Motor fault diagnosis method based on voiceprint feature extraction and transfer learning

Abstract

The invention discloses a motor fault diagnosis method based on voiceprint feature extraction and transfer learning, and belongs to the technical field of equipment fault diagnosis, and the method comprises the following steps: S1, obtaining a sound signal of a working state of a motor, and extracting voiceprint features through an empirical mode decomposition and Mel-frequency cepstral coefficient method; s2, constructing a voiceprint feature data set according to the extracted voiceprint features; and S3, inputting the voiceprint feature data set into a CNN + ResNet network to obtain a motor fault diagnosis result. According to the method, the limitation of a traditional contact type method can be overcome, the maintenance cost is reduced, more accurate diagnosis information can be provided at the early stage of a motor fault, and the operation safety and reliability of motor equipment are improved. Data analysis and fault classification are carried out through a voiceprint feature extraction technology in combination with a deep learning model, a new direction is developed for motor fault diagnosis, and the method has important research value and application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault diagnosis, and specifically relates to a motor fault diagnosis method based on voiceprint feature extraction and transfer learning. Background Art

[0002] As deep learning technology becomes more mature, the use of deep learning for motor fault diagnosis has received widespread attention, especially in processing complex nonlinear and high-dimensional data. Compared with traditional methods, deep learning can automatically learn motor fault characteristics through a large amount of data, and has strong adaptability and generalization capabilities.

[0003] However, although deep learning has shown great potential and advantages in motor fault diagnosis, the existing diagnostic methods based on contact data acquisition still have some significant limitations. For example, the deep learning model training process needs to rely on a large amount of fault data, and the current motor fault diagnosis methods based on deep learning usually rely on obtaining the working status data of the motor through contact sensors (such as vibration sensors, accelerometers, current sensors, etc.). These data include vibration signals, high-frequency current signals, SPM pulse currents, etc. of the motor in a healthy or faulty state, which constitute the data set for training deep learning models. The core of these methods is to analyze the complex data collected by these sensors through deep learning models, automatically extract the characteristics of motor faults, and further perform fault diagnosis.

[0004] However, this reliance on contact sensors has certain limitations and challenges: (1) Sensors need to be attached, and non-contact detection cannot be achieved: Existing fault diagnosis methods generally rely on installing sensors directly on the surface or inside the motor, which not only requires physical contact with the motor, but also requires frequent maintenance and calibration. This method limits its application in high-risk or difficult-to-access environments, and in some cases may have a certain impact on the operation of the motor itself.

[0005] (2) High diagnostic cost: Contact sensors usually need to be installed at key locations on the motor, and different types of faults may require different sensors to collect multiple data, which greatly increases the installation and maintenance costs of the equipment. In addition, the sensor itself has a certain price and is easily affected by environmental factors (such as temperature, humidity, vibration, etc.), resulting in data instability and inaccuracy, further increasing the cost of the system.

[0006] (3) The installation location is sensitive to the fault diagnosis results: The installation location of the sensor, the selection of the sensor and its configuration may have a significant impact on the final diagnosis results. It is also sensitive to environmental interference. Under some extreme working conditions (such as high temperature, high humidity, strong vibration, etc.), the stability and reliability of the sensor may be affected, resulting in reduced accuracy of data collection, thereby affecting the credibility of the diagnosis results. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the present invention provides a motor fault diagnosis method based on voiceprint feature extraction and transfer learning, which solves the problems of the prior art method using contact sensors, such as the need to attach sensors, the inability to achieve non-contact detection, high diagnosis costs, and the installation position being sensitive to fault diagnosis results.

[0008] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a motor fault diagnosis method based on voiceprint feature extraction and transfer learning, comprising the following steps: S1. Obtain the acoustic signal of the motor working state, and extract the voiceprint features through empirical mode decomposition and Mel cepstral coefficient method; S2. Construct a voiceprint feature dataset based on the extracted voiceprint features; S3. Input the voiceprint feature dataset into the CNN+ResNet network to obtain the motor fault diagnosis result.

[0009] Further: S1 comprises the following sub-steps: S11, performing empirical mode decomposition on the acoustic signal to obtain a number of eigenmode components, and selecting three eigenmode components with the highest energy contribution as the main components of the acoustic signal; S12. Calculate the Mel-frequency cepstral coefficient of each eigenmode component to obtain the cepstral feature of the eigenmode component in the Mel frequency band, and generate the voiceprint feature by combining the Mel-frequency cepstral coefficient of the main component. The voiceprint feature includes multi-scale and multi-time-frequency feature information.

[0010] Further: S2 is specifically: A three-channel data set is constructed according to the voiceprint features as the voiceprint feature data set.

[0011] Further: S3 includes the following sub-steps: S31, input the voiceprint feature data set into the CNN model, perform a first-stage feature extraction, and obtain the time-frequency feature information of the voiceprint features of each channel; S32, inputting the time-frequency feature information and the voiceprint feature data set into the pre-trained ResNet network, performing two-stage feature extraction, and obtaining the motor fault feature; S33, performing feature mapping according to the motor fault features to obtain a motor fault diagnosis result.

[0012] Further: the S32 is specifically: The time-frequency feature information is input into the pre-trained ResNet network for training. Based on the training results and the task similarity between the voiceprint feature dataset and the image field, the front-end parameters of ResNet are frozen, and the terminal parameters are fine-tuned using the voiceprint feature dataset to adapt to the characteristic scenarios of motor fault diagnosis and output the motor fault features.

[0013] The beneficial effects of the present invention are as follows: the present invention provides a motor fault diagnosis method based on voiceprint feature extraction and transfer learning, proposes a voiceprint feature extraction method integrating empirical mode decomposition and Mel cepstral coefficients to construct a voiceprint feature data set, and then makes full use of transfer learning and ResNet pre-training features to achieve motor fault prediction and diagnosis in a small amount of voiceprint feature data scenario. Compared with the prior art, the present invention has the following advantages: (1) Non-contact, remote monitoring: Acoustic signals can be collected remotely through a microphone array without direct contact with the motor, which can effectively avoid the complexity of installation and maintenance of contact sensors. In particular, effective fault monitoring and diagnosis of the motor can still be performed in high temperature, high pressure or other dangerous environments.

[0014] (2) Low diagnostic cost: The acoustic signal of the motor in working state can be collected by microphone array. Its hardware cost is lower than that of vibration sensor, high-frequency current transformer, etc. Even the microphone of handheld terminals such as mobile phones can be used to collect acoustic signals for fault diagnosis, which has low diagnostic cost.

[0015] (3) Strong environmental adaptability: In some special working environments (such as environments with strong electromagnetic interference, extremely high or extremely low temperatures), the accuracy and stability of vibration sensors may be affected, while the acquisition of acoustic signals is not easily affected by these external factors. Therefore, in complex environments, acoustic signals can still provide reliable information and avoid the influence of the installation position of the attached sensor on fault diagnosis.

[0016] (4) Real-time and high efficiency: The acquisition and processing of acoustic signals are relatively simple, and real-time monitoring and analysis can be performed during operation to promptly detect potential motor faults and ensure the safety and efficiency of the equipment. In particular, the combination of the CNN+ResNet two-stage transfer learning fault diagnosis strategy can greatly reduce the dependence on fault data based on the pre-training results of ResNet, realize a lightweight motor fault diagnosis solution, obtain motor fault diagnosis results more quickly, and achieve more effective fault handling and decision-making basis.

[0017] In summary, the motor fault diagnosis scheme based on voiceprint feature extraction and transfer learning proposed in the present invention can overcome the limitations of traditional contact methods, reduce maintenance costs, and provide more accurate diagnostic information in the early stages of motor failures, thereby improving the operational safety and reliability of motor equipment. The voiceprint feature extraction technology combined with deep learning models for data analysis and fault classification will open up new directions for motor fault diagnosis, which has important research value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a motor fault diagnosis method based on voiceprint feature extraction and transfer learning of the present invention.

[0019] Figure 2 This is a schematic diagram of the overall technical route.

[0020] Figure 3 This is the workflow for voiceprint feature extraction based on EMD and MFCC.

[0021] Figure 4 Construct a pipeline for the voiceprint dataset.

[0022] Figure 5 This is the two-stage transfer learning fault diagnosis strategy process based on CNN+ResNet. DETAILED DESCRIPTION

[0023] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0024] like Figure 1 As shown, in one embodiment of the present invention, a motor fault diagnosis method based on voiceprint feature extraction and transfer learning includes the following steps: S1. Obtain the acoustic signal of the motor working state, and extract the voiceprint features through empirical mode decomposition and Mel cepstral coefficient method; S2. Construct a voiceprint feature dataset based on the extracted voiceprint features; S3. Input the voiceprint feature dataset into the CNN+ResNet network to obtain the motor fault diagnosis result.

[0025] In this embodiment, the present invention proposes a motor fault diagnosis method based on voiceprint feature extraction and transfer learning, and its overall solution architecture is as follows: Figure 2As shown in the figure, the proposed method mainly includes three execution steps: (1) voiceprint feature extraction based on EMD and MFCC; (2) construction of voiceprint feature dataset based on MFCC of each intrinsic mode component (IMF); (3) construction of two-stage transfer learning fault diagnosis strategy based on CNN+ResNet.

[0026] The S1 comprises the following sub-steps: S11, performing empirical mode decomposition on the acoustic signal to obtain a number of eigenmode components, and selecting the three eigenmode components with the highest energy contribution (Ecd) as the main components of the acoustic signal; S12. Calculate the Mel-frequency cepstral coefficient of each eigenmode component to obtain the cepstral feature of the eigenmode component in the Mel frequency band, and generate the voiceprint feature by combining the Mel-frequency cepstral coefficient of the main component. The voiceprint feature includes multi-scale and multi-time-frequency feature information.

[0027] like Figure 3 As shown, in this embodiment, the present invention proposes a voiceprint feature extraction method based on empirical mode decomposition (EMD) and Mel-frequency cepstral coefficients (MFCC), performs empirical mode decomposition on the acoustic signal to obtain each eigenmode component, calculates the energy contribution of all eigenmode components according to the formula, arranges the eigenmode components in descending order according to the energy contribution, and selects the first three eigenmode components in descending order as the main components of the acoustic signal. The energy contribution of the eigenmode components The specific expression is:

[0028] In the formula, For the The energy of the eigenmode components, is the sum of the energies of all eigenmode components; The S2 is specifically: A three-channel data set is constructed according to the voiceprint features as the voiceprint feature data set.

[0029] like Figure 4 As shown, in this embodiment, a large number of data sets need to be constructed in the process of motor fault diagnosis using deep learning. Based on the extracted voiceprint feature information, a three-channel data set similar to a color image is constructed for deep learning model training and motor fault diagnosis.

[0030] The S3 comprises the following sub-steps: S31, input the voiceprint feature data set into the CNN model, perform a first-stage feature extraction, and obtain the time-frequency feature information of the voiceprint features of each channel; S32, inputting the time-frequency feature information and the voiceprint feature data set into the pre-trained ResNet network, performing two-stage feature extraction, and obtaining the motor fault feature; S33, performing feature mapping according to the motor fault features to obtain a motor fault diagnosis result.

[0031] The S32 is specifically: The time-frequency feature information is input into the pre-trained ResNet network for training. Based on the training results and the task similarity between the voiceprint feature dataset and the image domain, the front-end parameters of the ResNet are frozen, and the front-end convolutional layer parameters of the ResNet pre-trained model are retained to avoid updating these parameters during the training process. The terminal parameters are trained and fine-tuned using the voiceprint feature dataset to adapt them to the characteristic scenarios of motor fault diagnosis and output motor fault features.

[0032] like Figure 5 As shown, in this embodiment, the present invention fully integrates the feature extraction advantages of CNN and ResNet, and proposes a two-stage transfer learning fault diagnosis strategy based on CNN+ResNet. The CNN model is used to perform one-stage feature extraction on the three-channel voiceprint feature data to extract the time-frequency feature information of each channel voiceprint feature; relying on the large-scale parameter training results of ResNet and the task similarity between the three-channel voiceprint feature data and the image field, the front-end parameters of ResNet are frozen, and the voiceprint feature data set is used to fine-tune the terminal parameters to adapt them to the characteristic scenarios of motor fault diagnosis, so as to realize the task migration of motor fault diagnosis.

[0033] In the description of the present invention, it is necessary to understand that the orientation or positional relationship indicated by the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used only for descriptive purposes, and cannot be understood as indicating or implying the relative importance or the number of implicitly specified technical features. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of the features.

Claims

1. A motor fault diagnosis method based on voiceprint feature extraction and transfer learning, characterized in that: The following steps are involved: S1. Obtain the acoustic signal of the motor working state, and extract the voiceprint features through empirical mode decomposition and Mel cepstral coefficient method; S2. Construct a voiceprint feature dataset based on the extracted voiceprint features; S3. Input the voiceprint feature dataset into the CNN+ResNet network to obtain the motor fault diagnosis result.

2. The motor fault diagnosis method based on voiceprint feature extraction and transfer learning according to claim 1 is characterized in that: The S1 comprises the following sub-steps: S11, performing empirical mode decomposition on the acoustic signal to obtain a number of eigenmode components, and selecting three eigenmode components with the highest energy contribution as the main components of the acoustic signal; S12, calculate the Mel cepstral coefficient of each eigenmode component to obtain the cepstral feature of the eigenmode component in the Mel frequency band, The voiceprint features are generated by combining the Mel-frequency cepstral coefficients of the main components. The voiceprint features include multi-scale and multi-time-frequency feature information.

3. The motor fault diagnosis method based on voiceprint feature extraction and transfer learning according to claim 1 is characterized in that: The S2 is specifically: A three-channel data set is constructed according to the voiceprint features as the voiceprint feature data set.

4. The motor fault diagnosis method based on voiceprint feature extraction and transfer learning according to claim 3 is characterized in that: The S3 comprises the following sub-steps: S31, input the voiceprint feature data set into the CNN model, perform a first-stage feature extraction, and obtain the time-frequency feature information of the voiceprint features of each channel; S32, inputting the time-frequency feature information and the voiceprint feature data set into the pre-trained ResNet network, performing two-stage feature extraction, and obtaining the motor fault feature; S33, performing feature mapping according to the motor fault features to obtain a motor fault diagnosis result.

5. The motor fault diagnosis method based on voiceprint feature extraction and transfer learning according to claim 4 is characterized in that: The S32 is specifically: The time-frequency feature information is input into the pre-trained ResNet network for training. Based on the training results and the task similarity between the voiceprint feature dataset and the image field, the front-end parameters of ResNet are frozen, and the terminal parameters are fine-tuned using the voiceprint feature dataset to adapt to the characteristic scenarios of motor fault diagnosis and output the motor fault features.

Images