, d F =100.

[0100]

[0101] In the formula, k h is a constant coefficient, the greater the value, the fewer the identified wave peaks, and in this example, k h =6.

[0102] The peak value identification is applied to obtain the main frequency band of the fault sound frequency, that is, the subscript corresponding to F(k), and the subscript is returned in an array form, denoted as:

[0103] p i =p1,p2,...,p M ,i=1,2,...,M

[0104] In the formula, M is the number of identified frequency domain peaks, p i is the frequency position of the i-th wave peak, and the identification result is shown in Figure 2 .

[0105] Further, based on the fault sound frequency band and the normal sound frequency band, the fault feature sample and the normal feature sample are obtained, including:

[0106] Based on the fault sound frequency band and the normal sound frequency band, the peak value list is determined;

[0107] Based on the peak value list, the peak center is determined;

[0108] Based on the peak center, the interval width is obtained;

[0109] The interval width is weighted by using a double exponential weighting method to obtain a sample frequency domain feature statistical index;

[0110] Based on the sample frequency domain feature statistical index, obtain the fault feature sample and the normal feature sample.

[0111] Specifically, sample fault feature extraction:

[0112] ①For the segmented audio sample, the fault feature extraction process is as follows: first, the sample signal is processed according to steps 1-4 to obtain the sample frequency domain feature F(k).

[0113] ②Combine the main frequency band p i =p1,p2,...,p M extracted in step 5, and statistically analyze the amplitude of each peak value position frequency band.

[0114] ③Select the statistical interval width: in actual field, the fault noise presents characteristics in a certain frequency band, which is not absolutely fixed, but may exist in a certain interval, and there is energy out-of-band radiation. Therefore, when statistically analyzing the characteristics, the peak value should be determined as the center, and the statistical interval width W should be selected as 100 in this example.

[0115] ④Set the weight of different positions in the statistical interval: according to the characteristics of the fault audio, the frequency domain signal is most representative at the positioned peak value. When statistically analyzing, the frequency domain amplitude in the interval should not be simply added, but the peak value should have the maximum gain, and the gain at the left and right ends of the interval should be the lowest. The weight of different positions should be reasonably configured.

[0116] In this step, there are three weighting schemes:

[0117] One is step weighting:

[0118] That is, the frequency domain [p i -W, p i +W] interval is divided into 2n+1 segments, each with a width of The center segment has the highest weight k w , and the weight decreases step by step to the two sides, and the weight of the two edge segments is The weight of each segment closer to the center m increases by The weight of the center segment is the highest, which is

[0119] Two is linear weighting:

[0120] That is, in the frequency domain [p i -W, p i +W] interval, the weight of different positions increases linearly, and the weight of the edge is The weight of the center peak value is k wThis scheme can be regarded as a special case of the ladder weighting, i.e. the ladder weighting with W=n.

[0121] Thirdly, the exponential weighting:

[0122] The exponential curve is used to increase or decrease in geometric progression. The double exponential curve is used to reasonably weight the two sides of the frequency domain [p i -W, p i +W] interval. In order to avoid the large difference in order of magnitude, the exponential coefficient is reasonably adjusted, and k m is set as the constant coefficient of the main frequency amplification.

[0123] Where p i is the center of the peak value, and the weight value is e The edge of the interval is p i -W and p i +W, and the weight value is e 0 =1, k m The larger the value, the higher the weight value of the amplitude at the main frequency band. In this example, k m =2.

[0124] The edge is the 0th position, and the center is the Wth position. The frequency weighting coefficient of the jth position close to the center is

[0125] The schematic diagram of the three schemes is shown in Figure 3 :

[0126] ⑤Statistical index calculation: Because the exponential curve has more prominent intermediate frequency characteristics, combined with the frequency distribution characteristics of motor fault noise, the exponential weighting scheme is selected, and at the peak value of the corresponding frequency domain p i , the sample frequency domain feature statistical index formula is defined as:

[0127]

[0128] In the formula, p i -W+j<0, substitute F(p i -W+j)=0, p i +j>N m , substitute F(p i +j)=0.

[0129] ⑥Loop calculation: the statistical index calculation of step ⑤ is performed on all fault frequency bands p i , and the sample fault feature sequence is obtained, denoted as:

[0130] T=T(p1), T(p2),..., T(p M ).

[0131] The extracted normal audio and fault audio feature comparison example is shown in Figure 4

[0132] Machine learning data set is made:

[0133] Actual on-site audio collection is performed on the measured motor, and normal motor audio and fault motor audio under the on-site environment are recorded respectively, audio data is segmented and processed, and a large number of audio samples can be obtained.

[0134] The method in step 6 is applied to extract all sample feature sequences as machine learning feature quantity inputs; normal samples are marked as 0, and fault samples are marked as 1 as corresponding result labels.

[0135] Using the established machine learning data set, the SVM support vector machine algorithm is applied to train the fault diagnosis model, and the trained algorithm model is stored for motor fault monitoring.

[0136] Through test test, in the SVM algorithm, when the penalty coefficient is set to 1, the linear kernel function, the polynomial kernel function and the Gaussian kernel function all have very excellent performance, in the test data set, the diagnosis accuracy is very high, up to 100%, using audio samples for test, the diagnosis result is shown in Figure 5

[0137] In practical application, according to the difference of audio features and the existence of unknown interference, the diagnosis accuracy may decrease, and targeted parameter adjustment can be made according to the scheme of the application to obtain better effect.

[0138] Real-time fault monitoring of motor:

[0139] After the fault diagnosis model is established, it can be applied to the field for real-time fault monitoring of motor, only real-time audio of motor is collected through sensor, audio sample is input to host computer, host computer performs feature transformation on the collected audio according to the method in step 6, and then the transformed features are input to the fault diagnosis model, so that the motor state monitoring result can be obtained.

[0140] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the application, which should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.​​

Claims

1. A method of motor fault monitoring based on key audio recognition, characterized in that, The method comprises the following steps: acquiring a motor audio sample; acquiring a signal frequency domain feature based on the motor audio sample; locating a fault audio segment and a normal audio segment in the signal frequency domain feature; locating a fault audio segment and a normal audio segment in the signal frequency domain feature comprises: acquiring a minimum horizontal distance and a peak height of adjacent peaks in the signal frequency domain feature; automatically performing peak searching based on the minimum horizontal distance and the peak height of the adjacent peaks to acquire the fault audio segment and the normal audio segment; the method for acquiring the peak height in the signal frequency domain feature is: wherein, is the peak height, is the constant coefficient, is the signal frequency domain feature, is N / 2 rounded, i.e. half the length of the sample sequence, and k is the frequency component number; acquiring a fault feature sample and a normal feature sample based on the fault audio segment and the normal audio segment; training an SVM support vector machine using the fault feature sample and the normal feature sample to acquire a fault monitoring model, which is used for motor fault monitoring.

2. The key audio recognition based motor fault monitoring method of claim 1, wherein, acquiring a signal frequency domain feature based on the motor audio sample comprises: performing normalization preprocessing on the motor audio sample to acquire a normalized audio sample; performing signal filtering processing on the normalized audio sample to acquire a filtered audio sample; acquiring the signal frequency domain feature based on the filtered audio sample.

3. The key audio recognition based motor fault monitoring method of claim 2, wherein, The method for acquiring the signal frequency domain feature based on the filtered audio sample is: performing fast Fourier transform on the filtered audio sample to acquire a frequency complex spectrum; acquiring a sample frequency domain feature based on the frequency complex spectrum; acquiring the signal frequency domain feature based on the sample frequency domain feature.

4. The key audio recognition based motor fault monitoring method of claim 3, wherein, The method for performing fast Fourier transform on the filtered audio sample to acquire a frequency complex spectrum is: wherein is the filtered output signal, k is the frequency component number, i is the sample sequence number, N is the sample sequence length, is the kth frequency component complex spectrum, j is the imaginary unit.

5. The key audio recognition based motor fault monitoring method of claim 1, wherein, acquiring a fault feature sample and a normal feature sample based on the fault audio segment and the normal audio segment comprises: determining a peak list based on the fault audio segment and the normal audio segment; determining a peak center based on the peak list; acquiring an interval width based on the peak center; weighting the interval width using a double exponential weighting method to acquire a sample frequency domain feature statistical index; acquiring the fault feature sample and the normal feature sample based on the sample frequency domain feature statistical index.

6. The key audio recognition based motor fault monitoring method of claim 5, wherein, The method for weighting the interval width using a double exponential weighting method to acquire a sample frequency domain feature statistical index is: wherein, is a sample frequency domain feature statistical indicator, is a statistical interval width, is a main frequency amplification constant coefficient, is a frequency position of the i-th peak, is is a frequency component amplitude at is is a frequency component amplitude at

Images