Motor fault monitoring method based on key audio recognition

Abstract

The invention discloses a motor fault monitoring method based on key audio recognition. The method comprises the following steps: acquiring a motor audio sample; acquiring a signal frequency domain feature based on the motor audio sample; positioning a fault audio segment and a normal audio segment in the signal frequency domain features; based on the fault audio segment and the normal audio segment, obtaining a fault feature sample and a normal feature sample; and training an SVM (Support Vector Machine) by using the fault feature sample and the normal feature sample to obtain a fault monitoring model which is used for motor fault monitoring. According to the invention, the sound sensor is used to collect the motor operation sound in real time, the normal audio and the fault audio are used to establish the training sample, the audio features are extracted, the machine learning algorithm model is trained, the motor fault is automatically diagnosed in real time, and the motor maintenance efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault monitoring, and in particular relates to a motor fault monitoring method based on key audio recognition. Background Art

[0002] Motors are prone to failure during long-term uninterrupted operation. For example, auxiliary heat dissipation motors in industrial sites often run 24 hours a day due to the nature of the equipment. If they are not maintained in time after a failure, the equipment will fail and cause serious consequences.

[0003] Traditionally, motor fault monitoring has been done manually and periodically, while automated methods often use motor current monitoring or motor vibration signal monitoring.

[0004] Monitoring methods based on current or vibration signals require advance point selection and placement of corresponding sensors, combined with electrical or mechanical model analysis of the motor to establish a fault diagnosis algorithm. The implementation process is relatively complex, and when implemented in an already installed and working motor, there are problems such as difficulty in installing and placing sensors.

[0005] The manual inspection method based on sound signals requires manual on-site inspection, which is labor-intensive, inefficient, and unable to detect faults in a timely and effective manner.

[0006] Existing automatic diagnosis methods based on sound signals use relatively conventional feature extraction methods, often applying voiceprint features such as Mel-frequency cepstral coefficients, which cover all time or frequency domain features. In practical applications, especially in industrial sites with complex sound environments, sudden noise interference not considered during model training, such as occasional human voices, bird calls, thunder, and short-term noise from on-site operating equipment, will have a significant impact on sample characteristics. These methods require more comprehensive fault audio collection to ensure effective fault identification. Alternatively, manual selection is required during model training to determine the typical fault audio frequency domain to further ensure accuracy. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a motor fault monitoring method based on key audio recognition, which can utilize an automated method to automatically measure the sound signals emitted by the motor using a sound acquisition device, and process and analyze the signals in real time, thereby improving the efficiency of motor fault monitoring and reducing manual workload.

[0008] To achieve the above objectives, the present invention provides a motor fault monitoring method based on key audio recognition, comprising:

[0009] Get motor audio samples;

[0010] Based on the motor audio sample, obtaining signal frequency domain features;

[0011] Locating a fault audio segment and a normal audio segment in the signal frequency domain characteristics;

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

[0013] The fault feature samples and the normal feature samples are used to train an SVM support vector machine to obtain a fault monitoring model, and the fault monitoring model is used to perform motor fault monitoring.

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

[0015] Performing normalization preprocessing on the motor audio sample to obtain a normalized audio sample;

[0016] Performing signal filtering on the normalized audio samples to obtain filtered audio samples;

[0017] The signal frequency domain features are obtained based on the filtered audio samples.

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

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

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

[0021] Based on the sample frequency domain features, the signal frequency domain features are obtained.

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

[0023]

[0024] Among them, 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 complex spectrum of the kth frequency component, and j is the imaginary unit.

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

[0026] Obtaining the minimum horizontal distance between adjacent peaks and the peak search height in the signal frequency domain characteristics;

[0027] Automatic peak search is performed based on the minimum horizontal distance between adjacent peaks and the peak search height to obtain the fault audio segment and the normal audio segment.

[0028] Optionally, a method for obtaining the peak height in the signal frequency domain feature is:

[0029]

[0030] Among them, h F is the peak height, k h is a constant coefficient, F(k) is the frequency domain characteristic of the signal, N m N / 2 is rounded up to half the length of the sample sequence, and k is the frequency component number.

[0031] Optionally, obtaining a fault feature sample and a normal feature sample based on the fault audio segment and the normal audio segment includes:

[0032] Determining a peak list based on the fault audio segment and the normal audio segment;

[0033] Based on the peak list, determining a peak center;

[0034] Based on the peak center, obtaining the interval width;

[0035] The interval width is weighted using a double exponential weighting method to obtain a statistical index of sample frequency domain characteristics;

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

[0037] Optionally, a method for weighting the interval width using a double exponential weighting method to obtain a statistical index of sample frequency domain characteristics is as follows:

[0038]

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

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

[0041] 1. This invention utilizes an automated method, using sound signals to monitor motor faults, effectively improving motor maintenance efficiency. The application of sound monitoring simplifies equipment installation and deployment, does not damage the motor's original working structure, and as a non-contact measurement method, the monitoring system is less susceptible to interference from primary-side high voltage.

[0042] 2. The present invention locates the main frequency band of the fault audio through an automated method. Compared with traditional voiceprint recognition such as MFCC (Mel-frequency Cepstral Coefficient) features, the fault feature extraction is more accurate and representative. The training model uses fewer feature parameters and has a smaller amount of calculation.

[0043] 3. The present invention determines the feature extraction interval by analyzing the typical fault audio frequency domain signal. In application, it has a strong anti-interference ability against occasional noise in other frequency domains that appears on site, and the recognition result is more accurate.

[0044] 4. The feature extraction method proposed in the present invention takes into account the influence of slight offset in the fault audio domain. The extraction method has the effect of amplifying the energy in the center of the frequency domain and considering the energy nearby.

[0045] 5. The method of the present invention is simple and easy to implement, does not require a priori electrical or mechanical models, only requires a small amount of data samples, and is easy to implement.

[0046] 6. The present invention uses a machine learning method to automatically establish a fault diagnosis model. According to different monitoring objects required, process parameters can be pre-adjusted during the model establishment process to improve the monitoring accuracy, and no other human intervention is required in actual applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0048] Figure 1 is a flow chart of a motor fault monitoring method based on key audio recognition according to an embodiment of the present invention;

[0049] Figure 2 This is a peak recognition result diagram according to an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of weighted results of different solutions of an embodiment of the present invention;

[0051] Figure 4 2. This is a schematic diagram comparing the characteristics of normal audio and faulty audio according to an embodiment of the present invention;

[0052] Figure 5 2 is a schematic diagram of test results of an embodiment of the present invention. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

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

[0055] This embodiment proposes a motor fault monitoring method based on key audio recognition, such as Figure 1 As shown, the specific steps include:

[0056] Get motor audio samples;

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

[0058] Locating fault audio segments and normal audio segments in the signal frequency domain characteristics;

[0059] Obtaining a fault feature sample and a normal feature sample based on the fault audio segment and the normal audio segment;

[0060] The fault feature samples and normal feature samples are used to train the SVM support vector machine to obtain a fault monitoring model, which is used to monitor motor faults.

[0061] Specifically, motors often emit specific fault sound signals when certain faults occur. For example, industrial auxiliary cooling motors produce distinct high-frequency fault sounds when mechanical faults such as bearing damage occur. Experienced workers can often use this to determine whether the motor is functioning properly by listening to the sound on-site. This embodiment utilizes this principle to propose a motor fault monitoring method based on key audio recognition. This method uses a sound sensor to collect motor operating sounds in real time, and uses normal and faulty audio to create training samples, extract audio features, and train a machine learning algorithm model. This method enables automated, real-time diagnosis of motor faults, improving motor maintenance efficiency.

[0062] Furthermore, based on the motor audio sample, the signal frequency domain features are obtained including:

[0063] Perform normalization preprocessing on the motor audio sample to obtain the normalized audio sample;

[0064] Performing signal filtering on the normalized audio samples to obtain filtered audio samples;

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

[0066] Specifically, collect motor sound signal samples:

[0067] Use sound signal acquisition equipment to record the motor under test. It should be installed near the motor under test and can effectively collect the sound signals emitted by the motor. In order to effectively collect the high-frequency fault sound of the motor, the acquisition equipment should have a high sampling frequency. In this example, the sampling frequency f s =16kHz as an example.

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

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

[0070] Where i is the sampling sequence number, and N is the sample length of the sampling sequence. The sample length should remain consistent during implementation. In this example, each sample segment is 1s long. Under the 16Khz sampling frequency, N = 16000.

[0071] Normalization preprocessing of sound signals:

[0072] During the sound monitoring process, the distance and direction of the collection equipment are different, which will lead to differences in the decibel value of the collected sound signal. In order to eliminate this influence, the signal should be normalized and scaled to the range of -1 to 1 based on the maximum positive and negative values.

[0073] The formula is:

[0074]

[0075] Where i is the sampling sequence number, x′ i is the original sound signal sample, x″ i Preprocess the output sequence for this step.

[0076] Signal filtering:

[0077] In the actual application of the monitoring system on site, there is a lot of interference from environmental noise, such as cooling fan noise, and preliminary research has found that specific fault sounds are mainly high-frequency, so the signal is filtered.

[0078] Because motor fault sounds are usually high-frequency, this solution performs pre-emphasis filtering using the formula:

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

[0080] Where i is the sampling sequence number, x″ i is the output sequence of the preprocessed normalized signal, x i is the output sequence of pre-emphasis in this step. c1 is the pre-emphasis filter coefficient, and its value range should be 0<c1<1. In this embodiment, c1=0.97.

[0081] Furthermore, based on the audio samples after filtering, the method for obtaining the signal frequency domain characteristics is as follows:

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

[0083] Based on the frequency complex spectrum, obtain the frequency domain characteristics of the sample;

[0084] Based on the sample frequency domain features, the signal frequency domain features are obtained.

[0085] Specifically, calculate the signal frequency domain characteristics:

[0086] First, perform a Fast Fourier Transform (FFT) on the signal:

[0087]

[0088] Where 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 complex spectrum of the kth frequency component, and j is the 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:

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

[0091] Where N m Round to N / 2.

[0092] Furthermore, the fault audio segment and the normal audio segment in the frequency domain characteristics of the positioning signal include:

[0093] Obtain the minimum horizontal distance between adjacent peaks and the peak height in the signal frequency domain characteristics;

[0094] Automatic peak retrieval is performed based on the minimum horizontal distance between adjacent peaks and the peak retrieval height to obtain fault audio segments and normal audio segments.

[0095] Specifically, locate the main frequency band of the fault sound:

[0096] Because fault sounds and long-term stable environmental noise on site are mainly reflected in several fixed frequency bands, while occasional noise interference on site is basically presented in other frequency bands.

[0097] To effectively locate the frequency bands in which the fault sound is mainly reflected in the frequency domain, first collect typical fault audio samples. The collection should be carried out at the actual application site of the motor and in the presence of continuous background noise. Record the faulty motor to obtain samples, and process the samples according to the aforementioned steps 1-4 to obtain the signal frequency domain features F(k).

[0098] Perform automatic peak search on F(k) to obtain a peak list.

[0099] In order to effectively identify higher frequency domain energy, the minimum horizontal distance between two adjacent peaks is defined to be no less than d F , the height of the peak to be found is not less than h F In this example, take d F =100;

[0100]

[0101] Where k h is a constant coefficient. The larger the value, the fewer peaks are identified. In this example, k h =6.

[0102] Apply peak recognition to obtain the main frequency bands of the fault audio frequency domain characteristics, that is, the corresponding subscripts in F(k). The subscripts are returned in array form and recorded as:

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

[0104] Where M is the number of frequency domain peaks identified, p i That is the frequency position of the i-th peak, the recognition result is as follows Figure 2 shown.

[0105] Furthermore, obtaining a fault feature sample and a normal feature sample based on the fault audio segment and the normal audio segment includes:

[0106] Determine a peak list based on the fault audio segment and the normal audio segment;

[0107] Based on the peak list, determine the peak center;

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

[0109] The double exponential weighting method is used to weight the interval width to obtain the statistical indicators of the sample frequency domain characteristics;

[0110] Based on the statistical indicators of the sample frequency domain characteristics, fault feature samples and normal feature samples are obtained.

[0111] Specifically, sample fault feature extraction:

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

[0113] ② Combine the main frequency band p of the fault audio frequency domain features extracted in step 5 i =p1,p2,...,p M , and statistics are performed on the amplitude of the frequency band where each peak is located.

[0114] ③ Select the statistical interval width: In actual field conditions, the characteristics of fault noise in a certain frequency band are not absolutely fixed. Instead, there may be slight deviations within a certain interval, and there may be out-of-band radiation of energy. Therefore, when statistically analyzing the characteristics, the peak value should be taken as the center, and statistics should be performed within a reasonable interval on both sides of it. W is selected as the statistical interval width. In this example, W = 100 is selected.

[0115] ④ Set weights for different positions within the statistical interval: According to the characteristics of the fault audio, its frequency domain signal is most representative at the located peak. When performing statistics, the frequency domain amplitude within the interval cannot be simply accumulated directly. Instead, the gain should be maximum at the peak, and the gain should be lowest at the left and right ends of the interval. The weights at different positions should be reasonably configured.

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

[0117] One is step weighting:

[0118] About frequency domain [p i -W,p i +W] interval, divided into 2n+1 segments, each segment width is The center segment with the highest weight is k w , gradually decreasing to both sides, the weights on both sides of the edge are As it approaches the middle m levels, its weight increases The most central section The interval weight is the highest,

[0119] The second is linear weighting:

[0120] That is, in the frequency domain [p i -W,p i +W] interval, the weights at different positions increase linearly, and the weights at the edge are The weight at the central peak is k w, this scheme can be regarded as a special case of ladder weighting, namely, ladder weighting with W=n.

[0121] The third is index weighting:

[0122] Taking advantage of the characteristic of exponential curve that increases or decreases in geometric series, a double exponential curve is used to analyze the frequency domain [p i -W,p i +W] interval on both sides are reasonably weighted, and in order to avoid a large gap in the order of magnitude, the exponential coefficient is reasonably adjusted, and k is set m A constant factor for main frequency amplification.

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

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

[0125] The difference between the three scheme indexes is shown in the following diagram: Figure 3 As shown:

[0126] ⑤ Calculation of statistical indicators: Since the exponential curve has the characteristic of highlighting the intermediate frequency, combined with the frequency distribution characteristics of the motor fault noise, the exponential weighting scheme is selected. i At the corresponding frequency domain peak, the statistical indicator formula of the sample frequency domain characteristics is defined as:

[0127]

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

[0129] ⑥ Cyclic calculation: for all fault frequency bands p i By calculating the statistical indicators in step ⑤, we can obtain the fault characteristic sequence of this sample, which is recorded as:

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

[0131] The comparison example of extracted normal audio and fault audio features is as follows: Figure 4 As shown:

[0132] Creating a machine learning dataset:

[0133] Actual on-site audio collection is performed on the motor under test. The normal motor audio and the faulty motor audio in the on-site environment are recorded separately. The audio data is segmented and processed to obtain a large number of audio samples.

[0134] Apply the method in step 6 to extract all sample feature sequences as machine learning feature input; normal samples are marked as 0 and fault samples are marked as 1 as their corresponding result label input.

[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] After experimental testing, in the SVM algorithm, when the penalty coefficient is set to 1, the three kernel functions of linear kernel function, polynomial kernel function and Gaussian kernel function all have very excellent performance. In the test data set, the diagnosis accuracy is extremely high, up to 100%. The diagnosis results are as follows: Figure 5 shown.

[0137] In actual applications, the diagnostic accuracy may decrease depending on the different audio characteristics and the presence of unknown interference. Targeted parameter adjustments can be made according to the solution of the present invention to achieve better results.

[0138] Real-time fault monitoring of motors:

[0139] After the fault diagnosis model is established, it can be applied on-site to perform real-time fault monitoring of the motor. It only needs to collect motor audio through sensors in real time and input the audio samples into the host computer. The host computer performs feature transformation on the collected audio according to the method in step 6, and then inputs the transformed features into the fault diagnosis model to obtain the motor status monitoring results.

[0140] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A motor fault monitoring method based on key audio recognition, characterized in that: include: Get motor audio samples; Based on the motor audio sample, obtaining signal frequency domain features; Locating a fault audio segment and a normal audio segment in the signal frequency domain characteristics; Based on the fault audio segment and the normal audio segment, obtaining a fault feature sample and a normal feature sample; The fault feature samples and the normal feature samples are used to train an SVM support vector machine to obtain a fault monitoring model, and the fault monitoring model is used to perform motor fault monitoring.

2. The motor fault monitoring method based on key audio recognition according to claim 1 is characterized in that: Based on the motor audio sample, obtaining the signal frequency domain features includes: Performing normalization preprocessing on the motor audio sample to obtain a normalized audio sample; Performing signal filtering on the normalized audio samples to obtain filtered audio samples; The signal frequency domain features are obtained based on the filtered audio samples.

3. The motor fault monitoring method based on key audio recognition according to claim 2 is characterized in that: The method for obtaining the frequency domain characteristics of the signal based on the filtered audio sample is: Performing a fast Fourier transform on the filtered audio samples to obtain a frequency complex spectrum; Acquiring a sample frequency domain feature based on the frequency complex spectrum; Based on the sample frequency domain features, the signal frequency domain features are obtained.

4. The motor fault monitoring method based on key audio recognition according to claim 3 is characterized in that: The method of performing fast Fourier transform on the audio samples after the filtering process to obtain the complex frequency spectrum is as follows: Among them, x i is the output signal after filtering, k is the frequency component number, i is the sampling sequence number, N is the sample sequence length, X(k) is the complex spectrum of the kth frequency component, and j is the imaginary unit.

5. The motor fault monitoring method based on key audio recognition according to claim 4 is characterized in that: Locating the fault audio segment and the normal audio segment in the signal frequency domain feature includes: Obtaining the minimum horizontal distance between adjacent peaks and the peak search height in the signal frequency domain characteristics; Automatic peak search is performed based on the minimum horizontal distance between adjacent peaks and the peak search height to obtain the fault audio segment and the normal audio segment.

6. The motor fault monitoring method based on key audio recognition according to claim 5 is characterized in that: The method for obtaining the peak height in the signal frequency domain characteristics is: Among them, h F is the peak height, k h is a constant coefficient, F(k) is the frequency domain characteristic of the signal, N m N / 2 is rounded up to half the length of the sample sequence, and k is the frequency component number.

7. The motor fault monitoring method based on key audio recognition according to claim 6, characterized in that: Acquiring a fault feature sample and a normal feature sample based on the fault audio segment and the normal audio segment includes: Determining a peak list based on the fault audio segment and the normal audio segment; Based on the peak list, determining a peak center; Based on the peak center, obtaining the interval width; The interval width is weighted using a double exponential weighting method to obtain a statistical index of sample frequency domain characteristics; Based on the sample frequency domain feature statistical indicators, the fault feature samples and normal feature samples are obtained.

8. The motor fault monitoring method based on key audio recognition according to claim 7 is characterized in that: The method of weighting the interval width using the double exponential weighting method to obtain the statistical index of the sample frequency domain characteristics is as follows: Among them, T(p i ) is the statistical index of the sample frequency domain characteristics, W is the statistical interval width, k m is the main frequency amplification constant coefficient, p i is the frequency position of the i-th peak, F(p i -W+l) is p i -W+l frequency component amplitude, F(p i +l) is p i +l is the amplitude of the frequency component at the location.

Images