Deep learning fan fault diagnosis method based on double-wavelet fusion and CEEMDAN decomposition

Abstract

A deep learning fan fault diagnosis method based on double wavelet fusion and CEEMDAN decomposition comprises the following steps: S1, collecting vibration signals of a fan motor driving end, a fan driving end and a non-driving end, and constructing a data set; s2, performing improvement on the basis of wavelet packet noise reduction to form a double-wavelet fusion noise reduction algorithm; s3, carrying out noise reduction processing on the acquired signals; s4, carrying out sample division on the noise reduction signals according to a fixed time window, and carrying out classification according to measurement points; s5, randomly dividing the samples into a training set, a verification set and a test set; s6, decomposing all the samples by using CEEMDAN, and extracting IMFs; s7, sending the sample into the time-frequency domain joint feature extraction model of the corresponding measuring point for modeling; s8, training to obtain a diagnosis model and exporting a weight file; and S9, loading the model detection test set, and comprehensively judging the operation state of the fan. According to the invention, the signal processing quality and the fault identification accuracy are improved.

Description

Technical Field

[0001] The invention belongs to the field of mechanical equipment fault diagnosis and relates to a fan fault diagnosis method. Background Art

[0002] Fan units operate around the clock in subway stations, but operating in confined underground spaces poses numerous safety risks. Therefore, real-time monitoring of the fan's health and fault diagnosis are crucial.

[0003] Faults in wind turbines typically include belt faults and bearing faults. These two types of fault signals are typically characterized by non-stationarity and nonlinearity, which restricts the effective extraction of fault features. Traditional noise reduction and feature extraction methods suffer from modal aliasing and over-decomposition. Bearing fault diagnosis has received widespread attention in the field of neural networks. This method first performs noise reduction and feature extraction on the original vibration signal, and then uses the classification capabilities of convolutional neural networks to perform fault diagnosis. However, the diagnostic results still result in one or two segments of fault signals that cannot be correctly identified. Although many deep learning-based fault diagnosis methods have been studied and have achieved good diagnostic results, the diagnostic performance of these methods usually depends on the quality of the signals collected under different operating conditions.

[0004] When a fan is running, changes in air pressure and overall machine vibration inevitably cause significant noise in the collected signals. Convolutional and pooling kernels in neural networks typically act as local feature extractors. In the presence of signal noise, local feature extractors struggle to detect relevant fault signatures, significantly impacting the neural network's feature learning capabilities.

[0005] In recent years, CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) has become an effective tool for processing nonlinear and nonstationary signals. It decomposes the original signal into a series of intrinsic mode functions (IMFs), each representing the local characteristics of different frequency components in the signal, and exhibits excellent time-frequency localization. However, CEEMDAN is sensitive to noise and can easily generate false modes due to high-frequency interference in the original signal, affecting subsequent feature extraction and classification accuracy.

[0006] Wavelet packet decomposition (WPD), as a noise reduction preprocessing tool, has multi-resolution properties and can effectively retain useful components and filter out noise. However, when a single wavelet function is used to process complex vibration signals, problems often arise, such as insufficient matching and loss of local details. Summary of the Invention

[0007] In order to overcome the shortcomings of the existing technology, the present invention provides a deep learning fan fault diagnosis method based on dual wavelet fusion and CEEMDAN decomposition, which improves the quality of the collected signal and enhances the accuracy of fault identification.

[0008] The technical solution adopted by the present invention to solve its technical problem is:

[0009] A deep learning method for fan fault diagnosis based on dual wavelet fusion and CEEMDAN decomposition includes the following steps:

[0010] S1. Using an acceleration vibration sensor and an acquisition card to collect acceleration vibration signals of the fan group under different operating conditions and create a data set, the acceleration vibration signal includes vibration signals of the motor drive end, the fan drive end, and the fan non-drive end of the fan;

[0011] S2, improving the wavelet packet denoising algorithm as the basic denoising algorithm to obtain an improved denoising algorithm;

[0012] S3, using the improved dual wavelet denoising algorithm to denoise the data set in S1;

[0013] S4. Divide the signal after the noise reduction processing in step S3 into a fixed-length time window; set the length of each sample and the overlap between adjacent samples; and divide the generated samples into three categories based on the positions of the three measurement points of the fan: motor drive end samples, fan drive end samples, and fan non-drive end samples;

[0014] S5. After the sample division is completed, the three categories of samples are randomly divided independently to obtain their corresponding training sets, validation sets, and test sets to ensure the fairness and accuracy of model training;

[0015] S6. Use CEEMDAN to decompose all samples of the training set, validation set, and test set in S5 to decompose IMFs to prepare for subsequent model input;

[0016] S7, input the training set and validation set samples after CEEMDAN decomposition processing into the corresponding three time-frequency domain joint feature extraction models according to their measurement point categories for feature extraction and subsequent modeling;

[0017] S8. Use the time-frequency domain synchronous feature extraction model to train the training set of the three measurement points in S6 and verify the validation set to obtain the fan fault diagnosis model and output a weight file in .pt format;

[0018] S9. Load the three weight files generated in S8 and use the test sets of the three measurement points in S6 as input samples to detect the time-frequency domain synchronous extraction feature model. The model outputs of the three measurement points together obtain the detection results of the wind turbine's operating status, and the results include various operating states of the wind turbine.

[0019] Furthermore, in S1, the acceleration vibration sensor parameters used include a sensitivity of 100mv / g and a range of 50g. The data acquisition card used is Advantech USB-4711A, which is connected to the acceleration sensor in a single-ended manner. The sampling frequency is 12000Hz, and the measurement range of the acquisition card is set to ±5V. The formula for converting the voltage signal to the acceleration signal is shown below:

[0020]

[0021] Where: a(t): acceleration signal, unit: g; V(t): collected voltage signal, unit: mV; S: sensitivity of the sensor, unit: mV / g.

[0022] Furthermore, the process of S2 is as follows:

[0023] S21, using sym8 wavelet packet function and bior3.3 wavelet packet function to decompose a section of signal respectively;

[0024] S22. By performing weighted calculation on the normalized combination of energy, variance and Shannon entropy, the fractions of wavelet coefficients decomposed from the two wavelet packets are obtained;

[0025] S23. Set a threshold. Wavelet coefficients will only be retained when their fraction is greater than the threshold.

[0026] Preferably, in S2, energy reflects the energy distribution of the signal in different frequency bands; variance describes the fluctuation amplitude or discreteness of the signal; and Shannon entropy reflects the complexity or disorder of the signal. Considering the importance of energy and variance in identifying signal features, energy and variance are given higher weights and Shannon entropy is given a relatively lower weight in the feature fusion process to improve the effectiveness of feature representation and the accuracy of fault identification.

[0027] The wavelet packet transform is expressed in the following recursive formula:

[0028] Assume the original signal is x(t), and the wavelet packet function is Then the wavelet packet coefficient c j+1,2n (k), c j+1,2n+1 (k) is expressed as:

[0029]

[0030] Where: c j,n(k): wavelet packet coefficient at the jth layer and the nth node; h(m): low-pass filter coefficient (reconstruction approximation); g(m): high-pass filter coefficient (reconstruction details); k: discrete position index; j: decomposition layer number; n: node number (each layer contains 2 j nodes).

[0031] The energy calculation formula is shown as follows:

[0032]

[0033] Where: E: signal energy; x i : the amplitude of the i-th sampling point; N: the total length of the signal;

[0034] The variance calculation formula is as follows:

[0035]

[0036] Where: V: variance of the signal; x i : the amplitude of the i-th sampling point; The mean of the signal; N: the length of the signal.

[0037] Shannon entropy formula is shown as follows:

[0038]

[0039] Where: H: Shannon entropy, used to measure the complexity or uncertainty of the signal; p i is the probability value in the i-th histogram segment (normalized to form a probability density function); N: the total number of components of the histogram;

[0040] The formula of the combined weighted scoring function is shown as follows:

[0041]

[0042] Where: E: energy ratio; E max : The maximum energy value in the current signal set; V: variance; V max : The maximum variance in the current signal set; H: Shannon entropy; H max : The maximum value w of Shannon entropy in the current signal set E , w V , w H : Their respective weights satisfy:

[0043] w E +w V +w H =1.

[0044] Furthermore, in S6, the number of IMFs decomposition is 7, and the process is as follows:

[0045] S61. If the number of IMFs decomposed from the sample is less than 7, the sample is directly deleted;

[0046] S62. If the number of IMFs decomposed from the sample is equal to 7, save the sample directly;

[0047] S63. If the number of IMFs decomposed from the sample is greater than 7, then based on the weighted average frequency of each IMF, the two IMFs with the closest weighted average frequencies are merged to reach the final number of IMFs.

[0048] Preferably, in S6, the CEEMDAN IMF extraction formula is as follows:

[0049]

[0050] Where: x(t): original time series signal; c i (t): the i-th IMF (intrinsic mode function); N: the number of IMFs; r N (t): final residual or trend term, usually a monotonically changing signal;

[0051] The IMF weighted average frequency calculation formula is as follows:

[0052] Assume that the spectrum of a certain IMF component is X(f), then the weighted average frequency f of the IMF is avg Defined as:

[0053]

[0054] Where: f i : the i-th frequency point; |X(f i )| 2 : Power spectrum density (amplitude squared) at the frequency point; N: Total number of frequency points.

[0055] Furthermore, the process of S7 is as follows:

[0056] S71, the time domain part uses 1D-CNN and LSTM in series to extract features of the time domain signal;

[0057] S72, the frequency domain part uses FFT transformation and CNN in series to convert the time domain signal into the frequency domain signal and then extract the frequency domain features;

[0058] S73. Concatenate the LSTM time domain features and the CNN frequency domain features, and finally output the classification results.

[0059] In S7 and S71, two layers of 1D-CNN use 3×3 convolution kernels to locally extract the temporal structural features of IMFs; LSTM models the implicit temporal dependencies of IMFs;

[0060] In S72, a one-dimensional fast Fourier transform is performed on the IMFs of each signal, and its complex modulus (amplitude spectrum) is taken as the input feature. The three-layer CNN uses a 3×3 convolution kernel to extract the local frequency domain features between the IMFs;

[0061] In S73, the features extracted from the time domain and the frequency domain are combined and fused, and input into a classifier to complete fault identification.

[0062] The beneficial effects of the present invention are mainly manifested in:

[0063] 1. This paper proposes a method suitable for fan fault diagnosis. It innovatively introduces two different wavelet basis functions and effectively improves signal reconstruction quality and noise suppression capabilities through an energy-, entropy-, and variance-weighted fusion strategy. Compared with traditional single wavelet packet denoising methods, this method combines the decomposition advantages of the dominant and auxiliary wavelets, balancing the preservation of high-frequency details with the suppression of low-frequency noise. It achieves superior signal denoising in both the time and frequency domains, significantly improving the accuracy and robustness of subsequent fault feature extraction.

[0064] 2. The present invention further proposes a time-frequency domain joint feature extraction and fault diagnosis method with point-by-point independent modeling. For the three key measuring points in the fan system (motor drive end, fan drive end, and fan non-drive end), independent diagnostic models are constructed respectively to fully explore the local fault feature information of each measuring point. Each model takes the vibration signal of the corresponding measuring point as input, first performs double wavelet packet weighted denoising, combines CEEMDAN decomposition to extract IMF components, and then synchronously extracts features in the time domain and frequency domain, ultimately achieving efficient and accurate point fault identification. This method avoids the interference problem caused by the mixed processing of signals from multiple measuring points, and significantly improves the model's sensitivity to local anomalies and diagnostic accuracy.

[0065] Compared with the traditional single-model multi-channel input method, the present invention improves the independent judgment ability of fault information at each measuring point through a point-by-point modeling strategy, while retaining the advantages of time-frequency feature fusion, providing an innovative solution for achieving high-reliability, high-resolution intelligent diagnosis of wind turbine systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a schematic diagram of data collection points for fault diagnosis objects of the present invention;

[0067] Figure 2 Schematic diagram of the process of the fault diagnosis method of the present invention;

[0068] Figure 3 This is a schematic diagram of the improved dual wavelet algorithm flow of the present invention;

[0069] Figure 4 Schematic diagram of the IMFs merging process decomposed by CEEMDAN in the present invention;

[0070] Figure 5 This is an overall schematic diagram of the time-frequency domain synchronous feature extraction model of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be further described below with reference to the accompanying drawings.

[0072] Reference Figures 1 to 5 , a deep learning method for fan fault diagnosis based on dual wavelet fusion and CEEMDAN decomposition, such as Figure 1 As shown, acceleration vibration sensors are magnetically installed at the motor drive end, fan drive end, and fan non-drive end, serving as motor drive end collection point 1, fan drive end collection point 2, and fan non-drive end collection point 3.

[0073] like Figure 2 As shown in FIG, the deep learning fan fault diagnosis method based on dual wavelet fusion and CEEMDAN decomposition includes the following steps:

[0074] S1. Use an accelerometer and an acquisition card to collect vibration signals at three points on the fan (motor drive end, fan drive end, and fan non-drive end) under different operating conditions. The sampling frequency of the data acquisition card is 12kHz, and the sampling range is ±5V. The parameters of the accelerometer vibration sensor are: sensitivity 100mV / g, range 50g;

[0075] The collected data is converted from voltage to acceleration and saved as a CSV file. Each column represents a channel of the data collector, that is, a data collection point of the wind turbine. The data collector model used is Advantech USB-4711A, and the connection method with the acceleration sensor is single-ended. The voltage-acceleration conversion formula is shown below:

[0076]

[0077] Where: a(t): acceleration signal, unit: g; V(t): collected voltage signal, unit: mV; S: sensor sensitivity, unit: mV / g.

[0078] S2. Improve the wavelet packet denoising algorithm as the basic denoising algorithm to obtain an improved denoising algorithm; the process is as follows:

[0079] S21, using sym8 wavelet packet function and bior3.3 wavelet packet function to decompose a section of signal respectively;

[0080] S22, by performing weighted calculation on the normalized combination of energy, variance and Shannon entropy, the fraction of wavelet coefficients decomposed from the two wavelet packets is obtained;

[0081] S23. Set a threshold. Wavelet coefficients will only be retained when their fraction is greater than the threshold.

[0082] In S2, energy reflects the energy distribution of the signal in different frequency bands; variance describes the fluctuation amplitude or discreteness of the signal; Shannon entropy reflects the complexity or disorder of the signal. Considering the importance of energy and variance in identifying signal features, energy and variance are given high weights in the feature fusion process, while Shannon entropy is given low weights.

[0083] The wavelet packet transform is expressed in the following recursive formula:

[0084] Assume the original signal is x(t), and the wavelet packet function is Then the wavelet packet coefficient c j+1,2n (k), c j+1,2n+1 (k) is expressed as:

[0085]

[0086] Where: c j,n (k): wavelet packet coefficient at the jth layer and the nth node; h(m): low-pass filter coefficient; g(m): high-pass filter coefficient; k: discrete position index; j: decomposition layer number; n: node number;

[0087] The energy calculation formula is shown as follows:

[0088]

[0089] Where: E: signal energy; x i : the amplitude of the i-th sampling point; N: the total length of the signal;

[0090] The variance calculation formula is as follows:

[0091]

[0092] Where: V: variance of the signal; x i : the amplitude of the i-th sampling point; The mean of the signal; N: signal length;

[0093] Shannon entropy formula is shown as follows:

[0094]

[0095] Where: H: Shannon entropy, used to measure the complexity or uncertainty of the signal; pi is the probability value in the i-th histogram segment; N: the total component of the histogram;

[0096] The formula of the combined weighted scoring function is shown as follows:

[0097]

[0098] Where: E: energy ratio; E max : The maximum energy value in the current signal set; V: variance; V max : The maximum variance in the current signal set; H: Shannon entropy; H max : The maximum value w of Shannon entropy in the current signal set E , w V , w H : Their respective weights satisfy:

[0099] w E +w V +w H =1.

[0100] S3. Perform dual wavelet packet denoising on the converted acceleration signal and create a data set. The dual wavelet denoising process is as follows: Figure 3 As shown, the process is as follows:

[0101] S31. Select two wavelet basis functions (sym8 and bior3.3) with complementary characteristics to perform wavelet packet decomposition on each column of signals in the CSV file. Set the decomposition level to 6. Finally, each wavelet basis function will obtain 64 wavelet packet components.

[0102] S32, performing energy calculation, variance calculation, and Shannon entropy calculation on each wavelet packet component, and calculating a combined weighted score;

[0103] S33, determining a fractional threshold of the wavelet packet components, and retaining only the wavelet packet components greater than the fractional threshold;

[0104] S34, reconstructing the retained wavelet packet components using their respective wavelet basis functions;

[0105] S35. For the signals reconstructed by the two wavelet basis functions, the two signal segments are cut to the same length, and then a point-to-point weighted average is directly performed on the two reconstructed signals;

[0106] S4. Divide the signal after the noise reduction processing in step S3 into a fixed-length time window; set the length of each sample to 1024 points, and there is a 50% overlap between adjacent samples; and divide the generated samples into three major categories based on the different positions of the three measurement points of the fan: motor drive end samples, fan drive end samples, and fan non-drive end samples;

[0107] S5. After the sample division is completed, the three types of samples are randomly divided into the ratio of 7:2:1 to obtain their corresponding training set, validation set and test set to ensure the fairness and accuracy of model training;

[0108] S6. Perform CEEMDAN decomposition on all samples. The number of IMFs decomposed is set to 7 to prepare for subsequent model input. The process of setting the number of IMFs decomposition to 7 is as follows:

[0109] S61. If the number of IMFs decomposed from the sample is less than 7, the sample is directly deleted;

[0110] S62. If the number of IMFs decomposed from the sample is equal to 7, save the sample directly;

[0111] S63. If the number of IMFs decomposed from the sample is greater than 7, then based on the weighted average frequency of each IMF, the two IMFs with the closest weighted average frequencies are merged to finally reach the number of IMFs. The formula for calculating the weighted average frequency is as follows:

[0112] Assume that the spectrum of a certain IMF component is X(f), then the weighted average frequency f of the IMF is avg Defined as:

[0113]

[0114] Where: f i : the i-th frequency point;

[0115] |X(f i )| 2 : Power spectrum density (amplitude square) at the frequency point;

[0116] N: total number of frequency points.

[0117] S7. The training set and validation set samples after CEEMDAN decomposition are input into the three corresponding time-frequency domain joint feature extraction models according to their measurement point categories for feature extraction and subsequent modeling. The construction process of the time-frequency domain simultaneous extraction feature model is as follows:

[0118] S71. Time-domain feature extraction of the IMFs of the vibration signal at each measuring point is performed using a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM) in series. First, a two-layer 1D-CNN is used to extract local temporal features. Then, a two-layer LSTM network is used to model the long-term dependencies of the time series to obtain a complete temporal dynamic feature representation.

[0119] S72. Extract frequency domain features from the IMFs of the vibration signal at each measuring point. The IMFs are converted to the frequency domain using a fast Fourier transform (FFT) and input into a frequency domain convolutional neural network branch. A three-layer convolutional neural network (CNN) is used to extract spectral structure features. Multi-layer convolution-activation-pooling modules are connected in series to model frequency distribution features and extract key frequency domain information.

[0120] S73: Concatenate and fuse the LSTM output features obtained in step S71 with the CNN frequency-domain features obtained in step S72 to construct a joint feature vector. This vector is then fed into a fully connected classifier for dimensionality reduction and discrimination, ultimately outputting a fault classification result. The classifier incorporates a ReLU activation function and a Dropout regularization module to enhance the model's nonlinear expression capabilities and generalization performance.

[0121] S8. Use the time-frequency domain synchronous extraction feature model to train the training set of the three measurement points in S6 and verify the validation set to obtain the motor drive end fault diagnosis model, the fan drive end fault diagnosis model, and the fan non-drive end fault diagnosis model and output their respective .pt format weight files;

[0122] S9. Load the three weight files generated in S8 and use the test sets of the three measurement points in S6 as input samples to detect the time-frequency domain synchronous extraction feature model. The model outputs of the three measurement points together obtain the detection results of the wind turbine's operating status, and the results include various operating states of the wind turbine.

[0123] This embodiment proposes a fan fault diagnosis method based on dual wavelet packet denoising and CEEMDAN decomposition, and establishes independent time-frequency feature extraction models for multiple key measurement point signals in the fan system to achieve high-precision and strong robust point fault identification.

[0124] This method first performs dual wavelet packet denoising on vibration signals collected from key locations in the fan system (e.g., the motor drive end, fan drive end, and fan non-drive end). During the denoising phase, the symmetric wavelet basis function sym8 and the approximately symmetric wavelet basis function bior3.3 are selected to perform wavelet packet decomposition on the signals. A weighted fusion strategy is then constructed based on energy, Shannon entropy, and variance. This strategy, combining the decomposition advantages of the two wavelet basis functions, preserves high-frequency details while suppressing low-frequency noise, significantly improving the signal reconstruction quality and signal-to-noise ratio.

[0125] The denoised signal is further decomposed using the CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) method to extract intrinsic mode functions (IMFs), retaining the local non-stationary characteristics of the signal in a multi-scale form and enhancing the ability to express subtle abnormal changes.

[0126] Then, for each measurement point, an independent time-frequency domain joint feature extraction and fault diagnosis model is constructed. The model uses the IMF component of the measurement point as input and extracts features from both the time domain and frequency domain perspectives:

[0127] Time domain channel: Use 1D-CNN-LSTM structure to capture the time series pattern of vibration signals;

[0128] Frequency domain channel: First, perform Fourier transform on IMFs, and then extract spectrum features through convolutional neural network;

[0129] Joint channel: The features extracted from the time domain and frequency domain are combined and fused, and then input into the classifier to complete fault identification.

[0130] Each measurement point model is independently trained and output, ultimately generating a .pt format model weight file for subsequent deployment and online fault detection. Compared to traditional multi-channel joint modeling, this invention adopts a point-by-point modeling strategy, which effectively avoids interference between measurement points and improves the diagnostic sensitivity and accuracy of local anomalies.

[0131] The method of the present invention can realize high-resolution, multi-position, and highly robust intelligent fault identification during the operation of the wind turbine. It is particularly suitable for actual working conditions such as the superposition of multiple source weak faults and complex signal transmission paths. It has strong practical value and promotion prospects.

[0132] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. A deep learning method for fan fault diagnosis based on dual wavelet fusion and CEEMDAN decomposition, characterized in that: The method comprises the following steps: S1. Using an acceleration vibration sensor and an acquisition card to collect acceleration vibration signals of the fan group under different operating conditions and create a data set, the acceleration vibration signal includes vibration signals of the motor drive end, the fan drive end, and the fan non-drive end of the fan; S2, improving the wavelet packet denoising algorithm as the basic denoising algorithm to obtain an improved denoising algorithm; S3, using the improved dual wavelet denoising algorithm to denoise the data set in S1; S4. Divide the signal after the noise reduction processing in step S3 into a fixed-length time window; set the length of each sample and the overlap between adjacent samples; and divide the generated samples into three categories based on the positions of the three measurement points of the fan: motor drive end samples, fan drive end samples, and fan non-drive end samples; S5. After the sample division is completed, the three categories of samples are randomly divided independently to obtain their corresponding training sets, validation sets, and test sets to ensure the fairness and accuracy of model training; S6. Use CEEMDAN to decompose all samples of the training set, validation set, and test set in S5 to decompose IMFs to prepare for subsequent model input; S7, input the training set and validation set samples after CEEMDAN decomposition processing into the corresponding three time-frequency domain joint feature extraction models according to their measurement point categories for feature extraction and subsequent modeling; S8. Use the time-frequency domain synchronous feature extraction model to train the training set of the three measurement points in S6 and verify the validation set to obtain the fan fault diagnosis model and output a weight file in .pt format; S9. Load the three weight files generated in S8 and use the test sets of the three measurement points in S6 as input samples to detect the time-frequency domain synchronous extraction feature model. The model outputs of the three measurement points together obtain the detection results of the wind turbine's operating status, and the results include various operating states of the wind turbine.

2. The deep learning blower fault diagnosis method according to claim 1, characterized in that: In S1, the acceleration vibration sensor parameters used include a sensitivity of 100mv / g and a range of 50g. The data acquisition card used is Advantech USB-4711A, which is connected to the acceleration sensor in a single-ended manner. The sampling frequency is 12000Hz, and the measurement range of the acquisition card is set to ±5V. The formula for converting the voltage signal to the acceleration signal is shown below: Where: a(t): acceleration signal, unit: g; V(t): collected voltage signal, unit: mV; S: sensitivity of the sensor, unit: mV / g.

3. The deep learning blower fault diagnosis method according to claim 1 or 2, characterized in that: The process of S2 is as follows: S21, using sym8 wavelet packet function and bior3.3 wavelet packet function to decompose a section of signal respectively; S22. By performing weighted calculation on the normalized combination of energy, variance and Shannon entropy, the fractions of wavelet coefficients decomposed from the two wavelet packets are obtained; S23. Set a threshold. Wavelet coefficients will only be retained when their fraction is greater than the threshold.

4. The deep learning blower fault diagnosis method according to claim 3, characterized in that: In S2, energy reflects the energy distribution of the signal in different frequency bands; variance describes the fluctuation amplitude or discreteness of the signal; Shannon entropy reflects the complexity or disorder of the signal. Considering the importance of energy and variance in identifying signal features, energy and variance are given high weights in the feature fusion process, while Shannon entropy is given low weights. The wavelet packet transform is expressed in the following recursive formula: Assume the original signal is x(t), and the wavelet packet function is Then the wavelet packet coefficient c j+1,2n (k), c j+1,2n+1 (k) is expressed as: Where: c j,n (k): wavelet packet coefficient at the jth layer and the nth node; h(m): low-pass filter coefficient; g(m): high-pass filter coefficient; k: discrete position index; j: decomposition level number; n: node number; The energy calculation formula is shown as follows: Where: E: signal energy; x i : the amplitude of the i-th sampling point; N: the total length of the signal; The variance calculation formula is as follows: Where: V: variance of the signal; x i : the amplitude of the i-th sampling point; The mean of the signal; N: signal length; Shannon entropy formula is shown as follows: Where: H: Shannon entropy, used to measure the complexity or uncertainty of the signal; p i is the probability value in the i-th histogram segment; N: the total component of the histogram; The formula of the combined weighted scoring function is shown as follows: Where: E: energy ratio; E max : The maximum energy value in the current signal set; V: variance; V max : The maximum variance in the current signal set; H: Shannon entropy; H max : The maximum value w of Shannon entropy in the current signal set E , w V , w H : Their respective weights satisfy: In E +in V +in H =1。 5. The deep learning blower fault diagnosis method according to claim 1 or 2, characterized in that: In S6, the number of IMFs decomposition is 7, and the process is as follows: S61. If the number of IMFs decomposed from the sample is less than 7, the sample is directly deleted; S62. If the number of IMFs decomposed from the sample is equal to 7, save the sample directly; S63. If the number of IMFs decomposed from the sample is greater than 7, then based on the weighted average frequency of each IMF, the two IMFs with the closest weighted average frequencies are merged to reach the final number of IMFs.

6. The deep learning blower fault diagnosis method according to claim 5, characterized in that: In S6, the CEEMDANIMF extraction formula is as follows: Where: x(t): original time series signal; c i (t): the i-th intrinsic mode function IMF; N: the number of IMFs; r N (t): final residual or trend term, usually a monotonically changing signal; The IMF weighted average frequency calculation formula is as follows: Assume that the spectrum of a certain IMF component is X(f), then the weighted average frequency f of the IMF is avg Defined as: Where: f i : the i-th frequency point; |X(f i )| 2 : Power spectrum density at the frequency point; N: Total number of frequency points.

7. The deep learning blower fault diagnosis method according to claim 1 or 2, characterized in that: The process of S7 is as follows: S71, the time domain part uses 1D-CNN and LSTM in series to extract features of the time domain signal; S72, the frequency domain part uses FFT transformation and CNN in series to convert the time domain signal into the frequency domain signal and then extract the frequency domain features; S73. Concatenate the LSTM time domain features and the CNN frequency domain features, and finally output the classification results.

8. The deep learning blower fault diagnosis method according to claim 7, characterized in that: In S7 and S71, two layers of 1D-CNN use 3×3 convolution kernels to locally extract the temporal structural features of IMFs; LSTM models the implicit temporal dependencies of IMFs; In S72, a one-dimensional fast Fourier transform is performed on the IMFs of each signal, and its complex modulus is taken as the input feature. The three-layer CNN uses a 3×3 convolution kernel to extract the local frequency domain features between the IMFs; In S73, the features extracted from the time domain and the frequency domain are combined and fused, and input into a classifier to complete fault identification.

Images