Bearing diagnosis method with deep attention mechanism enhanced fault sensitive feature extraction

Abstract

This invention belongs to the field of equipment fault monitoring and diagnosis technology, specifically relating to a bearing diagnosis method that enhances fault-sensitive feature extraction using a deep attention mechanism. The method includes the following steps: acquiring time-domain signal data of the bearing under different fault categories; inputting the squared envelope of the time-domain signal into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information; filtering the key frequency band information as a filtering criterion to obtain a filtered attention-enhanced signal; performing a maximum overlap discrete wavelet transform on the attention-enhanced signal to obtain a sub-signal; optimizing the demodulation frequency band using the autocorrelation kurtosis as an indicator to obtain the optimal demodulated sub-signal; and calculating the squared envelope spectrum of the optimal demodulated sub-signal to achieve enhanced extraction and fault diagnosis of bearing fault-sensitive features. The beneficial effects are: further enhancing information strongly correlated with the fault in the signal and effectively extracting bearing fault signal features under strong background noise.

Description

Technical Field

[0001] This invention belongs to the field of equipment fault monitoring and diagnosis technology, specifically relating to a bearing diagnosis method that enhances fault-sensitive feature extraction using a deep attention mechanism. Background Technology

[0002] Gearboxes, as a crucial component of power units, often receive significant attention. Rotating components within gearboxes, such as bearings, shafts, and gears, frequently operate under complex variable load conditions. Consequently, these components often develop minor faults early on due to fatigue damage, and these fault signals are often difficult to detect in strong background noise. Allowing these minor faults to gradually worsen can not only lead to severe damage to mechanical equipment but may even cause personal injury or death. Therefore, research into condition monitoring and fault diagnosis methods for critical components such as bearings is of great importance for the safe and reliable operation of equipment.

[0003] With the rapid development of sensor and computer technologies, vibration signal feature extraction methods, as a crucial step in fault diagnosis, are also constantly improving. Designing and constructing new methods suitable for extracting fault features based on the operating characteristics of mechanical equipment, changes in vibration signal characteristics, and the advantages of different algorithms has become a research hotspot in the field of fault diagnosis. Meanwhile, with the increasing size and intelligence of power units, the vibration signal transmission paths are becoming more complex, background noise is becoming stronger, and fault feature extraction is becoming more difficult. Therefore, research on weak feature extraction and fault diagnosis of complex vibrations under strong noise will be an important trend in future research. Summary of the Invention

[0004] The purpose of this invention is to provide a bearing diagnosis method that enhances fault-sensitive feature extraction through a deep attention mechanism, which helps to improve the accuracy of bearing fault diagnosis in actual field operations.

[0005] The technical solution of the present invention is as follows: a bearing diagnosis method based on deep attention mechanism-enhanced fault-sensitive feature extraction, comprising the following steps:

[0006] Step 101: Collect time-domain signal data of the bearing under different fault categories;

[0007] Step 102: After squaring the time-domain signal, input it into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information;

[0008] Step 103: Use the key frequency band information as the basis for filtering to obtain the filtered attention-enhanced signal;

[0009] Step 104: Perform maximum overlap discrete wavelet transform on the attention enhancement signal to obtain sub-signals;

[0010] Step 105: Calculate the unbiased autocorrelation kurtosis of the squared envelope of the sub-signal;

[0011] Step 106: Using the autocorrelation kurtosis as an indicator, optimize the demodulation frequency band to obtain the optimal demodulated sub-signal;

[0012] Step 107: Calculate the squared envelope spectrum of the optimal demodulation sub-signal to achieve enhanced extraction and fault diagnosis of bearing fault sensitive features.

[0013] In step 101, the fault time-domain signal is acquired through a bearing fault simulation test bench to obtain time-domain signal data under different fault categories. The bearing model used is a 30310 tapered roller bearing. Before the experiment, the bearings in four healthy states, including inner ring fault, outer ring fault, rolling element fault, and normal bearing, are processed by wire cutting. The dimensions of the above faults are all 1 mm in width and 0.5 mm in depth. Under the working conditions of 1200 r / min and 60 N·m load, the time-domain data of the four healthy states are acquired at a sampling frequency of 1000 kHz.

[0014] In step 102, the time-domain signal is squared and then input into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information.

[0015] In step 102, the bearing time-domain signal is subjected to Hilbert envelope, and the square of the envelope spectrum amplitude is taken as the input of the deep attention mechanism network; the fault signal length N = 1,000,000 collected by the sensor is squared, and the first 4,000 points of the envelope spectrum are taken as the neural network input. The construction method of the multi-scale CNN network based on the deep attention mechanism is as follows: First, a sparse attention layer with weights is established. This layer uses the fully connected layer as a template and defines its weights W. s The absolute value is restricted to a certain numerical range [0,1]. An additional vector m is introduced, which has only two elements: 0 and 1. After multiplying the vector m by the corresponding position of the weight, the output of the layer is calculated using Formula 1.

[0016] Formula 1:

[0017] x out =m⊙(W s ⊙x),m,x,W s ∈R N

[0018] Where ⊙ indicates that the calculation method is to multiply corresponding positions;

[0019] The m-vector enables controllable sample output, thus achieving a sparse attention mechanism. The m-vector serves as the external input, and its iterative update method is as follows: First, initialize m and W. s After the first training cycle, obtain the trained W value.s Based on W s Update m according to the sparse rule and perform the second training cycle; and so on, until the predetermined number of training cycles T or the accuracy drops to a specified threshold.

[0020] Secondly, W in the sparse attention layer s The absolute value of each weight is a measure of the importance of its corresponding input sequence in the entire model operation. Therefore, a threshold β is defined in the sparsity rule, setting the β term with the smallest absolute value of weight in m to 0, and the rest to 1. That is, the features most sensitive to faults are selected through iterative sparsity, as shown in Formula 2.

[0021] Formula 2:

[0022]

[0023] The sparse attention mechanism requires specifying the total number of training iterations T, a sparsity rule based on the threshold β and the number of training iterations T; and defining the current number of training iterations t, where the sparsity increases with the number of training iterations, as shown in Formula 3.

[0024] Formula 3:

[0025]

[0026] Among them, s p Sparsity refers to the proportion of zero elements in vector m.

[0027] Finally, after training at t=T, the optimal result is selected based on accuracy as the primary metric and sparsity as the secondary metric to obtain key frequency band information.

[0028] In step 103, the key frequency band information is used as the filtering basis to obtain the filtered attention enhancement signal. Based on the key frequency band [0, 18kHz] obtained by the CNN network based on the deep attention mechanism in step 102, the original signal is low-pass filtered to obtain the attention enhancement signal containing only the key frequency band [0, 18kHz].

[0029] In step 104, the attention-enhanced signal is subjected to a maximum overlap discrete wavelet transform (MODWPT) to obtain a sub-signal. The filtered attention-enhanced time-domain signal is then subjected to a MODWPT, which is a highly redundant non-orthogonal wavelet transform. The scaling filter in MODWPT... and wavelet filter The relationship is shown in Formula 4:

[0030] Formula 4:

[0031]

[0032]

[0033] Among them, g l and h l These are the scaling filter and wavelet filter in Discrete Wavelet Transform (DWT), respectively.

[0034] The scaling coefficients V at scale j are obtained using Mallat's algorithm. j,t and wavelet transform coefficients W j,t As shown in Formula 5;

[0035] Formula 5:

[0036]

[0037]

[0038] Where N is the dimension of the original signal, mod represents the remainder when the two numbers are divided, and j represents the number of decomposition layers;

[0039] The MODWPT decomposition coefficients are calculated as shown in Formula 6;

[0040] Formula Six:

[0041]

[0042] Where, if nmod4 = 0 or 3, then If nmod4 = 1 or 2, then

[0043] After the above calculations, the sub-signals after the maximum overlap discrete wavelet transform can be obtained.

[0044] In step 105, the unbiased autocorrelation kurtosis of the squared envelope of the sub-signal is calculated, the sub-signal is squared, and the unbiased autocorrelation of the enveloped signal is calculated according to Formula 7.

[0045] Formula 7:

[0046]

[0047] Where X(t) i ) represents the subsignal x(t) i The square envelope of ); τ = q / f s Indicates time delay; N represents the number of sampling points; q = 0, 1, ..., N-1;

[0048] The autocorrelation kurtosis is calculated as shown in Formula 8;

[0049] Formula 8:

[0050]

[0051] In step 106, the demodulation frequency band is optimized using the autocorrelation kurtosis as an index to obtain the optimal demodulation sub-signal. After calculating the autocorrelation kurtosis of all sub-signals, the node with the largest autocorrelation kurtosis is found, and the signal of the demodulation frequency band of the node is selected as the optimal demodulation sub-signal.

[0052] In step 107, the optimal demodulated sub-signal squared envelope spectrum is calculated to achieve enhanced extraction and fault diagnosis of bearing fault sensitive features. The optimal frequency band sub-signal is subjected to Hilbert envelope, and the amplitude of the envelope spectrum is squared to obtain the envelope spectra of bearing signals with inner ring fault, outer ring fault, and rolling element fault after processing by the proposed method, thereby achieving enhanced extraction and fault diagnosis of bearing fault sensitive features.

[0053] The beneficial effects of this invention are as follows: This invention addresses the issue of strong background noise and low signal-to-noise ratio in early rolling bearing fault signals. The proposed signal enhancement method for bearing fault diagnosis utilizes a multi-scale CNN network based on sparse attention mechanism to mine key frequency bands. The autocorrelation kurtosis of the square envelope is used as the frequency band optimization index to further enhance information strongly correlated with the fault in the signal and effectively extract bearing fault signal features under strong background noise. Attached Figure Description

[0054] Figure 1 This is a flowchart of the bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism, provided in an embodiment of the present invention.

[0055] Figure 2 This is a normal bearing time-domain diagram and spectrum diagram of the bearing diagnosis method with enhanced fault-sensitive feature extraction using a deep attention mechanism provided in this embodiment of the invention.

[0056] Figure 3 This is a time-domain diagram and spectrum diagram of an inner ring fault bearing in the bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism, as provided in this embodiment of the invention.

[0057] Figure 4 This is a time-domain diagram and spectrum of an outer ring fault bearing in a bearing diagnosis method that enhances fault-sensitive feature extraction using a deep attention mechanism, as provided in this embodiment of the invention.

[0058] Figure 5 This is a time-domain diagram and spectrum diagram of a bearing with rolling element faults, provided in an embodiment of the present invention, using a deep attention mechanism to enhance the extraction of fault-sensitive features.

[0059] Figure 6This is a deep attention mechanism network structure diagram of the bearing diagnosis method that enhances fault-sensitive feature extraction using a deep attention mechanism, as provided in this embodiment of the invention.

[0060] Figure 7 This is the envelope spectrum of the inner ring fault bearing signal processed by the proposed method in the bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism provided in this embodiment of the invention.

[0061] Figure 8 This is the envelope spectrum of an outer ring fault bearing signal processed by the proposed method in a bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism, as provided in this embodiment of the invention.

[0062] Figure 9 The envelope spectrum of the rolling element fault bearing signal processed by the proposed method is provided in the bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism, as provided in this embodiment of the invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0064] like Figure 1 As shown, the bearing diagnosis method using a deep attention mechanism to enhance fault-sensitive feature extraction includes the following steps:

[0065] Step 101: Collect time-domain signal data of the bearing under different fault categories;

[0066] Step 102: After squaring the time-domain signal, input it into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information;

[0067] Step 103: Use the key frequency band information as the basis for filtering to obtain the filtered attention-enhanced signal;

[0068] Step 104: Perform maximum overlap discrete wavelet transform on the attention enhancement signal to obtain sub-signals;

[0069] Step 105: Calculate the unbiased autocorrelation kurtosis of the squared envelope of the sub-signal;

[0070] Step 106: Using the autocorrelation kurtosis as an indicator, optimize the demodulation frequency band to obtain the optimal demodulated sub-signal;

[0071] Step 107: Calculate the squared envelope spectrum of the optimal demodulation sub-signal to achieve enhanced extraction and fault diagnosis of bearing fault sensitive features.

[0072] In step 101, in this embodiment, the fault time-domain signal is acquired through a bearing fault simulation test bench to obtain time-domain signal data under different fault categories. In this embodiment, the bearing model used is a 30310 tapered roller bearing. Before the experiment, wire cutting is used to process bearings in four healthy states, including inner ring fault, outer ring fault, rolling element fault, and normal bearing. The dimensions of the above faults are all 1mm wide and 0.5mm deep. Under the conditions of 1200r / min speed and 60N·m load, time-domain data of the four healthy states are acquired at a sampling frequency of 1000kHz. Figures 2-5 The images show the time-domain waveforms and frequency spectra of bearings with normal operation, inner ring faults, outer ring faults, and rolling element faults, respectively, using a bearing diagnostic method based on a deep attention mechanism to enhance fault-sensitive feature extraction, as provided in this embodiment. It can be seen that the time-domain waveforms of different fault categories exhibit certain differences.

[0073] In step 102, the time-domain signal is squared and then input into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information. The neural network structure based on the deep attention mechanism is as follows: Figure 6 As shown in Table 1, the network structure parameters are as follows.

[0074] Table 1. Neural network model structure parameters for deep attention mechanism

[0075]

[0076] The bearing time-domain signal is subjected to a Hilbert envelope, and the square of the envelope spectrum amplitude is used as the input to the deep attention mechanism network. In this embodiment, the fault signal length N = 1,000,000 is collected by the sensor. After squaring the envelope, the first 4,000 points of the envelope spectrum (the frequency range to be focused on) are used as the neural network input. The construction method of the multi-scale CNN network based on the deep attention mechanism is as follows: First, a sparse attention layer with weights is established. This layer uses a fully connected layer as a template and defines its weights W. s The absolute value is restricted to a certain numerical range [0,1]. An additional vector m is introduced, which has only two elements: 0 and 1. After multiplying the vector m by the corresponding position of the weight, the output of the layer is calculated using Formula 1.

[0077] Formula 1:

[0078] x out =m⊙(W s ⊙x),m,x,W s ∈R N

[0079] Here, ⊙ indicates that the calculation method is to multiply corresponding positions.

[0080] The m-vector enables controllable sample output, thus achieving a sparse attention mechanism. The m-vector serves as the external input, and its iterative update method is as follows: First, initialize m and W. s After the first training cycle, obtain the trained W value. s Based on W s Update m according to the sparse rule and perform a second training iteration. Continue in this manner until the predetermined number of training iterations T or the accuracy drops to a specified threshold.

[0081] Secondly, W in the sparse attention layer s The absolute value of each weight is a measure of the importance of its corresponding input sequence in the entire model operation. Therefore, a threshold β is defined in the sparsity rule, setting the β term with the smallest absolute value of weight in m to 0, and the rest to 1. That is, the features most sensitive to faults are selected through iterative sparsity, as shown in Formula 2.

[0082] Formula 2:

[0083]

[0084] The sparse attention mechanism requires specifying the total number of training iterations T, and a sparsity rule based on the threshold β and the number of training iterations T. The current number of training iterations t is defined, and the sparsity increases with the number of training iterations, as shown in Formula 3.

[0085] Formula 3:

[0086]

[0087] Among them, s p Sparsity refers to the proportion of zero elements in the m-vector.

[0088] Finally, after training at t=T, the optimal result is selected based on accuracy as the primary metric and sparsity as the secondary metric to obtain key frequency band information.

[0089] In step 103, the key frequency band information is used as the filtering basis to obtain the filtered attention-enhanced signal. Based on the key frequency band [0, 18kHz] obtained by the CNN network based on the deep attention mechanism in step 102, the original signal is low-pass filtered to obtain an attention-enhanced signal containing only the key frequency band [0, 18kHz].

[0090] In step 104, the attention-enhanced signal is subjected to a maximum overlap discrete wavelet transform (MODWPT) to obtain a sub-signal. The filtered attention-enhanced time-domain signal is then subjected to MODWPT. MODWPT is a highly redundant non-orthogonal wavelet transform, and the scaling filter in MODWPT... and wavelet filter The relationship is shown in Formula 4:

[0091] Formula 4:

[0092]

[0093]

[0094] Among them, g l and h l These are the scaling filter and wavelet filter in Discrete Wavelet Transform (DWT), respectively.

[0095] The scaling coefficients V at scale j are obtained using Mallat's algorithm. j,t and wavelet transform coefficients W j,t As shown in Formula 5;

[0096] Formula 5:

[0097]

[0098]

[0099] Where N is the dimension of the original signal, mod represents the remainder when the two numbers are divided, and j represents the number of decomposition layers.

[0100] The MODWPT decomposition coefficients are calculated as shown in Formula 6;

[0101] Formula Six:

[0102]

[0103] Where, if nmod4 = 0 or 3, then If nmod4 = 1 or 2, then

[0104] After the above calculations, the sub-signals after the maximum overlap discrete wavelet transform can be obtained.

[0105] In step 105, the unbiased autocorrelation kurtosis of the squared envelope of the sub-signal is calculated. The sub-signal is squared, and the unbiased autocorrelation of the enveloped signal is calculated according to Formula 7.

[0106] Formula 7:

[0107]

[0108] Where X(t) i ) represents the subsignal x(t) i The square envelope of ); τ = q / f s q represents time delay; N represents the number of sampling points; q = 0, 1, ..., N-1.

[0109] The autocorrelation kurtosis is calculated as shown in Formula 8;

[0110] Formula 8:

[0111]

[0112] In step 106, the demodulation frequency band is optimized using the autocorrelation kurtosis as an indicator to obtain the optimal demodulated sub-signal. After calculating the autocorrelation kurtosis for all sub-signals, the node with the largest autocorrelation kurtosis is found, and the signal of the demodulation frequency band of that node is selected as the optimal demodulated sub-signal.

[0113] In step 107, the optimal demodulated sub-signal squared envelope spectrum is calculated to achieve enhanced extraction and fault diagnosis of bearing fault sensitive features. The optimal frequency band sub-signal is subjected to Hilbert envelope processing, and the envelope spectrum amplitude is squared to obtain the envelope spectra of bearing signals with inner ring faults, outer ring faults, and rolling element faults processed by the proposed method, as shown below. Figures 7-9 As shown, this enables enhanced extraction of bearing fault-sensitive features and fault diagnosis.

[0114] This invention addresses the issue of strong background noise and low signal-to-noise ratio in early rolling bearing fault signals. It proposes a signal enhancement method for bearing fault diagnosis, which mines key frequency bands using a multi-scale CNN network based on sparse attention mechanism and uses the autocorrelation kurtosis of the squared envelope as a frequency band optimization index to further enhance information strongly correlated with the fault in the signal and effectively extract bearing fault signal features under strong background noise.

[0115] The above description is merely a preferred embodiment of this application and an explanation of the technical principles used. Those skilled in the art should understand that the scope involved in this application is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the inventive concept. For example, technical solutions formed by replacing the above-mentioned features with technical features with similar functions disclosed in this application (but not limited to) each other.

[0116] Apart from the technical features described in the specification, the other technical features are known to those skilled in the art. To highlight the innovative features of this invention, the other technical features will not be described in detail here.

Claims

1. A bearing diagnosis method that enhances fault-sensitive feature extraction through deep attention mechanism, characterized in that, Includes the following steps: Step 101: Collect time-domain signal data of the bearing under different fault categories; Step 102: After squaring the time-domain signal, input it into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information; In step 102, the time-domain signal is squared and then input into a multi-scale CNN network based on a deep attention mechanism to mine key frequency band information. In step 102, the bearing time-domain signal is subjected to a Hilbert envelope, and the square of the envelope spectrum amplitude is used as the input to the deep attention mechanism network; the length of the fault signal acquired by the sensor... =1000000, after squaring the envelope, the first 4000 points of the envelope spectrum are taken as the input of the neural network. The construction method of the multi-scale CNN network based on the deep attention mechanism is as follows: First, a sparse attention layer with weights is established. This layer uses the fully connected layer as a template and defines its weights. The absolute value is restricted to a certain numerical range [0,1], introducing an additional vector. , The layer contains only two elements, 0 and 1. After multiplying them by the corresponding positions of the weights, the output of the layer is calculated using Formula 1. Formula 1: ， in, This indicates that the calculation method involves multiplying corresponding positions. pass Vectors enable controllable sample output, thus achieving a sparse attention mechanism; The vector is used as external input, and its iterative update method is as follows: First, initialize... , After the first training cycle, obtain the post-training data. ;based on Update according to sparse rules Then proceed to the second cycle of training; and so on, until the predetermined number of cycles is reached. Or the accuracy drops to a specified threshold; Secondly, in the sparse attention layer The absolute value of each weight is a measure of the importance of its corresponding input sequence in the entire model operation. Therefore, a threshold is defined in the sparsity rule. ,make The one with the smallest absolute value of weight The term is 0, and the rest are 1, that is, the most sensitive features to faults are selected by iterative sparsity, as shown in Formula 2; Formula 2: ， The sparse attention mechanism described above requires specifying the total number of training iterations. Threshold-based and number of training cycles The sparsity rule; defining the current number of training iterations. The sparsity increases with the number of training iterations, as shown in Formula 3. Formula 3: ， in, Sparsity, i.e. The proportion of zero elements in a vector; Ultimately, when After training, with accuracy as the primary metric and sparsity as the secondary metric, the optimal result is selected to obtain key frequency band information. Step 103: Use the key frequency band information as the basis for filtering to obtain the filtered attention-enhanced signal; Step 104: Perform maximum overlap discrete wavelet transform on the attention enhancement signal to obtain sub-signals; Step 105: Calculate the unbiased autocorrelation kurtosis of the squared envelope of the sub-signal; Step 106: Using the autocorrelation kurtosis as an indicator, optimize the demodulation frequency band to obtain the optimal demodulated sub-signal; Step 107: Calculate the squared envelope spectrum of the optimal demodulation sub-signal to achieve enhanced extraction and fault diagnosis of bearing fault sensitive features.

2. The bearing diagnosis method using deep attention mechanism to enhance fault-sensitive feature extraction as described in claim 1, characterized in that: In step 101, the fault time-domain signal is acquired through a bearing fault simulation test bench to obtain time-domain signal data under different fault categories. The bearing model used is a 30310 tapered roller bearing. Before the experiment, the bearings in four healthy states, including inner ring fault, outer ring fault, rolling element fault, and normal bearing, are processed by wire cutting. The dimensions of the above faults are all 1 mm in width and 0.5 mm in depth. Under the working conditions of 1200 r / min and 60 N·m load, the time-domain data of the four healthy states are acquired at a sampling frequency of 1000 kHz.

3. The bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism as described in claim 1, characterized in that: In step 103, the key frequency band information is used as the filtering basis to obtain the filtered attention enhancement signal. Based on the key frequency band [0, 18kHz] obtained by the CNN network based on the deep attention mechanism in step 102, the original signal is low-pass filtered to obtain the attention enhancement signal containing only the key frequency band [0, 18kHz].

4. The bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism as described in claim 1, characterized in that: In step 104, the attention-enhanced signal is subjected to a maximum overlap discrete wavelet transform (MODWPT) to obtain a sub-signal. The filtered attention-enhanced time-domain signal is then subjected to a MODWPT, which is a highly redundant non-orthogonal wavelet transform. The scaling filter in MODWPT... and wavelet filter The relationship is shown in Formula 4: Formula 4: ， ， in, and These are the scaling filter and wavelet filter in the discrete wavelet transform, respectively. The scale is calculated using Mallat's algorithm. Scale transformation coefficients under and wavelet transform coefficients As shown in Formula 5; Formula 5: ， ， in, The dimension of the original signal. This represents the remainder when two numbers are divided. Indicates the number of decomposition levels; The MODWPT decomposition coefficients are calculated as shown in Formula 6; Formula Six: ， Among them, if ,but ;if ,but ; After the above calculations, the sub-signals after the maximum overlap discrete wavelet transform can be obtained.

5. The bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism as described in claim 1, characterized in that: In step 105, the unbiased autocorrelation kurtosis of the squared envelope of the sub-signal is calculated, the sub-signal is squared, and the unbiased autocorrelation of the enveloped signal is calculated according to Formula 7. Formula 7: ， in, Represents sub-signal The square envelope; Indicates time delay; Indicates the number of sampling points; ; The autocorrelation kurtosis is calculated as shown in Formula 8; Formula 8: 。 6. The bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism as described in claim 1, characterized in that: In step 106, the demodulation frequency band is optimized using the autocorrelation kurtosis as an index to obtain the optimal demodulation sub-signal. After calculating the autocorrelation kurtosis of all sub-signals, the node with the largest autocorrelation kurtosis is found, and the signal of the demodulation frequency band of the node is selected as the optimal demodulation sub-signal.

7. The bearing diagnosis method for enhancing fault-sensitive feature extraction using a deep attention mechanism as described in claim 1, characterized in that: In step 107, the optimal demodulated sub-signal squared envelope spectrum is calculated to achieve enhanced extraction and fault diagnosis of bearing fault sensitive features. The optimal frequency band sub-signal is subjected to Hilbert envelope, and the amplitude of the envelope spectrum is squared to obtain the envelope spectra of bearing signals with inner ring fault, outer ring fault, and rolling element fault after processing by the proposed method, thereby achieving enhanced extraction and fault diagnosis of bearing fault sensitive features.

Images