Motor fault detection method based on audio signal multi-dimensional sensitive features and S-ReXNet

Abstract

The invention discloses a motor fault detection method based on audio signal multi-dimensional sensitive features and S-ReXNet, and relates to a motor fault detection method. The invention aims to solve the problem that the existing motor fault detection method is inconvenient for real-time monitoring and early fault discovery. According to the method, the characteristics of the internal magnetic field distortion of the motor rotor fault and the mechanical noise of the bearing are extracted by collecting the audio signals, the real-time fault diagnosis is realized through the S-ReXNet model, the fault diagnosis of the rotor and the bearing of the permanent magnet synchronous motor is finally realized, and the method is of great significance to the reliability of the motor and the control of the high-precision motor. The invention belongs to the technical field of motors.

Description

Technical Field

[0001] The invention relates to a motor fault detection method and belongs to the technical field of motors. Background Art

[0002] Permanent magnet synchronous motors have become the power core of many industrial fields due to their ultra-high efficiency, power density and torque density. However, due to the high power density, complex operating conditions, small installation space and poor heat dissipation environment of permanent magnet synchronous motors, permanent magnet demagnetization faults and rotor dynamic eccentricity faults often occur in permanent magnet synchronous motors. At the same time, due to the wear caused by the large operating load of the motor, excessive temperature, corrosive substances or particulate matter will damage the bearings. With the widespread use of high switching frequency devices such as silicon carbide, the number of bearing electrocorrosion faults has increased, which has also increased the mass and diversity of bearing fault audio signal data. Therefore, accurate diagnosis of permanent magnet demagnetization faults, rotor dynamic eccentricity faults and bearing faults is of great significance to ensure system safety, improve system efficiency and reduce losses.

[0003] There are two main types of traditional motor rotor fault diagnosis methods. The first method is to establish mathematical equations for the motor under different faults based on the motor model, and then derive the fault characteristics hidden in signals such as voltage, current, vibration and noise. Another method is called motor current / voltage feature analysis, which identifies the motor state based on the harmonic components of the motor current or voltage signal, but is often affected by the motor pole slot matching, and it is difficult to accurately distinguish between demagnetization faults and rotor eccentricity faults. The mainstream method for bearing faults is based on artificial intelligence classification of vibration signals, that is, to extract the fault information hidden in the vibration signal through models such as machine learning. Such methods often require special instruments and equipment to collect and analyze vibration signals, which increases the cost and usage threshold of the equipment, and invasive sensors will have additional impacts on the system under test. These methods are not convenient for real-time monitoring and early detection of faults.

[0004] In the field of audio feature extraction, time domain features generally include characteristic parameters such as short-time energy and short-time zero-crossing rate. Frequency domain features include parameters such as spectrum, frequency domain energy, and spectrum centroid. Cepstral domain features include Mel-frequency cepstral coefficients (MFCC) and linear prediction cepstral coefficients (LPCC). Using a single time-frequency domain feature cannot accurately express the fault audio features, and the accuracy is usually low. Most audio feature extraction studies choose to use MFCC for feature extraction. However, due to the unclear physical meaning of a single MFCC, it is not easy to describe the original fault characteristics, and the performance is easily affected in scenarios with a wide variety of faults and high noise.

[0005] In the field of visual neural networks, convolutional neural networks (CNNs) have been widely used due to their inherent excellent characteristics. However, CNN also has some defects: 1. The convolution operation in CNN can only capture local information, but cannot establish long-distance connections of the global image. 2. The existence of the pooling layer will lead to the loss of a lot of very valuable information, and will also ignore the relationship between the whole and the part. Although the visual Transformer model has achieved excellent results in large models, the model is large and complex, and requires a large number of parameters and pre-training. It performs generally in small data samples, and its application in embedded platforms will also be limited. In 2021, Dongyoon Han et al. studied a new method for designing lightweight models, ReXNets (Rank eXpandsion Networks), based on the lightweight network MobileNet, and further demonstrated significant tuning performance in COCO object detection, instance segmentation, and fine-grained classification.

[0006] This study usually uses a microphone to collect the audio signal emitted by the motor, and then transmits the collected audio signal to a single-chip microcomputer or laptop computer and other devices. These platforms calculate the main audio feature indicators, and then further determine whether a fault has occurred. According to investigations, the characteristic frequencies of electromagnetic noise and mechanical noise of motor and bearing failures usually do not exceed 5000Hz, and the human ear's hearing range is 20-20000Hz, so a microphone that collects voice signals can meet the conditions. The selection of a digital signal analysis system requires high integration, strong computing power, low power consumption, high real-time performance, strong anti-interference ability, and system stability. It is usually a platform such as a computer and FPGA. Summary of the invention

[0007] In order to solve the problem that existing motor fault detection methods are not convenient for real-time monitoring and early detection of faults, the present invention proposes a motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet.

[0008] The technical solution adopted by the present invention to solve the above problems is: the present invention specifically includes:

[0009] Step 1, collecting data; collecting audio signals of healthy motors, permanent magnet demagnetization motors, rotor dynamic eccentric motors and four types of bearing faults during operation under different speeds and load conditions;

[0010] Step 2: Process the signal;

[0011] Step 3: Build a convolutional autoencoder and pre-train it;

[0012] Step 4: Build the S-ReXNet network;

[0013] Step 5: Fault detection model training;

[0014] Step 6: Fault detection and performance evaluation.

[0015] Furthermore, in step 2, based on the data collected in step 1, the data processing process includes:

[0016] Step 201, pre-processing the original audio signal by frame division and windowing to construct a multi-dimensional sensitive feature matrix; the MDS matrix consists of six parts: short-time energy, short-time zero-crossing rate, vibration entropy, linear prediction cepstrum coefficient, Mel frequency cepstrum coefficient and Delta coefficient;

[0017] Step 202: reorganize the feature vectors obtained by audio processing, align the time frames, and arrange different features in sequence to obtain the RGB feature spectrum of MDS.

[0018] Furthermore, step 201 specifically includes:

[0019] Step A: Use the Hamming window w[n] as the window function. The formula is as follows:

[0020]

[0021] In formula (1), n ​​is the sampling point index in the window, and L is the length of the window;

[0022] Step B: Since the energy of the audio signal changes with time, the short-time energy is analyzed. The short-time energy expression is:

[0023]

[0024] In formula (2), x(m) represents a frame of discrete sound sampling data, w(n) represents the window function, and N represents the window length;

[0025] Step C: Let the audio signal be x[n], and the i-th frame speech signal obtained after frame processing be y i [n], the frame length is L, and the short-time zero-crossing rate calculation formula is:

[0026]

[0027] In formula (3), y n (m-1) is the previous sampling point of the n-th frame signal;

[0028] Step D: Assume the frequency domain signal of the frame signal is X k , the calculation formula of vibration entropy is:

[0029]

[0030] Step E: Use LPCC to extract features. After preprocessing, extract LPC feature parameters:

[0031]

[0032] In formula (5), LPC is used to derive the LPCC characteristic parameters, p represents the LPC order, n is the delay, m represents the LPCC order, x(n) represents the covariance function, represents the LPC coefficient;

[0033]

[0034] In formulas (6) and (7), c m represents the m-th order LPCC coefficient, c k represents the k-th order LPCC coefficient, x m represents the m-th order LPC coefficient, x m-k represents the mkth order LPC coefficient;

[0035] Step F, extracting MFCC coefficients. MFCC feature extraction includes framing, preprocessing, discrete Fourier transform, Mel bandpass filtering, discrete cosine transform and other steps, as follows:

[0036] The relationship between physical frequency and Mel frequency is:

[0037] Mel(f)=2595lg(1+f / 700)(8),

[0038] In formula (8), f represents;

[0039] Step G: The calculation formula of the differential parameter is:

[0040]

[0041] In formula (14), d(t) represents the tth first-order difference; c(t) represents the tth cepstral coefficient, K represents the time difference of the first-order derivative, which is 1 or 2, c(t-1) represents the t-1th cepstral coefficient, t represents the Mel-frequency cepstral coefficient index, and M represents the order of the cepstral coefficient.

[0042] Further, step F specifically includes:

[0043] Step (1), perform discrete Fourier transform on each frame of the preprocessed signal; let the preprocessed time domain signal be S(n), and the frequency domain signal S(k) after DFT transformation be expressed as:

[0044]

[0045] Step (2), calculate the square of S(k) to obtain the energy spectrum, and then use M Mel bandpass filters to filter. The transfer function of the mth filter is:

[0046]

[0047] In formula (10), f(m) is the center frequency of the triangular filter, and:

[0048]

[0049] Step (3), calculate the logarithmic power of each filter group. The logarithmic energy of the mth filter group is:

[0050]

[0051] Step (4), the Mel cepstrum coefficients can be obtained by discrete cosine transform:

[0052]

[0053] In formula (13), M is the number of Mel filters and also the dimension of MFCC features.

[0054] Furthermore, step 3 specifically includes:

[0055] Step 301, the encoder is composed of three convolutional layers, each of which is followed by a ReLU function to introduce nonlinear features and improve the representation ability of the convolutional autoencoder;

[0056] The three convolutional layers learn the spatial information of the input data and the association between different features, gradually compressing the features and achieving feature dimensionality reduction;

[0057] The decoder part performs layer-by-layer deconvolution operations on the encoded features, and finally reconstructs the features to obtain an output of the same size as the input image;

[0058] Step 302: The loss function uses the mean square error loss function to measure the difference between the model prediction and the actual label. The Adam optimizer is used, and the L2 regularization is set with a parameter of 0.00001, a learning rate of 0.001, a batch size of 64, and a total of 60 iterations.

[0059] Furthermore, in step 5, the audio with a sampling rate of 48000 Hz is evenly divided into audio slices of 80000 sample points, and frame division and windowing are performed according to a frame width of 25 ms and a step length of 10 ms.

[0060] Furthermore, in step 6, the data collected in step 1 is calculated in step 2 to obtain the audio features of the current state of the motor, which are compressed by the convolutional autoencoder and input into the fault diagnosis model built in step 4 to realize the fault diagnosis of the motor rotor and bearing.

[0061] The beneficial effects of the present invention are:

[0062] 1. The present invention proposes a method for constructing multi-dimensional sensitive features of permanent magnet synchronous motor audio. By introducing a method with clear physical principles, the fault features of the permanent magnet synchronous motor are extracted and reorganized, and static and dynamic information that can deeply represent the audio time-frequency domain and cepstrum domain is mined, thereby enhancing the dimension and sensitivity of feature representation;

[0063] 2. Compared with vibration signals, audio signals can be obtained in a non-invasive way without disassembling or modifying the motor, which means that real-time fault diagnosis can be performed when the motor is operating normally, reducing downtime and improving diagnostic efficiency;

[0064] 3. The present invention uses a convolutional autoencoder to achieve feature dimensionality reduction and feature compression, performs feature dimensionality reduction while retaining sensitive feature information, and reduces the original training time by 69.88%;

[0065] 4. The present invention proposes a permanent magnet synchronous motor fault diagnosis model based on the S-ReXNet network, and its performance is very stable in a high noise environment, leaving margin for feature separation in a high noise environment and fault diagnosis when the characteristics of minor faults are not obvious;

[0066] 5. The S-ReXNet network is a lightweight model with fewer parameters and computational complexity. It can still guarantee good performance without pre-training and a large number of data sets. It has broad application potential in resource-constrained application scenarios, such as mobile devices and embedded systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a hardware system block diagram;

[0068] Figure 2 It is the MDS feature extraction flow chart;

[0069] Figure 3 It is the convolutional autoencoder network structure;

[0070] Figure 4 is the convolutional autoencoder pre-training loss function value;

[0071] Figure 5 It is the seven-class confusion matrix and t-SNE feature map of motor fault diagnosis;

[0072] Figure 6 It is the accuracy and loss function curve of seven-classification of motor fault diagnosis;

[0073] Figure 7 It is the overall technology roadmap. DETAILED DESCRIPTION

[0074] Specific implementation method 1: Figures 1 to 7 As shown in the figure, the motor fault detection method based on the multi-dimensional sensitive features of audio signals and S-ReXNet specifically includes:

[0075] Step 1, collecting data; collecting audio signals of healthy motors, permanent magnet demagnetization motors, rotor dynamic eccentric motors and four types of bearing faults during operation under different speeds and load conditions;

[0076] Step 2: Process the signal;

[0077] Step 3: Build a convolutional autoencoder and pre-train it;

[0078] Step 4: Build the S-ReXNet network;

[0079] Step 5: Fault detection model training;

[0080] Step 6: Fault detection and performance evaluation.

[0081] The present invention uses an INMP441 microphone, which has a high-precision 24-bit digital I2S interface, a high signal-to-noise ratio of 61dBA, a high sensitivity of -26dBFS, and a flat frequency response of 60Hz to 15kHz. After dual-channel acquisition of the audio signal, it is converted into analog and filtered and input into the FPGA for further processing; Figure 1 As shown, the data collected in the present invention are all collected in a scene where the ambient noise sound pressure level is above 64dB.

[0082] In step 3, in order to speed up the training of the diagnostic model and save computing resources, this paper builds a convolutional autoencoder to reduce the dimension and compress the MDS feature spectrum, and finds the best model and optimization solution by balancing the accuracy and training speed.

[0083] In step 4, the ReXNet network is improved on the basis of the lightweight network MobileNetV2, and the feature performance bottleneck problem of the existing network is effectively alleviated by appropriate adjustments; when the input dimension of the traditional CNN is low, a lot of information will be lost after passing through a nonlinear function such as ReLU. When the input dimension is high enough, less information will be lost after passing through the nonlinear function. To address this problem, the inverted residual structure and linear bottleneck used in MobileNetV2, and the SE module proposed in SENet are the important basis for the lightweight neural network; in the context of the present invention, the MDS feature map can accurately represent the audio characteristics of the motor state, too many bottleneck structures will cause the model to overfit on the training data, and some bottleneck structures will increase redundant information, reduce the model accuracy and generalization ability, and the complex model structure also means higher computing resource loss, which will be limited in practical applications; the present invention simplifies and improves the existing modules of the ReXNet network to obtain the simplified ReXNet (Simplified ReXNet, S-ReXNet) network of this article, and the network structure is shown in Table 2,

[0084] Table 2 S-ReXNet model structure

[0085]

[0086] In step 5, the generated MDS spectrogram has different resolutions, so the original spectrogram size is first adjusted to (3, 224, 224) and normalized, and then compressed to a feature spectrogram of (3, 64, 64) by the convolutional autoencoder. The compressed MDS feature map is input into the S-ReXNet model, and the network parameters are updated using the back propagation algorithm. The hyperparameter values ​​of the present invention are shown in Table 3.

[0087] Table 3 ReXNet model parameters

[0088]

[0089] Specific implementation method 2: Figures 1 to 7 As shown, in step 2, based on the data collected in step 1, the data processing process includes:

[0090] Step 201, pre-processing the original audio signal by frame division and windowing to construct a multi-dimensional sensitive feature matrix; the MDS matrix consists of six parts: short-time energy, short-time zero-crossing rate, vibration entropy, linear prediction cepstrum coefficient, Mel frequency cepstrum coefficient and Delta coefficient;

[0091] Step 202: reorganize the feature vectors obtained by audio processing, align the time frames, and arrange different features in sequence to obtain the RGB feature spectrum of MDS.

[0092] Specific implementation method three: Figures 1 to 7 As shown, step 201 specifically includes:

[0093] Step A: Use the Hamming window w[n] as the window function. The formula is as follows:

[0094]

[0095] In formula (1), n ​​is the sampling point index in the window, and L is the length of the window;

[0096] Step B: Since the energy of the audio signal changes with time, the short-time energy is analyzed. The short-time energy expression is:

[0097]

[0098] In formula (2), x(m) represents a frame of discrete sound sampling data, w(n) represents the window function, and N represents the window length;

[0099] Step C: Let the audio signal be x[n], and the i-th frame speech signal obtained after frame processing be y i [n], the frame length is L, and the short-time zero-crossing rate calculation formula is:

[0100]

[0101] In formula (3), y n (m-1) is the previous sampling point of the n-th frame signal;

[0102] Step D: Assume the frequency domain signal of the frame signal is X k , the calculation formula of vibration entropy is:

[0103]

[0104] Step E: Use LPCC to extract features. After preprocessing, extract LPC feature parameters:

[0105]

[0106] In formula (5), LPC is used to derive the LPCC characteristic parameters, p represents the LPC order, n is the delay, m represents the LPCC order, x(n) represents the covariance function, represents the LPC coefficient;

[0107]

[0108] In formulas (6) and (7), c m represents the m-th order LPCC coefficient, c k represents the k-th order LPCC coefficient, xm represents the mth order LPC coefficient, x m-k represents the mkth order LPC coefficient;

[0109] Step F, extracting MFCC coefficients. MFCC feature extraction includes framing, preprocessing, discrete Fourier transform, Mel bandpass filtering, discrete cosine transform and other steps, as follows:

[0110] The relationship between physical frequency and Mel frequency is:

[0111] Mel(f)=2595lg(1+f / 700)(8),

[0112] In formula (8), f represents;

[0113] Step G: The calculation formula of the differential parameter is:

[0114]

[0115] In formula (14), d(t) represents the tth first-order difference; c(t) represents the tth cepstral coefficient, K represents the time difference of the first-order derivative, which is 1 or 2, c(t-1) represents the t-1th cepstral coefficient, t represents the Mel-frequency cepstral coefficient index, and M represents the order of the cepstral coefficient.

[0116] Wherein, in step C, in audio processing, the short-time zero-crossing rate is used as one of the reference indicators for sound feature processing, and the boundary or turning point in the audio signal is determined by analyzing the change of the short-time zero-crossing rate;

[0117] In step G, the obtained MFCC features only reflect the static characteristics of the signal. By making a first-order difference on this static feature, the corresponding MFCC Delta spectrum can be obtained, which helps to fully describe the dynamic characteristics of the audio in different motor states.

[0118] Specific implementation method four: Figures 1 to 7 As shown, step F specifically includes:

[0119] Step (1), perform discrete Fourier transform on each frame of the preprocessed signal; let the preprocessed time domain signal be S(n), and the frequency domain signal S(k) after DFT transformation be expressed as:

[0120]

[0121] Step (2), calculate the square of S(k) to obtain the energy spectrum, and then use M Mel bandpass filters to filter. The transfer function of the mth filter is:

[0122]

[0123] In formula (10), f(m) is the center frequency of the triangular filter, and:

[0124]

[0125] Step (3), calculate the logarithmic power of each filter group. The logarithmic energy of the mth filter group is:

[0126]

[0127] Step (4), the Mel cepstrum coefficients can be obtained by discrete cosine transform:

[0128]

[0129] In formula (13), M is the number of Mel filters and also the dimension of MFCC features.

[0130] Specific implementation method five: Figures 1 to 7 As shown, step 3 specifically includes:

[0131] Step 301, the encoder is composed of three convolutional layers, each of which is followed by a ReLU function to introduce nonlinear features and improve the representation ability of the convolutional autoencoder;

[0132] The three convolutional layers learn the spatial information of the input data and the association between different features, gradually compressing the features and achieving feature dimensionality reduction;

[0133] The decoder part performs layer-by-layer deconvolution operations on the encoded features, and finally reconstructs the features to obtain an output of the same size as the input image;

[0134] Step 302: The loss function uses the mean square error loss function to measure the difference between the model prediction and the actual label. The Adam optimizer is used, and the L2 regularization is set with a parameter of 0.00001, a learning rate of 0.001, a batch size of 64, and a total of 60 iterations.

[0135] Wherein, in step 301, if Figure 3 As shown in the figure, during the training process, the L2 norm distance between the decoded output and the input image is used as the loss function, and the back propagation algorithm is used to update the weight parameters in the convolutional autoencoder, thereby achieving lossy compression and reconstruction of features. The convolutional autoencoder parameters are shown in Table 1;

[0136] Table 1 Convolutional autoencoder structure parameters

[0137]

[0138] In step 302, the loss value changes with the number of iterations as follows: Figure 4As shown in the figure, after 60 rounds of iterative training, the mean square error between the decoded features output by the decoder and the input features is only 0.0009, which perfectly reproduces the original information of the input features. The present invention finally obtains a compressed feature map size of (3, 64, 64) by adjusting parameters and weighing performance.

[0139] Specific implementation method six: Figures 1 to 7 As shown, the data collected in step 1 in step 6 is calculated in step 2 to obtain the audio features of the current state of the motor, which are compressed by the convolutional autoencoder and input into the fault diagnosis model built in step 4 to realize the fault diagnosis of the motor rotor and bearing.

[0140] Among them, the changes in accuracy and loss function during the training process are as follows Figure 5 As shown, the diagnosis result confusion matrix and t-SNE (t-Distributed Stochastic Neighbor Embedding) feature map are as follows Figure 6 , the test set accuracy can reach 99.76%. At the same time, the introduction of the convolutional autoencoder greatly reduces the training time. The average training time per training cycle is reduced by 69.88% of the original, and the accuracy rate is also increased by 0.05%. The overall route of the present invention is as follows Figure 7 As shown;

[0141] Table 4 Convolutional autoencoder ablation experiment

[0142]

[0143] How it works

[0144] When a permanent magnet synchronous motor has a permanent magnet demagnetization fault or an eccentricity fault, the symmetry of the rotor magnetic field is destroyed, which will introduce fractional harmonics related to the pole pair number p, affecting the air gap flux density, and then affecting the distribution of radial electromagnetic force, causing the motor to generate vibration noise with fault characteristics. Based on this feature, the abnormality of the motor rotor can be identified. Bearing faults will produce different vibration characteristic frequencies depending on the location of occurrence, and then generate fault characteristic noise.

[0145] The fault feature audio is extracted and reorganized in multiple dimensions to achieve visualization. The features are learned through S-ReXNet's inverse residual, linear bottleneck, SE structure, etc., and the features and parameters are learned by training on a large amount of labeled image data. During the training process, the model parameters are updated through the back propagation algorithm, so that the model can gradually improve the accuracy of image feature extraction and classification. Finally, intelligent fault diagnosis of permanent magnet synchronous motor rotors and bearings is achieved.

[0146] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet, characterized in that: Specifically include: Step 1, collecting data; collecting audio signals of healthy motors, permanent magnet demagnetization motors, rotor dynamic eccentric motors and four types of bearing faults during operation under different speeds and load conditions; Step 2: Process the signal; Step 3: Build a convolutional autoencoder and pre-train it; Step 4: Build the S-ReXNet network; Step 5: Fault detection model training; Step 6: Fault detection and performance evaluation.

2. The motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet according to claim 1 is characterized in that: In step 2, based on the data collected in step 1, the data processing process includes: Step 201, pre-processing the original audio signal by frame division and windowing to construct a multi-dimensional sensitive feature matrix; the MDS matrix consists of six parts: short-time energy, short-time zero-crossing rate, vibration entropy, linear prediction cepstrum coefficient, Mel frequency cepstrum coefficient and Delta coefficient; Step 202: reorganize the feature vectors obtained by audio processing, align the time frames, and arrange different features in sequence to obtain the RGB feature spectrum of MDS.

3. The motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet according to claim 2 is characterized in that: Step 201 specifically includes: Step A: Use the Hamming window w[n] as the window function. The formula is as follows: In formula (1), n ​​is the sampling point index in the window, and L is the length of the window; Step B: Since the energy of the audio signal changes with time, the short-time energy is analyzed. The short-time energy expression is: In formula (2), x(m) represents a frame of discrete sound sampling data, w(n) represents the window function, and N represents the window length; Step C: Let the audio signal be x[n], and the nth frame of speech signal obtained after frame processing be y n [m], the frame length is L, and the short-time zero-crossing rate calculation formula is: In formula (3), y n (m-1) is the previous sampling point of the n-th frame signal; Step D: Assume the frequency domain signal of the frame signal is X n , the calculation formula of vibration entropy is: Step E: Use LPCC to extract features. After preprocessing, extract LPC feature parameters: In formula (5), LPC is used to derive the LPCC characteristic parameters, p represents the LPC order, n is the delay, m represents the LPCC order, x(n) represents the covariance function, represents the LPC coefficient; In formulas (6) and (7), c m represents the m-th order LPCC coefficient, c k represents the k-th order LPCC coefficient, x m represents the m-th order LPC coefficient, x m-k represents the mkth order LPC coefficient; Step F, extracting MFCC coefficients. MFCC feature extraction includes framing, preprocessing, discrete Fourier transform, Mel bandpass filtering, discrete cosine transform and other steps, as follows: The relationship between physical frequency and Mel frequency is: Mel(f)=2595lg(1+f / 700)(8), In formula (8), f represents; Step G: The calculation formula of the differential parameter is: In formula (14), d(t) represents the tth first-order difference; c(t) represents the tth cepstral coefficient, K represents the time difference of the first-order derivative, which is 1 or 2, c(t-1) represents the t-1th cepstral coefficient, t represents the Mel-frequency cepstral coefficient index, and M represents the order of the cepstral coefficient.

4. The motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet according to claim 3 is characterized in that: Step F specifically includes: Step (1), perform discrete Fourier transform on each frame of the preprocessed signal; let the preprocessed time domain signal be S(n), and the frequency domain signal S(k) after DFT transformation be expressed as: Step (2), calculate the square of S(k) to obtain the energy spectrum, and then use M Mel bandpass filters to filter. The transfer function of the mth filter is: In formula (10), f(m) is the center frequency of the triangular filter, and: Step (3), calculate the logarithmic power of each filter group. The logarithmic energy of the mth filter group is: Step (4), the Mel cepstrum coefficients can be obtained by discrete cosine transform: In formula (13), M is the number of Mel filters and also the dimension of MFCC features.

5. The motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet according to claim 1 is characterized in that: Step 3 specifically includes: Step 301, the encoder is composed of three convolutional layers, each of which is followed by a ReLU function to introduce nonlinear features and improve the representation ability of the convolutional autoencoder; The three convolutional layers learn the spatial information of the input data and the association between different features, gradually compressing the features and achieving feature dimensionality reduction; The decoder part performs layer-by-layer deconvolution operations on the encoded features, and finally reconstructs the features to obtain an output of the same size as the input image; Step 302: The loss function uses the mean square error loss function to measure the difference between the model prediction and the actual label. The Adam optimizer is used, and the L2 regularization is set with a parameter of 0.00001, a learning rate of 0.001, a batch size of 64, and a total of 60 iterations.

6. The motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet according to claim 1 is characterized in that: In step 5, the audio with a sampling rate of 48000 Hz is evenly divided into audio slices of 80000 sample points, and frame division and windowing are performed according to a frame width of 25 ms and a step length of 10 ms.

7. The motor fault detection method based on multi-dimensional sensitive features of audio signals and S-ReXNet according to claim 1 is characterized in that: In step 6, the data collected in step 1 is calculated in step 2 to obtain the audio features of the current state of the motor, which are compressed by the convolutional autoencoder and input into the fault diagnosis model built in step 4 to realize the fault diagnosis of the motor rotor and bearing.

Images