A bevel gear fault diagnosis method for an aero-engine based on a wavelet denoising network

Abstract

The present disclosure discloses a kind of wavelet denoising network-based aero-engine bevel gear fault diagnosis method, comprising the following steps: S100: the vibration signal of aero-engine bevel gear to be measured is collected and segmented;S200: constructing fault diagnosis model and training;S300: based on the vibration signal of the segmented vibration signal of the trained fault diagnosis model is detected, and the probability of different fault categories of vibration signal is output, wherein the maximum value of probability is the actual fault of bevel gear.This disclosure fully considers the advantages and disadvantages of signal processing technology and deep learning network for fault diagnosis, by introducing wavelet transform, zero denoising, different scale wavelet coefficient selection technology, the data driving ability of signal processing and deep learning network is organically combined, the anti-noise, diagnostic accuracy and diagnostic stability of fault diagnosis technology are improved.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the field of signal processing analysis and fault diagnosis, and particularly relates to a bevel gear fault diagnosis method for an aero-engine based on a wavelet denoising network. BACKGROUND

[0002] As the heart of an airplane, the reliable operation of an aero-engine is an important link to ensure the safety of an airplane. As a power transmission link of an aero-engine, real-time evaluation of the health status of a bevel gear is very important for evaluating the reliability of a transmission system. At present, with the development of intelligent fault diagnosis, deep networks play a dominant role in building intelligent fault diagnosis models. Although there are some intelligent algorithms for aero-engine bevel gear fault diagnosis at present, an aero-engine belongs to major equipment, and its installation and support technology needs to have sufficient reliability. However, a diagnosis model based on deep learning is a data-driven black box model, and the algorithm construction is difficult to have high stability and reliability as general signal processing technology, which cannot meet the needs of service safety support technology of an aero-engine.

[0003] At the same time, the complex structure of an aero-engine leads to a complex signal transmission path and much interference, there are few vibration measuring points of an engine, and multi-source signal coupling is serious. The above factors result in poor data quality and much noise interference. Although a deep learning network can realize end-to-end automatic feature extraction, it has insufficient noise resistance and is prone to overfitting to noise. In the case of poor data quality and much noise interference, it is difficult to extract fault features. Therefore, it does not have strong generalization, and it is difficult to be applied to actual fault diagnosis.

[0004] In addition, there are some technologies that use signal processing technology as pre-processing, and input the extracted features into a deep network to improve the generalization performance of the network. However, this diagnosis mode has high requirements for the data pre-processing stage, and needs strong professional knowledge in the relevant field and time-consuming parameter selection and adjustment to obtain good features through relevant signal processing technology.

[0005] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore can contain information that does not constitute prior art known to those of ordinary skill in the art in the country. SUMMARY

[0006] In view of the deficiencies in the prior art, the purpose of the present disclosure is to provide a bevel gear fault diagnosis method for an aero-engine based on a wavelet denoising network. The method can improve the noise resistance, diagnosis accuracy and diagnosis stability of fault diagnosis by introducing wavelet transform, zeroing denoising and different scale wavelet coefficient selection technology.

[0007] To achieve the above purpose, the present disclosure provides the following technical solutions:

[0008] The application discloses an aero-engine bevel gear fault diagnosis method based on a wavelet denoising network.

[0009] S100: collecting vibration signals of a to-be-tested aero-engine bevel gear and segmenting the vibration signals;

[0010] S200: constructing a fault diagnosis model and training the fault diagnosis model;

[0011] S300: detecting the segmented vibration signals based on the trained fault diagnosis model, and outputting probabilities of different fault categories of the vibration signals, wherein a maximum value of the probabilities is a fault category of the bevel gear.

[0012] Preferably, in the step S200, the fault diagnosis model comprises:

[0013] a wavelet convolution module, configured to perform one-dimensional wavelet convolution decomposition on the segmented vibration signals to obtain wavelet coefficients;

[0014] a denoising module, configured to perform zero denoising processing on the wavelet coefficients to obtain denoised wavelet coefficients;

[0015] a weighting module, configured to obtain weights capable of representing importance of the denoised wavelet coefficients, and to weight the denoised wavelet coefficients according to the weights to obtain weighted wavelet coefficients;

[0016] a recognition module, configured to obtain probabilities of different fault states according to the weighted wavelet coefficients, so as to realize fault recognition of the bevel gear.

[0017] Preferably, the segmented vibration signals are subjected to one-dimensional wavelet convolution decomposition through one-dimensional wavelet convolution, and the one-dimensional wavelet convolution is represented as:

[0018] W = x * ψ'(a)

[0019] wherein W is the wavelet coefficient obtained through one-dimensional convolution, the dimension of W is C*L, C is a number of wavelet bases, L is a length of the obtained wavelet coefficient, * is convolution, x is a vibration signal sample segmented and divided, ψ'(.) is a convolution kernel composed of linearly weighted fusion of a plurality of wavelet bases, and an independent variable of ψ'(.) is a scale parameter a.

[0020] Preferably, the wavelet coefficients are subjected to zero denoising processing through the following steps:

[0021] S2011: obtaining a scale-in relationship of the wavelet coefficients through one-dimensional convolution conv(.) and outputting a judgment of whether to perform zero processing;

[0022] S2012: obtaining a scale-out relationship of the wavelet coefficients through a full connection network fc(·) and outputting a judgment of whether to perform zero processing;

[0023] S2013: judging whether the in-scale and inter-scale zero processing is performed in steps S2011 and S2022, calculating the probability φ0 of performing zero processing and the probability φ1 of reserving the original value without performing zero processing corresponding to the wavelet coefficient;

[0024] S2014: performing zero denoising operation on the wavelet coefficient according to the probability [φ0, φ1] to obtain the denoised wavelet coefficient.

[0025] Preferably, the weighted wavelet coefficient is obtained by the following steps:

[0026] S2021: calculating the energy of the wavelet coefficient of each scale;

[0027] S2022: obtaining the importance index of the wavelet coefficient of each scale by one-dimensional convolution based on the energy of the wavelet coefficient of each scale;

[0028] S2023: performing linear weighting on the zero-processed wavelet coefficient of each scale based on the importance index to obtain the weighted wavelet coefficient.

[0029] Preferably, in step S200, the training process of the fault diagnosis model comprises the following steps:

[0030] S201: segmenting the vibration signal of known fault, dividing the segmented vibration signal into a training set and a test set, and labeling the corresponding fault state label;

[0031] S202: training the fault diagnosis model using the training set, when the network training loss is minimum, the model training is completed; otherwise, adjusting the network parameters and retraining until the network training loss reaches the minimum;

[0032] S203: testing the trained model using the test set, when the test accuracy rate reaches 90% or above, the test is completed, and the trained fault diagnosis model is obtained.

[0033] Preferably, in step S202, the network training loss is denoted as L, which is expressed as:

[0034]

[0035] wherein, L cls is the fault classification diagnosis loss calculated according to the cross entropy, N is the number of bevel gear fault categories, y i and are the label value and the network output fault probability corresponding to the i-th fault type respectively, L DR and L SKrespectively, are a spectrum kurtosis optimization loss for controlling the loss of noise reduction rate and scale optimization, and alpha and beta are hyperparameters that can be optimized through grid search, y = [y1, y2,..., y N ] is a true fault state label corresponding to the data x.

[0036] The present disclosure also provides an aero-engine bevel gear fault diagnosis device based on a wavelet denoising network, comprising:

[0037] A collection module is configured to collect vibration signals of a to-be-tested aero-engine bevel gear and segment the vibration signals.

[0038] A training module is configured to train the constructed fault diagnosis model.

[0039] A detection module is configured to detect the segmented vibration signals based on the trained fault diagnosis model to output probabilities of different fault categories of the vibration signals.

[0040] The present disclosure also provides a computer medium, comprising:

[0041] A memory is configured to store a plurality of computer instructions.

[0042] A processor is configured to instruct the computer instructions to implement the method according to any one of the preceding embodiments.

[0043] Compared with the prior art, the present disclosure has the following beneficial effects:

[0044] 1. Compared with the traditional method, the present disclosure fully considers the advantages and disadvantages of signal processing technology and deep learning network for fault diagnosis, and combines the data-driven ability of signal processing and deep learning network by introducing wavelet transform, zero denoising, and different scale wavelet coefficient selection technology, thereby improving the noise resistance, diagnosis accuracy, and diagnosis stability of the fault diagnosis technology.

[0045] 2. Considering the important role of the wavelet base in wavelet transform processing signals, expert experience is needed to judge the appropriate wavelet base in practical applications. By introducing the importance of the wavelet base and the linear weighted synthesis method, the wavelet transform can use the appropriate wavelet base for signal decomposition through calculation during the network training process.

[0046] 3. Considering that the wavelet scale corresponds to the fault feature extraction frequency band, expert experience is needed for selection. By learning the scale parameter, the appropriate wavelet scale is selected in a data-driven manner, and the corresponding spectrum kurtosis is used to constrain the learning process. After learning a series of appropriate scales, the deep learning network selects the wavelet coefficients corresponding to the scale again based on energy.

[0047] 4. Considering that the design of the zeroing denoising function requires expert experience, while taking into account the intra-scale and inter-scale relationships, a deep learning network is used as an empirical function to determine the zeroing denoising operation from the data. Attached Figure Description

[0048] Figure 1 This is a flowchart of a fault diagnosis method for aero-engine bevel gears based on wavelet denoising networks, provided in one embodiment of this disclosure;

[0049] Figure 2 This is a schematic diagram of a bevel gear vibration signal sample;

[0050] Figure 3 This is a schematic diagram of the network generating a new wavelet basis by weighting the candidate wavelet basis based on the weights;

[0051] Figure 4 yes Figure 2 A schematic diagram of wavelet coefficients generated by the wavelet convolution module of the signal;

[0052] Figure 5 yes Figure 4 A schematic diagram of wavelet coefficients after the signal is denoised by the denoising module;

[0053] Figure 6 yes Figure 5 A schematic diagram of the wavelet coefficients of the signal after weighting by the weighting module;

[0054] Figure 7 It is a fault probability histogram derived by the trained recognition module based on the 64 noise-reduced and weighted wavelet coefficients input.

[0055] Figure 8 This is a diagram comparing the diagnostic accuracy of various models under different signal-to-noise ratios;

[0056] Figure 9 This is a schematic diagram illustrating the stability of diagnostic accuracy for different models under varying operating conditions. Detailed Implementation

[0057] The following will refer to the appendix. Figures 1 to 9 Specific embodiments of this disclosure are described in detail. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0058] It should be noted that some terms are used in the description and claims to refer to certain components. Those skilled in the art can understand that the same component can be referred to by different terms. The description and claims of the present disclosure do not distinguish components based on differences in terminology, but rather on differences in function. As mentioned throughout the description and claims, "including" or "including" is an open term, and should be interpreted as "including but not limited to". The subsequent description is a preferred embodiment of implementing the present disclosure, and is for the purpose of illustrating the general principles of the description, and is not intended to limit the scope of the present disclosure. The scope of protection of the present disclosure is defined by the appended claims.

[0059] In order to facilitate the understanding of the embodiments of the present disclosure, the following will be further explained and described with specific examples combined with the accompanying drawings, and each drawing does not constitute a limitation to the embodiments of the present disclosure.

[0060] In one embodiment, as shown in Figure 1 The present disclosure proposes a wavelet denoising network-based aero-engine bevel gear fault diagnosis method, comprising the following steps:

[0061] S100: Collecting the vibration signal of the bevel gear of the aero-engine to be tested and segmenting;

[0062] S200: Building a fault diagnosis model and training;

[0063] S300: Detecting the segmented vibration signal based on the trained fault diagnosis model, and outputting the probability of different fault categories of the vibration signal, wherein the maximum value of the probability is the actual fault of the bevel gear.

[0064] The above embodiments constitute the complete technical solution of the present disclosure. The present disclosure fully considers the advantages and disadvantages of signal processing technology and deep learning network for fault diagnosis, and combines signal processing with the data-driven ability of deep learning network by introducing wavelet transform, zero denoising, and different scale wavelet coefficient selection technology, thereby improving the noise resistance, diagnosis accuracy, and diagnosis stability of the fault diagnosis technology.

[0065] In another embodiment, in step S200, the fault diagnosis model comprises:

[0066] A wavelet convolution module for one-dimensional wavelet convolution decomposition of the segmented vibration signal to obtain wavelet coefficients;

[0067] A denoising module for zero denoising processing of the wavelet coefficients to obtain denoised wavelet coefficients;

[0068] a weighting module configured to obtain a weight representing importance of the denoised wavelet coefficients, and weight the denoised wavelet coefficients according to the weight to obtain weighted wavelet coefficients;

[0069] a recognition module configured to obtain probabilities of different fault states according to the weighted wavelet coefficients to realize fault recognition of the bevel gear. The recognition module can be composed of a Batch Normalization layer, a ReLU activation function, a max-pooling layer with a step of 2, a one-dimensional convolution layer with a convolution kernel of 5, a Batch Normalization layer, a ReLU activation function, a self-adaptive pooling layer, a full connection layer, a ReLU activation function, a full connection layer with an output dimension of N, and a softmax probability normalization layer connected in sequence. The input of the module is the wavelet coefficients after the weighting processing, and the output is where N is a number of fault state categories, are probability values corresponding to the first fault state and the last fault state of the bevel gear, respectively.

[0070] In another embodiment, the segmented vibration signal is decomposed by one-dimensional wavelet convolution, specifically as follows:

[0071] W = x * ψ'(a)

[0072] where W is a wavelet coefficient obtained by one-dimensional convolution, the dimension of W is C x L, C is a number of wavelet bases, L is a length of the obtained wavelet coefficient, * is convolution, x is a segmented vibration signal sample, ψ'(·) is a convolution kernel composed of linearly weighted fusion of multiple wavelet bases, and the independent variable is a scale parameter a.

[0073] It should be noted that the scale parameter a is determined by network training, and the spectral kurtosis constraint loss L sk for optimization of the optimal scale, and the calculation formula is as follows:

[0074]

[0075]

[0076] where Hilbert(·) is a Hilbert transform, is a wavelet coefficient after denoising, is an envelope coefficient of is a wavelet coefficient at the i-th scale, and L is the length thereof.

[0077] It should be further noted that the C wavelet bases ψ' = {ψ'1, ψ'2,..., ψ' C} used in the wavelet convolution are linearly weighted fusion of O wavelet bases {ψ1, ψ2,..., ψ O} selected as candidates. C} are linearly weighted fusion of O wavelet bases {ψ1, ψ2,..., ψ O ​Linear weighting yields the i-th wavelet basis ψ′. i The calculation formula is:

[0078]

[0079] Where, ψ n For the nth candidate wavelet basis, p i =[p i，1 p i，2 , ..., p i，O ] represents the i-th wavelet basis ψ′ i The corresponding weight vector, where p i，n For the weights corresponding to the nth candidate wavelet basis, C wavelet bases require a total of C weight vectors {p1, p2, ..., p...} C}, p i，n From the candidate wavelet basis set {ψ1, ψ2, ..., ψ O The i-th wavelet basis ψ i With the nth wavelet basis ψ n The similarity is calculated using the following formula:

[0080] p i，n =Softmax(-10) 4 ||ψ i -ψ n ||2),i,n=1,2,...,O

[0081] Where |||2 is the L2 norm calculation, i.e., Euclidean distance; Softmax(·) is the softmax function.

[0082] O weight vectors {p1, p2, ..., p} are generated from O wavelet bases. O}, and select C weight vectors {p1, p2, ..., p} based on their importance evaluation. C Its importance is assessed as follows, using the KL divergence D. KL Measuring the nth vector p n With the weight vector set {p1, p2, ..., p O The difference between all weight vectors in} is taken as the average of the differences, which is used as a single weight vector p. n Importance Indicator I n Then the nth weight p n Corresponding importance index I n The calculation formula is as follows:

[0083]

[0084] Among them, D KL (·||·) is the formula for calculating the KL divergence of two vectors, p mLet {p1, p2, ..., p} O The m-th weight in}, p n，l For p n =[p n，1 p n，2 , ..., p n，O The l-th weight value in the matrix, p m，l For p m The m-th value in the weight vector.

[0085] According to importance, I = {I1, I2, ..., I...} O Sort the data in descending order and select the first C corresponding weight vectors to form a weight set {p1, p2, ..., p...} C}, generating the corresponding wavelet basis set ψ′.

[0086] In another embodiment, the wavelet coefficients are denoised by setting them to zero using the following steps:

[0087] S2011: Local convolution is performed using a one-dimensional convolution conv(·) to obtain the intra-scale relationship of wavelet coefficients and output the probability [ι0, ι1] of whether to perform zeroing, i.e., [ι0, ι1] = conv(W), where W is the wavelet coefficient, conv() is a one-dimensional convolutional layer, ι0 is the probability of zeroing the wavelet coefficient, and ι1 is the probability of keeping the coefficient unchanged. That is, when ι0 > ι1, the corresponding wavelet coefficient is zeroed; otherwise, the original value is retained. The dimensions of ι0 and ι1 are 1×L, and the dimensions of the wavelet coefficient W are C×L. ι0 and ι1 correspond one-to-one with the coefficient W in the length dimension L of the wavelet coefficient to give the judgment of whether to zero it. The same probability ι0 and ι1 are used in the full-scale dimension C.

[0088] S2012: The inter-scale relationship of wavelet coefficients is obtained through a fully connected network fc(·), and the probability [γ0, γ1] of whether to perform zeroing is output, i.e. Among them, W1, W C γ0 represents the wavelet coefficients corresponding to the first scale and the wavelet coefficients corresponding to the Cth scale, respectively. γ0 is the probability of setting the wavelet coefficients to zero, and γ1 is the judgment for keeping the coefficients unchanged. That is, when γ0 > γ1, the wavelet coefficients are set to zero, otherwise the original values ​​are retained.

[0089] S2013: Combine the probabilities of zeroing out the intra-scale and inter-scale relationships in steps S2011 and S2022, and calculate the probability of zeroing out φ0 and the probability of retaining the original value φ1 for the wavelet coefficients, i.e., [φ0, φ1] = Softmax(ι0 + γ0, ι1 + γ1), where the dimensions of φ0 and φ1 are both 1×L, and they have a one-to-one correspondence with the wavelet coefficients in the length dimension.

[0090] S2014: In the denoising network training process, the denoising constraint loss is constructed for the convolutional layer and the fully connected layer to achieve the specified denoising rate, so as to control the denoising ability of the network. The wavelet coefficients are zeroed according to the probability [φ0, φ1] to obtain the denoised wavelet coefficients The zeroing denoising operation is specifically as follows:

[0091]

[0092] wherein, is Hadamard product, H is a sampling result, and H is expressed as:

[0093] H = Gumbel-Softmax (φ0, φ1), H ∈ {0, 1}

[0094] wherein, Gumbel-Softmax (·, ·) is a sampling strategy in Gumbel distribution according to the probability, which samples in the Gumbel distribution according to the probability [φ0, φ1] and outputs the zeroing processing judgment H. When H = 0, the zeroing processing is performed, and when H = 1, the wavelet coefficient is retained. In the denoising network test set stage, when φ0> φ1, H = 0 is output, and when φ0< φ1, H = 1 is output.

[0095] In this step, the denoising constraint loss L DR is expressed as:

[0096]

[0097] wherein, r actual is the denoising rate of the current network, which is calculated by the current zeroing processing judgment H, H ∈ {0, 1}, the dimension of L is L, and L is also the length of the wavelet coefficient at a single scale. r target is the expected denoising rate of the network, which is set to be larger when the signal noise is large, and is set to be smaller when the signal noise is small. The expected denoising rate needs to be set in advance.

[0098] In another embodiment, the weighted wavelet coefficients are obtained by the following steps:

[0099] S2021: Calculate the wavelet coefficient energy E = {E1, E2,..., E C};

[0100] In this step, the wavelet coefficient energy is calculated by the following formula:

[0101]

[0102] wherein, is the zeroed wavelet coefficient of the first scale, E1 represents the wavelet coefficient The energy is calculated sequentially; the subscripts 1, 2, and C indicate the scale number.

[0103] S2022: Based on the energy of wavelet coefficients at each scale, the importance index ω of wavelet coefficients at each scale is obtained through one-dimensional convolution;

[0104] In this step, the importance index ω is represented as:

[0105] ω = sigmoid(conv(E))

[0106] Among them, ω=[ω1,ω2,...,ω C ],ω1,ω C These are the wavelet coefficient importance indices for the 1st and Cth scales, respectively. sigmoid(·) is the sigmoid activation function. The input dimension of the one-dimensional convolution conv(·) is 1×C, and the output dimension is also 1×C.

[0107] It should be noted that the kernel of the one-dimensional convolution used to calculate the importance index ω should be small, with a kernel size of 3×1 and a stride of 1.

[0108] S2023: Based on the importance index, the wavelet coefficients at each scale after being zeroed are linearly weighted to obtain the weighted wavelet coefficients.

[0109] In this step, the weighted wavelet coefficients Represented as:

[0110]

[0111] Wherein, ω is an indicator of the importance of wavelet coefficients; These are the wavelet coefficients after being set to zero.

[0112] In another embodiment, step S200, the training process of the fault diagnosis model includes the following steps:

[0113] S201: Segment the vibration signal of the known fault, divide the segmented vibration signal into training set and test set, and label the corresponding fault state label.

[0114] In this step, the collected vibration signals of the aero-engine bevel gears are segmented into non-overlapping samples according to a fixed sampling time, i.e., ... Where N train With N test These represent the number of samples in the training and test sets, typically set to a 4:1 ratio. Each fault state label corresponds one-to-one with a sample. For a single sample x and its corresponding label y, y needs to be converted to one-hot encoding format y = [y1, y2, ..., y3] during network training. N ], where N is the number of fault state categories. If x belongs to the k-th fault state, then:

[0115]

[0116] In addition, it should be noted that the fault model has a requirement for the length of the input signal, which is a fixed value. Generally, the input length is set to 1024 points, so the collected signal needs to be segmented before it can be input into the model for training.

[0117] S202: Train the fault diagnosis model using the training set. The model training is complete when the network training loss is minimized; otherwise, adjust the network parameters and retrain until the network training loss is minimized.

[0118] S203: Test the trained model using the test set. When the test accuracy reaches 90% or higher, the test is complete, and the trained fault diagnosis model is obtained.

[0119] In another embodiment, in step S202, the network training loss is denoted as L and expressed as:

[0120]

[0121] Among them, L cls The fault classification and diagnosis loss is calculated based on cross-entropy, where N is the number of bevel gear fault categories, and y... i and Let L be the label value corresponding to the i-th fault type and the fault probability output by the network, respectively. DR With L SK These represent the loss for controlling the noise reduction rate and the loss for controlling the spectral kurtosis optimization, respectively. α and β are hyperparameters that can be optimized via grid search, and y = [y1, y2, ..., y]. N [] represents the actual fault state label corresponding to data x, which uses one-hot encoding. If x belongs to the k-th fault state, then:

[0122]

[0123] In another embodiment, this disclosure also provides a fault diagnosis device for aero-engine bevel gears based on a wavelet denoising network, comprising:

[0124] The acquisition module is used to acquire and segment the vibration signal of the bevel gear of the aircraft engine under test.

[0125] The training module is used to train the constructed fault diagnosis model;

[0126] a detection module configured to detect the segmented vibration signal based on the trained fault diagnosis model to output probabilities of different fault categories of the vibration signal.

[0127] In another embodiment, the present disclosure also provides a computer medium comprising:

[0128] a memory configured to store a plurality of computer instructions;

[0129] a processor configured to execute the computer instructions to implement the method according to any one of the preceding embodiments.

[0130] The technical solutions of the present disclosure will be further described below. Figures 2 to 9 The technical solutions of the present disclosure will be further described below.

[0131] Four different fault states of bevel gear experiments were set on a certain type of aero-engine bevel gear fault test test bench, each group was run at five different speeds, and the data set is shown in Table 1.

[0132] Table 1

[0133]

[0134] Figure 2 is the bevel gear vibration signal sample of the input network, the fault is small end collapse, and it is collected at a speed of 1500r / min, and its length is 1024.

[0135] Figure 3 is a schematic diagram of the network generating a new wavelet basis using a candidate wavelet basis based on weights, the left side of the figure is 44 different scales of Laplace wavelets, 42 different scales of Morlet wavelets, and 42 different scales of Mexhat wavelets selected in the model, which together form a library of 128 candidate wavelet bases. The model proposed in the present disclosure calculates the importance weight p according to the difference between the wavelet bases, and then generates 64 new wavelet bases by weighted summation for wavelet convolution and obtains 64 wavelet coefficients. The right side of the figure is a new wavelet generated according to the first candidate wavelet basis weight p 1，1 to the last candidate wavelet basis weight P 1，128 is a schematic diagram of generating the first new wavelet basis ψ'1. The remaining network training setting parameters are shown in Table 2.

[0136] Table 2

[0137] Training epoch 30 Optimizer Adam Noise reduction rate 0.2 Network input sample length 1024 Learning rate 0.0001 BatchSize 64 α 0.05 β 0.005

[0138] Figure 4 is one of the 64 wavelet coefficients generated by the trained wavelet convolution module from the signal shown in Figure 2 Figure 3 the right side.​Figure 2 The signal is convolved to generate 64 wavelet coefficients, with a total dimension of 64x995, where the length of the wavelet coefficients is 995. Figure 4 The input bevel gear signal shown is significantly different, with stronger periodic impact and more obvious characteristics. Figure 2 The input bevel gear signal shown is significantly different, with stronger periodic impact and more obvious characteristics.

[0139] Figure 5 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients. Figure 4 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients. Figure 4 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients. Figure 4 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients. Figure 5 The denoised signal has less noise, the coefficients are more sparse, the impact is more obvious (such as the reduction in amplitude between the two impacts at points 420-500, highlighting the impact signal), and the overall amplitude of the signal is increased, thereby improving the feature extraction ability of the diagnostic model and the robustness to noise.

[0140] Figure 6 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients. Figure 5 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients. Figure 5 The signal is denoised by the trained denoising module, and the denoising decision is determined by the signal itself and the remaining 63 wavelet coefficients.

[0141] Figure 7 The trained recognition module is based on the input of 64 denoised and weighted wavelet coefficients, and the fault probability bar chart is obtained. The output result is [0.02%, 0%, 0%, 99.98%], which belongs to the 4th type of fault shown in Table 1, consistent with the fault state of the input signal, and the diagnosis is correct.

[0142] Figure 8 The diagnostic accuracy of various models under different signal-to-noise ratios is shown in the diagram. DFAWNet is the diagnostic model based on wavelet denoising filtering proposed in the present disclosure. The comparison model is a CNN with a similar structure and number of layers as the present model, a wide kernel convolutional network WCNN, a deep residual shrinkage denoising network DSN, and a WN with a Laplace wavelet convolution kernel as the convolution layer, but with discrete and learnable scale and translation parameters. Since the test bench has a good collection environment, the noise is significantly insufficient compared to the actual operating conditions, so different signal-to-noise ratios (-2dB, 0dB, 2dB) of Gaussian noise are added to the original signal to simulate noisy signals and test the robustness of the method to noise. Figure 8It can be seen that the DFAWNet has the best performance under different signal-to-noise ratios, and the accuracy rate decreases less with the decrease of the signal-to-noise ratio. Under the condition of-2dB, the diagnostic accuracy is 26.5% higher than that of the ordinary CNN, and the WN network using ordinary wavelet function as the convolution kernel performs poorly under different noise, fully demonstrating that the method is robust to noise.

[0143] Figure 9 is a diagram showing the diagnostic accuracy stability of different models under variable operating condition test conditions. 1500rpm is the original operating condition, and other rotating speeds are set as variable operating condition test conditions. It can be seen from the figure that the method shown in the present disclosure not only has the highest diagnostic accuracy under variable operating conditions, but also has the best diagnostic stability. In addition, the diagnostic accuracy of all methods under stable operating conditions and variable operating conditions is compared, as shown in Table 3, the method provided by the present disclosure has the highest diagnostic accuracy under stable operating conditions, and the diagnostic accuracy decreases the least from stable operating conditions to variable operating conditions.

[0144] Table 3

[0145] Model Stable working condition Variable working condition Performance degradation CNN 92.11 65.69 26.43 WCNN 94.74 65.97 28.77 DSN 94.69 64.06 30.63 WN 93.37 65.72 27.62 DFAWNet 98.08 83.68 14.40

[0146] Although the present application has been described in detail in conjunction with the accompanying drawings and examples, the present application is not limited to the above specific embodiments and application fields, and various forms of substitution and modification of the present application are protected by the present application without departing from the principles and spirits of the present application.

Claims

1. A fault diagnosis method for bevel gears in aero-engines based on wavelet denoising networks, comprising the following steps: S100: Collects vibration signals of the bevel gear of the aircraft engine under test and segments them; S200: Construct and train a fault diagnosis model; S300: Based on a trained fault diagnosis model, the segmented vibration signal is detected, and the probability of different fault categories of the vibration signal is output. The maximum probability is the fault category of the bevel gear; In step S200, the fault diagnosis model includes: The wavelet convolution module is used to perform one-dimensional wavelet convolution decomposition on the segmented vibration signal to obtain wavelet coefficients. The denoising module is used to perform zero-setting denoising on the wavelet coefficients to obtain the denoised wavelet coefficients. The weighting module is used to obtain the weights that can characterize the importance of the denoised wavelet coefficients, and to weight the denoised wavelet coefficients according to the weights to obtain the weighted wavelet coefficients. The identification module is used to obtain the probability of different fault states based on the weighted wavelet coefficients in order to identify the faults of bevel gears. The wavelet coefficients are denoised by setting them to zero using the following steps: S2011: Through one-dimensional convolution Obtain the intra-scale relationship of wavelet coefficients and output the judgment on whether to perform zeroing. S2012: Through a fully connected network Obtain the inter-scale relationship of wavelet coefficients and output the judgment on whether to perform zeroing. S2013: Combine the probabilities of zeroing within and between scales from steps S2011 and S2022, and calculate the probability of zeroing corresponding to the wavelet coefficients. And the probability of retaining the original value without setting it to zero. ; S2014: Based on probability The wavelet coefficients are denoised by setting them to zero.

2. The method according to claim 1, wherein, The segmented vibration signal is solved by one-dimensional wavelet convolution, and the one-dimensional wavelet convolution is expressed as follows: ； in, These are the wavelet coefficients obtained through one-dimensional convolution. Dimensions , The number of wavelet bases, To obtain the wavelet coefficient length, For convolution, These are the vibration signal samples after segmentation. The convolution kernel is composed of multiple wavelet baselines weighted and fused, with the scale parameter as its independent variable. .

3. The method according to claim 1, wherein, The weighted wavelet coefficients are obtained through the following steps: S2021: Calculate the energy of wavelet coefficients at each scale; S2022: Based on the energy of wavelet coefficients at each scale, an importance index of wavelet coefficients at each scale is obtained through one-dimensional convolution; S2023: Based on the importance index, the wavelet coefficients of each scale after being zeroed are linearly weighted to obtain the weighted wavelet coefficients.

4. The method according to claim 1, wherein, In step S200, the training process of the fault diagnosis model includes the following steps: S201: Segment the vibration signal of the known fault, divide the segmented vibration signal into training set and test set, and label the corresponding fault state label. S202: Train the fault diagnosis model using the training set. The model training is complete when the network training loss is minimized; otherwise, adjust the network parameters and retrain until the network training loss is minimized. S203: Test the trained model using the test set. When the test accuracy reaches 90% or higher, the test is complete, and the trained fault diagnosis model is obtained.

5. The method according to claim 4, wherein, In step S202, the network training loss is denoted as L and expressed as: ； in, The loss for fault classification and diagnosis is calculated based on cross-entropy. For the number of bevel gear fault categories, and The first The label values ​​corresponding to each fault type and the fault probability output by the network. and These represent the loss for controlling the noise reduction rate and the loss for controlling the spectral kurtosis optimization, respectively. and These are hyperparameters that can be optimized through grid search. For data The corresponding actual fault status label.

6. A fault diagnosis device for aero-engine bevel gears based on wavelet denoising networks, comprising: The acquisition module is used to acquire and segment the vibration signal of the bevel gear of the aircraft engine under test. The training module is used to train the constructed fault diagnosis model; The detection module is used to detect the segmented vibration signal based on the trained fault diagnosis model, so as to output the probability of different fault categories of the vibration signal. The fault diagnosis model in the training module includes: The wavelet convolution module is used to perform one-dimensional wavelet convolution decomposition on the segmented vibration signal to obtain wavelet coefficients. The denoising module is used to perform zero-setting denoising on the wavelet coefficients to obtain the denoised wavelet coefficients. The weighting module is used to obtain the weights that can characterize the importance of the denoised wavelet coefficients, and to weight the denoised wavelet coefficients according to the weights to obtain the weighted wavelet coefficients. The identification module is used to obtain the probability of different fault states based on the weighted wavelet coefficients in order to identify the faults of bevel gears. The denoising module performs zeroing denoising on the wavelet coefficients through the following steps: S2011: Through one-dimensional convolution Obtain the intra-scale relationship of wavelet coefficients and output the judgment on whether to perform zeroing. S2012: Through a fully connected network Obtain the inter-scale relationship of wavelet coefficients and output the judgment on whether to perform zeroing. S2013: Combine the probabilities of zeroing within and between scales from steps S2011 and S2022, and calculate the probability of zeroing corresponding to the wavelet coefficients. And the probability of retaining the original value without setting it to zero. ; S2014: Based on probability The wavelet coefficients are denoised by setting them to zero.

7. A computer medium, comprising: Memory is used to store multiple computer instructions; A processor for instructing computer instructions to implement the method described in any one of claims 1-5.

Images