A signal extraction and classification method based on sound

Abstract

The present invention provides a signal extraction and classification method based on sound, comprising the following steps: step 1, collecting sound signals under different states of a motor; step 2, preprocessing the collected sound signals; step 3, extracting fault features using the sound signals preprocessed in step 2, obtaining a time-frequency spectrum of the sound and constructing a data set; step 4, building a CNN model, inputting the data set constructed in step 3 into the model for training and testing; step 5, diagnosing and classifying the motor rotor broken bar fault using the trained model and outputting the result; the present invention can effectively realize the diagnosis and classification of the motor rotor broken bar fault, provides a new research idea for the diagnosis of the motor rotor broken bar fault, and utilizes the effectiveness of the fault diagnosis and classification method combining short-time Fourier transform and deep learning to verify that the accuracy rate reaches 100%.

Description

Technical Field

[0001] The present invention belongs to the technical field of sound signal diagnosis and classification, and in particular relates to a sound-based signal extraction and classification method. Background Art

[0002] The MCSA method is a motor rotor broken bar fault diagnosis method based on stator current signal data. It is the most commonly used method and is also considered the most effective method. In addition to the fault information contained in the stator current signal, the sound signal when the motor is running smoothly also contains a lot of fault information, but the expression form is different from that of the stator current signal. Sound can reflect the running state and healthy operating conditions of the motor. At the same time, sound signals are also one of the data that is easiest to collect through non-contact methods. Therefore, sound signals are also commonly used data signals for fault diagnosis by scholars. At present, most of them use stator current signals to diagnose broken rotor bar faults. There are few studies on rotor broken bar faults using sound signals, and there is no research on motor rotor broken bar fault diagnosis using sound signals combined with deep learning. Summary of the invention

[0003] The present invention takes the sound signal of a motor running as a research object and proposes a signal extraction and classification method based on sound.

[0004] To achieve the above object, the present invention adopts the following technical solution, and the specific steps are:

[0005] Step 1, collecting sound signals of the motor under different states;

[0006] Step 2, preprocessing the collected sound signal;

[0007] Step 3, using the sound signal preprocessed in step 2 to extract fault features, obtain the time-frequency spectrum of the sound and construct a data set;

[0008] Step 4: Build a CNN model and input the data set constructed in step 3 into the model for training and testing;

[0009] Step 5: Use the trained model to diagnose and classify the motor rotor broken bar fault and output the results.

[0010] Preferably, the sound signals collected in different states of the motor in step 1 are divided into two rounds. The first round is to collect the sound signals of the healthy motor when it is running smoothly for 20 minutes, and the second round is to collect the sound signals of the motor when one rotor bar is broken and the sound signals of the motor when two rotor bars are broken for 20 minutes each.

[0011] Preferably, the specific process of the second round of sound signal collection is to disassemble the motor, separate the motor rotor, use an electric drill to drill a hole and cut off a rotor guide bar, reassemble the motor, and collect the sound signal of a broken rotor bar fault after the motor is powered on and runs smoothly, and collect sound data for 20 minutes; after the motor is powered off and cooled, disassemble the motor again, separate the motor rotor, use an electric drill to drill a hole and cut off two rotor guide bars, reassemble the motor, and collect the sound signals of two broken rotor bars fault after the motor is powered on and runs smoothly, and collect sound data for 20 minutes.

[0012] Preferably, the preprocessing of the collected sound signal in step 2 is divided into two steps:

[0013] S1: The collected sound signals of the motor in different states are converted into .wav format;

[0014] S2: Resample the entire sound signal at equal time intervals, and set the resampling time interval to 60 seconds.

[0015] Preferably, the fault feature extraction in step 3 is to obtain the time-frequency spectrum diagram under different fault states of the motor rotor broken bar by short-time Fourier transform, specifically, the three sound signals of the motor running healthily, having one rotor broken bar fault and having two rotor broken bars fault are subjected to short-time Fourier transform to extract the fault features and obtain the time-frequency spectrum diagram of the sound.

[0016] Preferably, the data set constructed in step 3 is randomly divided into a training set and a test set, and the ratio of the training set to the test set is 7:3.

[0017] Preferably, the CNN model built in step 4 is an AlexNet model.

[0018] Compared with the existing technology, the fault diagnosis and classification method based on sound signals and deep learning can effectively realize the diagnosis and classification of motor rotor broken bar faults, providing a new research idea for motor rotor broken bar fault diagnosis. The effectiveness of the fault diagnosis and classification method combined with short-time Fourier transform and deep learning is used to verify the accuracy rate of 100%. Compared with the other four CNN models, the AlexNet model has the shortest training time for rotor broken bar fault classification and is fast in fault classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Simulate the broken bar failure scenario for electric drill drilling.

[0020] Figure 2 Figure 2 shows rotor scenarios with different fault states.

[0021] Figure 3Acquire scene graph for signal.

[0022] Figure 4 It is the time domain image of the sound signal.

[0023] Figure 5 It is the frequency domain image of the sound signal.

[0024] Figure 6 It is the sound signal spectrogram.

[0025] Figure 7 Schematic diagram of the fault diagnosis and classification model structure.

[0026] Figure 8 This is a visualization diagram of the fault characteristics.

[0027] Fig. 9 It is the Loss curve graph.

[0028] Fig.10 It is a confusion matrix diagram. DETAILED DESCRIPTION

[0029] The invention will be further described below in conjunction with specific embodiments.

[0030] like Figures 1 to 10 As shown, a sound-based signal extraction and classification method comprises the following steps:

[0031] Step 1, collecting sound signals of the motor under different states;

[0032] Step 2, preprocessing the collected sound signal;

[0033] Step 3, using the sound signal preprocessed in step 2 to extract fault features, obtain the time-frequency spectrum of the sound and construct a data set;

[0034] Step 4: Build a CNN (convolutional neural network) model and input the data set constructed in step 3 into the model for training and testing;

[0035] Step 5: Use the trained model to diagnose and classify the motor rotor broken bar fault and output the results.

[0036] The test site of the present invention is in the 511 electrical laboratory of the engineering building of Qingdao Agricultural University, and the model of the squirrel cage asynchronous motor selected is YX3 80M1-4. The data collection time was from December 17, 2020 to December 25, 2020, and experiments were carried out for data collection both during the day and at night. During the experiment, the average temperature of the electrical laboratory environment was 17°C. During the experiment, the highest temperature reached by the motor was 56.3°C and the lowest temperature was 16.8°C. The conventional parameters of the YX3 80M1-4 squirrel cage three-phase asynchronous motor are shown in Table 1.1.

[0037] Table 1.1 YX3 80M1-4 motor parameter table

[0038]

[0039]

[0040] 1.1 Sound signal collection

[0041] The collection of sound signals is divided into two rounds. The first round is to collect the sound signals of the healthy motor running smoothly. The motor is powered by the three-phase AC power of the electrician's test bench, and is protected by a fuse in the middle and grounded. The sound signal of the healthy motor running smoothly collected 20 minutes of sound data. The second round is to collect the sound signals of the motor with one broken rotor bar failure and the sound signals of the motor with two broken rotor bars. Before collecting, the motor needs to simulate the broken rotor bar failure. To simulate the broken rotor bar failure of the motor, the asynchronous motor needs to be disassembled and assembled, and the rotor bars need to be drilled with an electric drill until the rotor bars are broken. Figure 1 It is the process of drilling and chiseling the rotor using an electric drill.

[0042] The second round of sound signal collection requires disassembly and assembly of the motor, which is a relatively cumbersome process. The specific process is: disassemble the three-phase asynchronous motor, separate the motor rotor, use an electric drill to drill a hole and cut off a rotor bar, reassemble the three-phase asynchronous motor, and after the motor is powered on and runs smoothly, collect the sound signal of a broken rotor bar fault, and collect sound data for 20 minutes; after the motor is powered off and cooled, disassemble the motor again, separate the motor rotor, use an electric drill to drill a hole and cut off two rotor bars, reassemble the motor, and after the motor is powered on and runs smoothly, collect the sound signal of two broken rotor bars fault, and collect sound data for 20 minutes. Schematic diagram of the rotor under different fault conditions is shown below. Figure 2 shown.

[0043] During the signal acquisition process, it was found that the occurrence of a motor rotor bar breakage will not only affect the waveform of the motor stator current signal, but the more directly felt impact is the change in motor temperature. The impact of different motor rotor bar breakage fault states (different numbers of rotor bar breakage faults) on temperature varies greatly. The more rotor bar breakage data there is, the faster the motor temperature rises, and the motor cannot run for a long time. Therefore, the process of collecting sound signals requires real-time monitoring, and the difficulty of collecting sound signals of motors with rotor bar breakage faults is greater than that of healthy motors.

[0044] The sound signal collection under different fault conditions was carried out in the actual laboratory environment. The sound data was recorded by the recorder in the Android phone, and the collected sound data was saved in the .mp4 format. During the collection process, a digital oscilloscope was used to observe the voltage waveform, and a thermometer was used to measure the temperature of the asynchronous motor housing in real time to prevent the motor from being affected by excessive temperature and aggravating the degree of the motor failure. Figure 3 It is an image acquisition method based on on-site sound signals.

[0045] 1.2 Fault feature extraction

[0046] The collected sound signals under different fault states of the motor are non-stationary signals, and the fault features of the sound signals are extracted by using the time-frequency analysis method. The present invention uses the short-time Fourier transform method to extract the features of the rotor broken bar fault of the motor sound signal.

[0047] 1.2.1 Sound signal preprocessing

[0048] Before extracting features from sound signals, it is necessary to preprocess the sound signals. This paper has two steps to preprocess the sound signals.

[0049] (1) Audio signal format conversion

[0050] In the actual on-site environment, the sound signal was collected using the recorder function in the Android phone. The sound signal can only be saved in .mp4 format. However, the prerequisite for converting the sound signal into a spectrogram is to first convert the sound signal into .wav format. Therefore, we uniformly converted the sound signals collected under different fault conditions of the motor into .wav format.

[0051] (2) Resampling the sound signal at equal time intervals

[0052] The sound signal collected in the actual on-site environment is a long-term continuous time domain signal, which needs to be made into multiple samples as the input of artificial intelligence. To this end, we resample the entire sound signal at equal time intervals and set the resampling time interval to 60 seconds.

[0053] 1.2.2 Sound signal indication

[0054] The sound signal directly obtained from the experiment is a time domain signal. In order to better diagnose and classify the fault, the frequency domain image of the sound signal is made. The time domain image and frequency domain image of the sound signal are as follows: Figure 4 and Figure 5 shown.

[0055] Depend on Figure 4 and Figure 5It can be seen that the time domain image and frequency domain image of the sound signal under different motor rotor bar fault states are almost indistinguishable, and the fault features cannot be effectively extracted, so the rotor bar fault cannot be diagnosed. In addition, from the frequency domain image of the sound signal under different motor rotor bar fault states, it can be seen that the sound energy of the motor rotor bar fault is mainly distributed in the medium frequency region below 1500Hz.

[0056] The spectrogram is a time-frequency spectrum obtained through time-frequency transformation, which can simultaneously reflect the time domain and frequency domain information of the sound signal. There are differences in the spectrogram features of the sound signal under different fault conditions, which is the premise of using image processing methods for sound recognition. The spectrogram can reflect the energy distribution of each short period of time at different frequencies. The present invention uses short-time Fourier transform to obtain the time-frequency spectrum of the motor under different fault conditions of broken rotor bars. The three sound signals of healthy motor operation, one broken rotor bar fault and two broken rotor bars fault are subjected to short-time Fourier transform to extract fault features, and the obtained sound spectrogram is as follows: Figure 6 shown.

[0057] The sound spectrograms under different fault conditions are used to construct a data set, in which the training set and the test set are randomly divided in a ratio of 7:3. The detailed information of the constructed data set is shown in Table 1.2.

[0058] Table 1.2 Motor rotor broken bar fault data set

[0059]

[0060]

[0061] 1.3 Fault Diagnosis and Classification Framework

[0062] Deep learning methods have powerful functions and can be trained on text, images, and sounds, but compared to sounds, learning on images has greater advantages. Based on image classification, CNN has great advantages compared to other deep learning methods. This paper extracts fault features from sound signals and proposes a motor rotor broken bar fault diagnosis and classification method based on the combination of short-time Fourier transform and deep learning. First, the acquired sound signal data is preprocessed by converting the storage format and resampling at equal time intervals. Then, the preprocessed sound signal is short-time Fourier transformed to obtain a two-dimensional image, and a data set is constructed and input into the AlexNet model for learning, thereby completing the diagnosis and classification of the motor rotor broken bar fault. Finally, the diagnosis and classification effects of the AlexNet model are compared and analyzed. The structural diagram of the motor rotor broken bar fault diagnosis and classification method is shown in the figure. Figure 7 shown.

[0063] Depend on Figure 7It can be seen that the entire motor rotor broken bar fault diagnosis and classification structure diagram mainly consists of five parts: sound signal acquisition, sound signal data storage format conversion and resampling at equal time intervals, data preprocessing, short-time Fourier transform to obtain time-frequency spectrum, CNN model classification and result analysis. The internal structure parameters of the AlexNet model are shown in Table 1.3.

[0064] Table 1.3 AlexNet model parameters

[0065]

[0066]

[0067] 1.4 Test results and analysis

[0068] 1.4.1 Feature Visualization

[0069] The features extracted from different layers of CNN determine the final accuracy of the model. Generally, visual examination of filter responses is considered to be a suitable method for evaluating model performance. This method can extract low-level and high-level features, thereby having an intuitive response to each layer of the model. Some filters applied in the final convolutional layer of the model The filters of the primary convolutional layer extract color features and directional edges; the filters of the intermediate layers encode simple textures prepared by the combination of color and edges; the final convolutional layer extracts image textures and clear patterns. Figure 8 Shows feature visualization of the first convolutional layer and the fully connected layers of the AlexNet model.

[0070] according to Figure 8 It can be found that the features of the fully connected layer are significantly richer than those of the first layer, which illustrates the advantages of CNN in feature extraction and classification.

[0071] 1.4.2 Result Analysis

[0072] The motor rotor broken bar fault sound spectrogram dataset is input into the AlexNet model for training. The Loss curve and confusion matrix obtained after training are as follows: Fig. 9 and Fig.10 shown.

[0073] The obtained loss curve and confusion matrix show the effectiveness of the fault diagnosis and classification method combining short-time Fourier transform and deep learning, and the verification accuracy rate reaches 100%. To prove the effectiveness of the method, this paper also compares the diagnosis and classification effects of four CNN models: VGG16, GoogLeNet, DenseNet201, and SqueezeNet.

[0074] In order to illustrate the rapidity of model classification, the training time of each model was counted. Table 1.4 shows the training time of each model under the parameters of Batch Size = 32 and MaxEpochs = 40.

[0075] Table 1.4 Model classification time table

[0076]

[0077] As can be seen from Table 1.4, the AlexNet model has the shortest training time for rotor bar fault classification, and the fault classification is fast. In addition, Table 1.5 shows the performance parameters of the AlexNet model and the other four models under different Batch Sizes and MaxEpochs.

[0078] Table 1.5 Model performance parameters

[0079]

[0080]

[0081]

[0082] Table 1.6 Model multi-classification performance parameters

[0083]

[0084]

[0085] From the classification and multi-classification performance of each model in Table 1.5 and Table 1.6, it can be seen that the fault classification effect of the AlexNet model under different Batch Sizes and Max Epochs is better than that of the other four models, and the classification accuracy is high. Combining Table 1.4 and Table 1.5, it can be obtained that the AlexNet model not only has a high classification accuracy, but also has the characteristics of rapid classification, which further illustrates the effectiveness of the proposed rotor broken bar fault diagnosis and classification method based on sound signals.

[0086] The present invention proposes a signal extraction and classification method based on sound for diagnosing and classifying motor rotor faults. First, the collected sound signal in .mp4 format is converted into .wav format and pre-processed by resampling at equal time intervals. Then, the time domain image and frequency domain image of the sound signal are made, and the short-time Fourier transform is used to extract the fault features of the pre-processed sound signal to obtain a two-dimensional sound spectrogram and construct a data set, which is input into the AlexNet model for training, thereby completing the diagnosis and classification of the broken rotor bar fault. Finally, the classification effects of four different CNN models are compared and analyzed for performance. The experimental results show that the fault diagnosis and classification method based on sound signals and deep learning can effectively realize the diagnosis and classification of motor rotor broken bar faults, providing a new research idea for the diagnosis of motor rotor broken bar faults.

[0087] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0088] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A sound-based signal extraction and classification method, characterized in that: The following steps are involved: Step 1, collecting sound signals of the motor under different states; Step 2, preprocessing the collected sound signal; Step 3, using the sound signal preprocessed in step 2 to extract fault features, obtain the time-frequency spectrum of the sound and construct a data set; specifically, the three sound signals of healthy motor operation, one rotor broken bar fault and two rotor broken bar fault are subjected to short-time Fourier transform to extract fault features and obtain the time-frequency spectrum of the sound; Step 4: Build a CNN model and input the data set constructed in step 3 into the model for training and testing. The CNN model built is the AlexNet model, and its specific structure is as follows: the AlexNet model consists of multiple convolutional layers, pooling layers, and fully connected layers. The model starts with convolutional layer 1, using an 11x11 convolution kernel, a step size of 4x1, a total of 96 convolution kernels, and an output size of 55x55x96. This is followed by a maximum pooling layer, using a 3x3 convolution kernel, a step size of 2x1, and also 96 convolution kernels, with an output size of 27x27x96. This is followed by a convolutional layer 2, using a 5x5 convolution kernel, a step size of 1x1, and 256 convolution kernels, with an output size of 27x27x 256, and again through the maximum pooling layer, using a 3x3 convolution kernel, a stride of 2x1, 256 convolution kernels, and an output size of 13x13x256; the model continues to include convolution layers 3 and 4, both using 3x3 convolution kernels, a stride of 1x1, 384 convolution kernels each, and an output size of 13x13x384; convolution layer 5 uses a 3x3 convolution kernel, a stride of 1x1, 256 convolution kernels, and an output size of 13x13x256, and again through the maximum pooling layer, using a 3x3 convolution kernel, a stride of 2x1, 256 convolution kernels, and an output size of 6x6x256; finally, the model contains three fully connected layers, each with an output size of 6x6x256; Step 5: Use the trained model to diagnose and classify the motor rotor broken bar fault and output the results.

2. A sound-based signal extraction and classification method as claimed in claim 1, characterized in that: In the step 1, the sound signals collected when the motor is in different states are divided into two rounds. The first round is to collect the sound signals of the healthy motor when it is running smoothly for 20 minutes, and the second round is to collect the sound signals of the motor when one rotor bar is broken and the sound signals of the motor when two rotor bars are broken for 20 minutes each.

3. A sound-based signal extraction and classification method as claimed in claim 2, characterized in that: The specific process of the second round of sound signal collection is to disassemble the motor, separate the motor rotor, use an electric drill to drill a hole and cut off a rotor bar, reassemble the motor, and collect the sound signal of a broken rotor bar fault after the motor is powered on and runs smoothly, and collect sound data for 20 minutes; after the motor is powered off and cooled, disassemble the motor again, separate the motor rotor, use an electric drill to drill a hole and cut off two rotor bars, reassemble the motor, and collect the sound signals of two broken rotor bars fault after the motor is powered on and runs smoothly, and collect sound data for 20 minutes.

4. A sound-based signal extraction and classification method as claimed in claim 1, characterized in that: The preprocessing of the collected sound signal in step 2 is divided into two steps: S1: The collected sound signals of the motor in different states are converted into .wav format; S2: Resample the entire sound signal at equal time intervals, and set the resampling time interval to 60 seconds.

5. The method for extracting and classifying signals based on sound according to claim 1, characterized in that: The data set constructed in step 3 is randomly divided into a training set and a test set, and the ratio of the training set to the test set is 7:3.

Images