Variable-speed motor fault diagnosis method based on spectral kurtosis fusion feature enhancement

Abstract

The invention discloses a variable-speed motor fault diagnosis method based on spectral kurtosis fusion feature enhancement. The method comprises a signal preprocessing module, a fusion feature enhancement module and a network training and recognition module. The signal preprocessing module is used for conducting self-correlation noise reduction, Fourier transformation and fast spectral kurtosis graph transformation processing on collected original vibration signals. Fusion feature enhancement is combined with a frequency domain signal and a spectral kurtosis graph, estimation weight fusion data is calculated through the spectral kurtosis graph, valuable frequency band intervals are screened, and weights are given to enhance features; the network training and recognition adopts a multi-input attention feature fusion network and introduces a spectral kurtosis channel weight to calculate a channel weight of a sample, and finally, the trained network is used for predicting a fault diagnosis result to realize fault diagnosis of the variable-speed motor; the method has the advantages that by highlighting the data features of the signal high-spectrum kurtosis frequency band interval, high recognition precision can still be achieved under the condition of few samples, and the method can be widely applied to the field of variable-speed motor fault diagnosis under the condition of small samples.

Description

Technical Field

[0001] The invention relates to a method for extracting fault features of a transmission, in particular to a method for extracting fault features of a transmission under a small sample. Background Art

[0002] As a key component of the transmission system, the speed changer is widely used in aerospace, weapons, ships, manufacturing and other industrial fields related to national security and national economy. Once a failure or performance degradation occurs, it is very easy to cause the performance of the entire mechanical system to decline or even shut down. In this context, the research and development and application of speed changer fault diagnosis technology is particularly important.

[0003] In the current research context, since the gear transmission system of the speed changer is mostly wrapped in layers of shells inside a complex device, affected by spatial accessibility, it is difficult to install the sensor at the most sensitive position away from the fault point, resulting in complex signal propagation paths, signal attenuation, and weak fault characteristics. Existing fault diagnosis methods often use multi-sensor data fusion methods to integrate data from different sensors and different time points, extract fault characteristics, and use signal enhancement and processing technology for signal attenuation and weak fault characteristics. However, the existing methods only perform adaptive weighted averaging fusion and filtering to remove noise from data from different sources, without fully considering the concentration of fault feature information in a specific area, and will ignore some key fault features, especially those abnormal transient impact components that only exist in the high spectral kurtosis value frequency band area of ​​the frequency domain signal.

[0004] At the same time, in engineering practice, the gearbox operates normally most of the time, and failures are relatively rare. In order to prevent accidents, the machine should be shut down and the faulty parts should be replaced immediately once a failure occurs. The collected normal signals are much more than the fault signals, resulting in an imbalance between the health status data of each category. In order to overcome the imbalance of sample size and maintain high diagnostic accuracy, the existing fault diagnosis network often introduces an attention mechanism to assign different weights to the extracted features for screening. However, the current attention mechanism is too single in calculating attention and cannot fully capture the complex features in the fault signal.

[0005] It is of great significance to study a transmission fault diagnosis method that effectively fuses multi-sensor data, enhances the fault characteristics of signal frequency bands, and uses an attention mechanism to focus the network model on the frequency bands with high spectral kurtosis values ​​of frequency domain signals. Summary of the invention

[0006] In order to address the shortcomings of insufficient data fusion and single attention mechanism in existing methods, the present invention proposes a transmission fault diagnosis method based on spectral kurtosis fusion feature enhancement, which adopts the following technical solutions: Signal preprocessing. This module usesF A vibration sensor collects the original vibration signal of the transmission, and performs autocorrelation noise reduction on the original vibration signal. W The original vibration signals from different signal sources are transformed by fast Fourier transform and fast spectral kurtosis diagram to obtain W The frequency domain signals R from different signal sources w and the highest decomposition level is M The spectral kurtosis plot K w , as shown below,

[0007] Where K w m For the w The signal source m Layer decomposition layer, K wj m For the w The spectral kurtosis diagram of the signal source m Layer decomposition layer j The spectral kurtosis value of the decomposition frequency band interval is obtained by signal preprocessing. W The frequency domain signals R from different signal sources w and spectral kurtosis plot K w Contains sufficient fault feature information to provide a data basis for subsequent feature extraction and fault diagnosis.

[0008] Fusion feature enhancement. This module aims to generate a fusion data set by fusing the frequency domain signal and spectral kurtosis diagram of the vibration sensor through feature enhancement, which includes two steps: data fusion and feature enhancement.

[0009] (1) Data fusion. The frequency domain signals and spectral kurtosis diagrams of different vibration sensors are integrated into fused frequency domain signals and fused spectral kurtosis diagrams. The data fusion method is used to W The frequency domain signals R from different signal sources 1 , R 2 , ..., R w ,...,R W and spectral kurtosis plot K 1 , K 2 , ..., K w , …,K W Data fusion is divided into two steps. The first step is to determine the kurtosis graph K of each spectrum. w The optimal decomposition level K in w m’ (i.e., spectral kurtosis graph K w The decomposition layer K where the maximum value is located w m’ ), the optimal decomposition layer is normalized to obtain the frequency domain signal R wThe estimated weight S w The second step is to W The frequency domain signals R of different vibration sensors w Each point r w n With the estimated weight S w The corresponding weight s w n Perform weighted average fusion to obtain the value r of the corresponding point of the fused frequency domain signal R n , as shown below,

[0010] Where S w For the w The frequency domain signal R of the signal source w The estimated weight of R is the fused frequency domain signal. Take all spectral kurtosis graphs K w The maximum value in is taken as the corresponding position K of the fused spectral kurtosis graph K j m The value is as follows,

[0011] Where K j m is the first m Tier j The frequency domain signal of the vibration sensor and the spectral kurtosis graph data are fused to generate the corresponding fused frequency domain signal R and spectral kurtosis graph K, which are combined into a fused data set O.

[0012] Compared with traditional methods, the data fusion method of the present invention aims at the problem that a single sensor cannot accurately reflect the fault details due to signal attenuation due to the complex structure of the transmission. The estimated weight is obtained by spectral kurtosis calculation and the frequency domain signal is weighted. The recognition of the fault characteristics in the high spectral kurtosis frequency band is enhanced as a whole. The weak or fuzzy fault signal data captured by multiple sensors are integrated into a frequency domain signal R and a fused spectral kurtosis graph K. The combined fusion data set O contains the fault information of different sensors, which is conducive to extracting the key fault characteristics of different sensors in the subsequent feature extraction work.

[0013] (2) Feature enhancement. The fused dataset is enhanced to a feature-enhanced dataset in two steps. The first step is to extract each data O from the fused dataset one by one. i , where each data O i It includes the fused spectral kurtosis graph K and the fused frequency domain signal R. For each layer of the fused spectral kurtosis graph K, m Normalization is performed to obtain the spectral kurtosis interval weight V m, and select each decomposition layer K in the spectral kurtosis graph K m The maximum value of the spectral kurtosis max(K m ), combined into the spectral kurtosis channel weight B, as shown below,

[0014] Where V m j The fusion spectrum kurtosis graph m The spectral kurtosis value of the layer is calculated j The second step is to convert the spectral kurtosis weight V into M Spectral kurtosis interval weight V m Multiplying the points of the fused frequency domain signal one by one obtains the feature-enhanced fault feature sample RV, as shown below:

[0015] Where RV is the feature enhancement sample of the frequency domain signal R, RV m For feature enhancement sample RV M The feature enhanced samples RV generated by the fusion data set O after feature enhancement and the spectral kurtosis channel weight B are combined into the feature enhanced data set Ov, where each feature enhanced data Ov in the feature enhanced data set i Contains feature enhanced samples RV and their corresponding spectral kurtosis channel weights B.

[0016] The feature enhancement method of the present invention combines the spectral kurtosis graph with the frequency domain signal, performs feature enhancement on the interval with higher spectral kurtosis value in the frequency domain signal, and outputs feature sample RV and spectral kurtosis channel weight B. The spectral kurtosis interval weight is weighted and multiplied with the frequency domain signal to generate equivalent feature enhancement samples RV with richer types and more obvious features, solving the problem that insufficient samples make it difficult to train the network under small samples. The spectral kurtosis channel weight B evaluates each subsample RV in RV. m The global features are used to calculate the channel weight of each sample and input into the fault network in the subsequent module for feature extraction and fusion of multiple samples.

[0017] The feature enhanced dataset Ov obtained through the two steps of data fusion and feature enhancement not only contains the fault features from each vibration sensor, but also enhances the fault features, laying the foundation for the subsequent fault identification work of the fault diagnosis network.

[0018] Network training and identification. Building and training a fault diagnosis network to achieve fault classification and identification is divided into three steps: (1) Network construction. The labeled data in the feature enhancement dataset is divided into training set and test set in a ratio of 7:3, and a multi-input attention feature fusion network is built as the fault diagnosis network, including three modules: feature extraction module, channel attention feature fusion module, and fault classification module. The feature extraction module consists of M One-dimensional convolutional pooling layers, each fault feature RV M Subsample RV m Through the feature extraction module M One-dimensional convolutional pooling layer feature extraction M The eigenvector γ (γ 1 , γ 2 , γ 3 …γ M ); The channel attention feature fusion module is composed of three parts: compression, excitation and feature fusion. Compared with the traditional channel attention feature fusion module, the improved channel attention feature fusion module introduces the spectral kurtosis channel weight B. The global vector C obtained by multiplying the spectral kurtosis channel weight B with the global pooling vector A is used to estimate the channel weights of each feature vector. Its various parts are as follows: In the compression part, M The feature vectors are averaged and pooled to obtain the average pooled vector A=(a 1 ,a 2 ,a 3 ,…,a M ), and take the corresponding spectral kurtosis channel weight B=(max(K 1 ) ,max(K 2 ) ,max(K 3 ) ,…, max(K M )), multiply vector A and vector B element by element to get 1×1× M The global vector C is compressed, as shown in the formula:

[0019] The excitation part inputs the global vector C into two fully connected layers with nonlinearity, and then uses the Sigmoid function to generate the attention channel weight D=(d 1 ,d 2 ,d 3 ,…,d M ), the attention channel weights D and M The eigenvector γ = (γ 1 , γ 2 , γ 3 …γ M ) are multiplied and weighted one by one to obtain the eigenvector ζ=(ζ 1 , 2 , 3 …ζM ) completes the incentive, as shown in the formula,

[0020] The feature fusion part will M The eigenvector ζ i The faults are added and fused into a one-dimensional fused fault feature vector η to complete the fault feature fusion, as shown in the formula:

[0021] M The eigenvector γ (γ 1 , γ 2 , γ 3 …γ M ) and the spectral kurtosis channel weights pass through the attention feature fusion module to generate a fused fault feature vector η that contains the feature information of different samples, and the spectral kurtosis channel weights are used to increase the network's attention to high spectral kurtosis features to highlight the fault features of the fused fault feature vector η; the fault classification module consists of two fully connected layers and an LSSVM classifier. The fused fault feature vector η passes through two fully connected layers and is input into the LSSVM classifier to obtain the predicted value p.

[0022] (2) Network training and optimization. When the network has not been trained, the false alarm rate is high, or new data labels are added, the network training process begins. The training samples are input into the established network for network training. The predicted label L obtained by the classifier is compared with the true label P, the loss function L1 is calculated, and the network parameters are optimized accordingly. ϕ ; Finally, the test set is input into the network and the accuracy h is calculated. When the accuracy h reaches the preset standard H, the network training ends.

[0023] (3) Fault identification. The actual data to be classified after feature enhancement is input into the trained fault diagnosis model for prediction. Through model prediction, the classification result of the fault category is obtained, thereby realizing the fault diagnosis of the transmission.

[0024] The beneficial effect of the present invention is as follows: the present invention proposes a transmission fault diagnosis method based on spectral kurtosis fusion feature enhancement. The method uses spectral kurtosis values ​​to estimate the sensitivity of different vibration sensors to transmission faults and integrates multiple sensor signals through multi-source data fusion to enhance the recognition of fault features, expand the diagnosis range and reduce errors. At the same time, the method enhances the fault features of high spectral kurtosis value bands in the frequency domain signal through the spectral kurtosis graph, and introduces the spectral kurtosis channel weight vector in the channel attention mechanism to increase the feature vector channel weight of high spectral kurtosis samples, so that the fault diagnosis model focuses on the frequency band interval of high spectral kurtosis values, thereby solving the problem of difficulty in training the network due to insufficient samples under small sample conditions, and realizing effective and accurate fault diagnosis of transmissions under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flowchart of the transmission fault diagnosis method based on spectral kurtosis fusion feature enhancement of the present invention.

[0026] Figure 2 It is a structural diagram of the speed changer of the present invention.

[0027] Figure 3 This is a flow chart of feature fusion enhancement of the present invention.

[0028] Figure 4 It is the data fusion flow chart of the present invention.

[0029] Figure 5 It is a feature enhancement flow chart of the present invention.

[0030] Figure 6 This is a structural diagram of a fault diagnosis model based on the present invention.

[0031] Figure 7 It is the optimization flow chart of the BKA optimization algorithm of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] See attached Figure 1 , which is a flowchart of the transmission fault diagnosis method based on spectral kurtosis fusion feature enhancement of the present invention. It includes three modules: signal preprocessing, fusion feature enhancement, network training and recognition, and is used for transmission fault diagnosis under small sample conditions.

[0034] Signal preprocessing. Considering the multi-directional vibration characteristics of the transmission case in a complex spatial vibration environment, two triaxial acceleration sensors are installed at both ends of the input shaft and output shaft of the case. This process can be seen in the attached Figure 2 The two sensors can fully capture the vibration of the gearbox in different directions, thus providing more detailed vibration data. The original vibration signals are collected from the two sensors on the input and output ends of the gearbox housing, and the signals are subjected to time-delay autocorrelation noise reduction. The autocorrelation function is defined as,

[0035] Where x(t) is the original signal; T is the signal period; τ is the delay step; and τ is the autocorrelation function of the original signal. After obtaining the autocorrelation sequence of the original signal, the two parts of autocorrelation data where τ = 0 and approaching τ→∞ are eliminated, which can effectively reduce the influence of random noise in the original fault vibration signal.

[0036] The original vibration signal after noise reduction is subjected to fast Fourier transform to obtain the frequency domain signal. At the same time, in order to analyze the characteristics of the vibration signal more deeply, the original vibration signal is decomposed to the highest level. M The fast spectral kurtosis transform is used to obtain the spectral kurtosis plot of each signal.

[0037] In the process of fast spectral kurtosis transformation, a high-pass filter and a low-pass filter are first designed. These two filters are used to filter the original vibration signal to obtain two filtering results. Then, each filtering result is downsampled by 2 times and iterated repeatedly. Each layer of filtering results contains the center frequency and bandwidth information. M The spectral kurtosis of the layer filtering results is calculated, and the calculation formulas for the high-pass filter and the low-pass filter are as follows:

[0038] In the formula, h 1 (t) is a high-pass filter, h 2 (t) is a low-pass filter, h(t) is a standard low-pass filter, and the cut-off frequency is ε. The spectral kurtosis calculation formula is:

[0039] In the formula, c j m Indicates m Tier j The frequency band signal obtained by the filter is its spectral kurtosis value, E(.) represents the expectation, and finally all the calculated spectral kurtosis are integrated to construct the fast spectral kurtosis diagram of the original vibration signal.

[0040] The original vibration signals of the two triaxial vibration sensors are preprocessed to obtain the frequency domain signals and spectral kurtosis diagrams of six different signal sources. These data will be used for subsequent fault diagnosis and analysis.

[0041] Fusion feature enhancement. In order to improve the performance of the fault diagnosis model, the frequency domain signal and spectral kurtosis diagram of the vibration sensor are fused and enhanced to generate a feature-enhanced data set. This process is mainly divided into two steps, such as Figure 3 As shown, the steps are as follows: (1) Data fusion. The frequency domain signals and spectral kurtosis diagrams of more than 6 different signal sources are fused into fused frequency domain signals and fused spectral kurtosis diagrams. See Appendix Figure 4 , which has two steps.

[0042] The first step is to take the frequency domain signal R of each data 1 , R 2 ,…, R 6 and spectral kurtosis plot K 1 , K 2 ,…, K 6 , each spectral kurtosis graph K w The decomposition layer where the mid-spectral kurtosis is the maximum is taken as the optimal decomposition layer K w m’ , and get the optimal decomposition layer K 1 m’ ,K 2 m’ ,…, K 6 m’ , normalize the optimal decomposition layer to calculate the estimated weight S 1 ,S 2 ,…, S 6 , each point r of the frequency domain signal w n and weights w n The weighted average fusion generates the fused frequency domain signal R as shown below.

[0043]

[0044] The second step is to convert the spectral kurtosis graph K 1 , K 2 ,…, K 6 For each data point K wj m The maximum value of the fused spectral kurtosis graph K corresponds to position K j m The value of generates the fused spectral kurtosis graph K as shown below:

[0045] The frequency domain signal and spectral kurtosis graph of the vibration sensor are fused to generate a fused frequency domain signal R and a fused spectral kurtosis graph K. The fused frequency domain signal and the fused spectral kurtosis graph are combined into a fused data set O.

[0046] (2) Feature enhancement. The fusion dataset O is transformed into a feature enhanced dataset O by feature enhancement. V , see attached Figure 5 , divided into two steps: The first step is to fuse each data O of the dataset i It is composed of the fused spectral kurtosis graph K and the fused frequency domain signal R. M Spectral kurtosis layer K M Normalization is performed to obtain the spectral kurtosis interval weight V of each layer M , take each decomposition level K of the spectral kurtosis graph K M The maximum value of the spectral kurtosis max(K m ) is combined into the spectral kurtosis channel weight B, as shown below, V j m The fusion spectrum kurtosis graph m The spectral kurtosis value of the layer is calculated j The spectral kurtosis interval weights of the decomposition frequency bands.

[0047] In the second step, the spectral kurtosis weight V is weighted to the frequency domain signal R to obtain the enhanced fault feature RV, as shown below:

[0048] The feature enhanced signal RV and the spectral kurtosis channel weight B generated after feature enhancement of the fused data set O are combined into a feature enhanced data set Ov; by performing feature enhancement on the fused data set to generate a feature enhanced data set, more powerful data support is provided for subsequent fault diagnosis.

[0049] Network training and identification. This includes building a fault diagnosis network, training the network, and completing fault identification. It is divided into three steps: (1) Build the network. The labels and data of the feature enhancement dataset are divided into training set and test set in a ratio of 7:3, and a multi-input attention feature fusion fault diagnosis network is built as the fault diagnosis network, including feature extraction module, channel attention feature fusion module, and fault classification module. See Appendix Figure 6 ; The feature extraction module consists of M The one-dimensional convolutional pooling layer is composed of multiple input feature enhancement samples RV. M Subsample RV M Through the feature extraction module, feature extraction is performed and the M The eigenvector γ (γ 1, γ 2 , γ 3 …γ M ). The channel attention feature fusion module is divided into three parts: compression, excitation and feature fusion. M The eigenvector γ (γ 1 , γ 2 , γ 3 …γ M ) and the spectral kurtosis channel weight B are input to the channel attention feature fusion module, and a new feature vector ζ is obtained after compression and excitation steps. The feature fusion part M The eigenvector ζ m Additive fusion generates a one-dimensional fused fault feature vector η. The fault classification module consists of two fully connected layers and an LSSVM classifier. The fused fault feature vector η is input into the LSSVM classifier after passing through two fully connected layers to obtain the predicted value p.

[0050] (2) Network training and optimization. When the network has not been trained, the false alarm rate is high, or new labels are added, start the network training process. Input the training set into the built network to obtain the predicted label L, compare the predicted label L with the true label P, and obtain the loss function L1, as shown below:

[0051] l i is the fault label L i The true fault label value, p i For the i The predicted value of each data set is used to continuously adjust the network parameters through the optimization algorithm. ϕ , in order to minimize the loss function L1 loss parameter, the test set is input into the network to calculate the accuracy h. When the accuracy reaches the preset threshold H, the network training is judged to be completed and the network training process is ended.

[0052] The BKA optimization algorithm is used for its optimization algorithm, as shown in the attached Figure 7 As shown. The BKA algorithm is also known as the Black Kite Algorithm. It is inspired by the survival strategy of the Black Kite. The BKA algorithm combines the migration strategy and the attack strategy to improve the global search capability and convergence speed of the algorithm. The BKA algorithm relies on the attack strategy and migration strategy of the Black Kite to realize the global search and local search stages. The algorithm flow is as follows: Initialize the BK matrix to express the number and position of black-winged kites. The BK matrix is ​​as follows:

[0053] Among them, pop is the number of potential solutions, dim is the size of the given problem dimension, BK bi.bj It is bi Black-winged Kite bj Dimensions; During flight, the black-winged kite exhibits two attack states (hovering and diving). As a predator of small grassland mammals and insects, the black-winged kite will adjust its wings and tail angle according to the wind speed during flight, hover quietly to observe its prey, and then quickly dive to attack. This strategy includes different attack behaviors for global exploration and search; on the one hand, the black-winged kite can hover in the air, spread its wings and maintain balance in a hovering attack state, and on the other hand, it can rush towards its prey at an extremely fast speed in a diving attack state. Its attack model is as follows:

[0054] The loss function L1 is used as its adaptability; among them, BK bi.bj bt+1 Indicates bi A black-winged kite bj th Dimensions and bt +1 position in iteration, br A random number represented as (0-1); p b A constant expressed as, T bk Indicates the total number of iterations; During the migration process, a leader L is required to be responsible for the team. The leader L is selected as the network parameter ϕ . Evaluate the fitness of the current population. If the fitness value of the current population is less than the fitness value of the random population, the leader will give up leadership and join the migration population, and vice versa. Its migration behavior model is as follows,

[0055] Among them, C(0,1) represents the Cauchy mutation strategy. bj bt Indicates bt The iteration bj Leader of the Black Winged Kite, F bri Indicates from bt Any black kite in the iteration gets the bj The fitness value of a random position in the dimension, dm is a constant. The BKA algorithm is used to adjust the network parameters ϕ Optimize. Take advantage of the BKA algorithm to solve the problem of local extreme values ​​and optimize network parameters. ϕ .

[0056] (3) Fault identification. The actual data to be classified after data enhancement and its spectral kurtosis channel weights are input into the trained fault diagnosis model. The model outputs the classification result of the fault category according to the input data, and the fault type of the transmission is determined according to the classification result to realize the fault diagnosis of the transmission.

[0057] The beneficial effect of the present invention is as follows: the present invention proposes a transmission fault diagnosis method based on spectral kurtosis fusion feature enhancement. The method uses spectral kurtosis values ​​to estimate the sensitivity of different vibration sensors to transmission faults and integrates multiple sensor signals through multi-source data fusion to enhance the recognition of fault features, expand the diagnosis range and reduce errors. At the same time, the method enhances the fault features of high spectral kurtosis value bands in the frequency domain signal through the spectral kurtosis graph, and introduces the spectral kurtosis channel weight vector in the channel attention mechanism to increase the feature vector channel weight of high spectral kurtosis samples, so that the fault diagnosis model focuses on the frequency band interval of high spectral kurtosis values, thereby solving the problem of difficulty in training the network due to insufficient samples under small sample conditions, and realizing effective and accurate fault diagnosis of transmissions under small sample conditions.

Claims

1. A transmission fault diagnosis method based on spectral kurtosis fusion feature enhancement, including three modules: signal preprocessing, fusion feature enhancement, and network training and recognition. The signal preprocessing is to use a vibration sensor to collect the original vibration signal of the transmission and perform autocorrelation noise reduction, and perform Fourier transform and fast spectral kurtosis map transform on the original vibration signal to generate a frequency domain signal and a spectral kurtosis map to eliminate noise interference and extract useful features of the signal. The method is characterized by: The fusion feature enhancement is to generate a feature enhancement data set by fusing the frequency domain signal and the spectral kurtosis diagram of the vibration sensor, including two steps of data fusion and feature enhancement. The estimated weight of the frequency domain signal is calculated by using the spectral kurtosis diagram to fuse the data layer information of data from different sources, and the valuable frequency band intervals in the fused frequency domain signal are screened out by fusing the spectral kurtosis diagram and assigned different weights to enhance the data features, laying a foundation for the accurate identification of the subsequent fault diagnosis network. The network training and recognition described above is to input the feature enhanced data set and the real label data into the multi-input attention feature fusion network for network training and recognition to realize transmission fault diagnosis, which is carried out in three steps: (1) building a multi-input attention feature fusion network to be used as a fault diagnosis network, (2) inputting the data set and the label data into the fault diagnosis network for network parameter optimization, and (3) inputting the actual data to be classified after the fusion feature enhancement into the trained fault diagnosis network for prediction to obtain the fault diagnosis result.

2. The method for fault diagnosis of a transmission based on spectral kurtosis fusion feature enhancement according to claim 1 is characterized in that The data fusion module integrates the signals of different sensors through multi-source signal fusion to generate a fusion data set O, which is mainly divided into two steps: the first step is to determine the decomposition layer where the maximum value of the spectral kurtosis in the spectral kurtosis diagram of different vibration sensors is located as the optimal decomposition layer, and normalize the optimal decomposition layer to obtain the estimated weight S of the corresponding frequency domain signal w ; The second step is to use the estimated weights S of each vibration sensor source w and frequency domain signal R w The fused frequency domain signal R is generated by weighted average fusion, and the spectral kurtosis graph K of the vibration sensor source is obtained. w The maximum value fusion method is used to generate a fused spectral kurtosis graph K, and the generated fused frequency domain signal R and the fused spectral kurtosis graph K are combined into a fused data set O.

3. The method for fault diagnosis of a transmission based on spectral kurtosis fusion feature enhancement according to claim 1 is characterized in that The feature enhancement module is to enhance the frequency band interval of the frequency domain signal through the spectral kurtosis diagram to generate the feature enhancement data set Ov, which is divided into two steps: the first step is to normalize the spectral kurtosis value of each layer in the fused spectral kurtosis diagram to obtain the spectral kurtosis interval weight, extract the maximum spectral kurtosis value of each layer in the fused spectral kurtosis diagram, and combine the maximum spectral kurtosis value of each layer into the spectral kurtosis channel weight V m ; The second step is to M The spectral kurtosis interval weights are multiplied element by element with the corresponding frequency domain signal to generate M Feature Enhancer Sample RV m , and combine the feature enhanced sample RV and its spectral kurtosis channel weight B into the feature enhanced dataset Ov.

4. The method for fault diagnosis of a transmission based on spectral kurtosis fusion feature enhancement according to claim 1 is characterized in that The multi-input attention feature fusion network includes three parts: feature extraction module, channel attention feature fusion module and fault classification module; the feature extraction module M Feature Enhancer Sample RV m Through M One-dimensional convolutional pooling layer for feature extraction; channel The attention feature fusion module introduces the spectral kurtosis channel weight B and the attention mechanism. M The feature vectors are weighted updated and then feature fused to output the fused fault feature vector; the fault classification module consists of two fully connected layers and an LSSVM classifier, which determines the fault type according to the fused fault feature vector and outputs the predicted value and the fault classification result of the data.

Images