Nuclear expansion GRDAU time-varying rotating speed finite sample gear fault diagnosis method based on interference suppression

Abstract

The invention discloses a kernel expansion GRDAU time-varying rotating speed finite sample gear fault diagnosis method based on interference suppression, and relates to the technical field of variable gear faults, and the method comprises the steps: collecting the vibration data of a time-varying rotating speed gear to be detected; inputting vibration data of the time-varying rotation speed gear to be tested into a trained fault diagnosis model, wherein the fault diagnosis model is composed of a kernel expansion GRDAU model and a classifier; the features of the vibration data are extracted through the kernel expansion GRDAU model to obtain a fault diagnosis result of the time-varying rotation speed gear to be detected, the fault diagnosis result is classified through the classifier to obtain a fault classification result of the time-varying rotation speed gear to be detected, and fault diagnosis and classification of the time-varying rotation speed gear under strong noise interference are achieved. According to the method, various fault types of the variable-speed gear under noise interference and limited samples can be effectively diagnosed, and the method is high in practicability and worthy of popularization.

Description

Technical Field

[0001] The present invention relates to the technical field of variable-speed gear faults, and particularly to a time-varying speed limited-sample gear fault diagnosis method. Background Art

[0002] In actual engineering, rotating mechanical equipment often operates at a variable speed during its operation, and its monitoring signals are also affected by background noise interference. Therefore, in its fault diagnosis, in addition to considering the influence of speed changes, it is also necessary to pay special attention to and eliminate the influence of relevant interference on the diagnosis, which is of crucial significance for the accurate fault diagnosis of rotating machinery to ensure its safe and stable operation.

[0003] Traditional time-varying speed signal feature extraction methods such as the order tracking algorithm rely too much on the use of a tachometer, and the order tracking algorithm without a tachometer still relies too much on expert experience; time-frequency analysis methods need to rely on artificially designed basis functions and require sufficient prior knowledge and diagnostic expertise. These traditional methods are gradually being replaced by deep learning and transfer learning methods that can perform end-to-end learning. Among them, by improving the deep learning algorithm to enhance the model's ability to automatically extract features from the original signal, the classification accuracy and stability can be further improved. Similarly, the method based on transfer learning can use domain adaptation technology to eliminate the influence of speed changes on feature extraction and fault diagnosis. However, the methods for deep learning and transfer learning mainly focus on the diagnostic accuracy under time-varying speed, while ignoring the influence of noise interference and limited samples on the diagnostic results. Therefore, how to effectively simultaneously process the problem of effective feature extraction under the time-varying speed condition with noise interference and limited samples has become an urgent problem to be solved.

[0004] The invention with the application number 202310315205.8 discloses a bearing fault prediction method and system based on a convolutional denoising autoencoder. The method includes collecting a vibration signal sequence of a rolling bearing and inputting it into a bearing fault diagnosis model, extracting vibration signal features through a convolutional denoising autoencoder including a gated recurrent unit (GRU), and classifying by a classifier to obtain a fault detection result of the rolling bearing; the convolutional denoising autoencoder including a gated recurrent unit (GRU) is an encoder-decoder structure composed of an encoding unit and a decoding unit connected in sequence. The encoding unit is composed of a noise addition layer, a first GRU network, a convolutional layer, and a pooling layer connected in sequence, and the decoding unit is composed of a transposed convolutional layer, an unpooling layer, and a second GRU network connected in sequence. The above invention can adaptively extract the time series features in the bearing time series vibration signal, enhance the feature extraction ability, reduce the influence of noise, and achieve accurate and efficient bearing fault prediction and diagnosis. Although the above invention adopts a convolutional denoising autoencoder including a gated recurrent unit (GRU), there is still a problem of limited feature extraction ability; in actual engineering, the monitoring signals of rotating mechanical equipment are disturbed by complex and diverse background noises, so the above invention has insufficient processing of noise interference; and the above invention is not effective enough in processing small sample data. Summary of the Invention

[0005] The present invention proposes a method for diagnosing faults of gears with time-varying speeds and limited samples based on interference suppression and an extended kernel GRDAU, aiming to solve the technical problem of inaccurate diagnosis of faults of variable-speed gears under limited sample numbers and noise interference.

[0006] In order to achieve the above object, the technical solution of the present invention is realized as follows:

[0007] A method for diagnosing faults of gears with time-varying speeds and limited samples based on interference suppression and an extended kernel GRDAU includes the following steps:

[0008] S1. Collect vibration data of the gear with time-varying speed to be measured;

[0009] S2. Input the vibration data of the gear with time-varying speed to be measured into a trained fault diagnosis model. The fault diagnosis model is composed of an extended kernel GRDAU model and a classifier; extract the features of the vibration data through the extended kernel GRDAU model to obtain a fault diagnosis result of the gear with time-varying speed to be measured, and classify the fault diagnosis result through the classifier to obtain a classification result of the fault of the gear with time-varying speed to be measured, so as to realize the fault diagnosis and classification of the gear with time-varying speed under strong noise interference;

[0010] The extended kernel GRDAU model is composed of an extended kernel convolutional layer and a gated recurrent dropout attention unit (GRDAU) connected in sequence.

[0011] Preferably, the extended kernel convolutional layer is obtained by constructing an extended kernel parameter in a traditional convolution.

[0012] Preferably, the gated recurrent dropout attention unit GRDAU is obtained by adding an attention gate to the gated recurrent unit GRU and adding recurrent dropout RD to the input and hidden state of the gated recurrent unit GRU.

[0013] Preferably, the feature extraction of vibration data by the dilated kernel GRDAU model is specifically as follows: the dilated convolution layer is used to complete the preliminary feature extraction of the input vibration data of the time-varying rotational speed gear to be measured and suppress high-frequency noise, and then the gated recurrent dropout attention unit GRDAU is used to achieve the extraction of potential sensitive features, completing the feature extraction.

[0014] Preferably, it is assumed that the input signal is x ∈ T n , the filter is w ∈ T m and the dilation rate of the dilated kernel parameter Dilation Rate = d, and the calculation process of the dilated convolution layer is:

[0015]

[0016] In the formula, b l is the bias value, z l is the linear activation vector of the l-th layer, σ is the non-linear activation function, x l is the output feature of the l-th layer; conv represents the convolution operation, and same represents zero-padding; Y l refers to the feature vector after convolution processing, x l-1 is the output feature of the l-1 layer, and wl refers to the size of the convolution kernel of the l-th layer; (y l (t), …, y l (n - m + 1)) ∈ T n-m+1 represents the convolution process, represents the convolution process y l (t)'s calculation process, w l (i) represents the kernel of the l-th layer, and x l-1 (t + i - 1) is the signal of the l-1 layer; T represents the time dimension of the signal or filter, n represents the time series length of the input signal, m represents the time series length of the filter, t represents the time index; i represents the index of the filter.

[0017] Preferably, the calculation process of the gated recurrent dropout attention unit GRDAU is as follows:

[0018] x t = RD(input)

[0019] h t-1 = RD(t - 1)

[0020] r t = σ(Ur x t +W r h t-1 +b r )

[0021] z t =σ(U z x t +W z h t-1 +b z )

[0022]

[0023] s(x t ,h t-1 )=V T tanh(W s x t +U s h t-1 )

[0024] α t =ELU(s(x t ,h t-1 ))

[0025]

[0026] Among them, RD represents the cyclic dropout operation on the input data input; x t represents the input data after being processed by cyclic dropout RD; h t-1 represents the hidden state at time t-1, and cyclic dropout RD is performed on the hidden state; tanh refers to the hyperbolic tangent function; r t and z t represent the outputs of the reset gate and the update gate, U r , W r , U z , W z , U h and W h are all weight matrices, b r , b z and b h are all bias matrices, represents the temporary hidden state at instant t, and W represents the dot product; s(x t ,h t-1 ) is the attention scoring function of the hidden state h t-1 ; V is the additive parameter in the attention mechanism, W s and U s are the weight matrices related to the attention scoring function, and are the temporary hidden states at instant t The weight matrix related to the output, α t represents the attention distribution vector; ELU represents the ELU function; h t is the hidden state output at the current time step t.

[0027] Preferably, the classifier is a SoftMax classifier.

[0028] Preferably, the expression for classifying the fault diagnosis result through the classifier to obtain the classification result of the fault of the time-varying rotational speed gear to be measured is:

[0029]

[0030] where C refers to the classification category, f(x) refers to the classification output, exp is the exponential function, N is the total number of classification categories, is the output feature or score of the classifier for the nth category at the lth layer, and p is the probability value that the input data x belongs to the Cth category.

[0031] Preferably, before the step S2, it further includes training the fault diagnosis model, and the implementation method is:

[0032] Collect the vibration data of the time-varying rotational speed gear with different fault types under time-varying working conditions, establish a data set suitable for input to the extended kernel GRDAU model composed of the vibration data of the time-varying rotational speed gear under each fault type under time-varying working conditions, and divide the data set into a training set and a test set;

[0033] Use the training set to train the extended kernel GRDAU model to obtain the trained extended kernel GRDAU model, and use the test set to adjust the parameters of the trained extended kernel GRDAU model to obtain the trained extended kernel GRDAU model;

[0034] Add a SoftMax classifier to the trained extended kernel GRDAU model to obtain the trained fault diagnosis model.

[0035] Preferably, during the process of training the extended kernel GRDAU model, the learning rate changes dynamically, and the calculation formula of the learning rate is:

[0036]

[0037] In the formula, are respectively the maximum and minimum values of the learning rate in the mth cycle;

[0038] When the parameter b ∈ [0, 1], the calculation formula of the parameter b is:

[0039] where △T is the cycle period length, m is the number of cycle periods, and K is the parameter of the size of the data set.

[0040] Advantages of the present invention compared with the prior art:

[0041] The present invention provides an innovative method for the fault diagnosis of time-varying speed gears with limited samples in a noisy environment. The high-precision fault diagnosis and classification of time-varying speed gears under strong noise interference are realized through a fault diagnosis model composed of an extended kernel GRDAU model and a classifier, which are successively connected by an extended kernel convolutional layer and a gated recurrent dropout attention unit (GRDAU).

[0042] The present invention can not only effectively handle the problems of noise interference and difficult fault diagnosis of time-varying speed gears with limited samples, but also accurately extract the fault feature components of time-varying speed gears, effectively diagnose various fault types of variable speed gears under noise interference and limited samples, and has strong practicability and is worthy of promotion. Description of the drawings

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0044] Figure 1 It is a flowchart of the present invention.

[0045] Figure 2 It is a schematic diagram of GRDAU in the present invention.

[0046] Figure 3 It is a graph showing the change of the learning rate under the new learning strategy in the present invention.

[0047] Figure 4 It is a graph of the diagnosis result of the vibration data of the time-varying speed gear obtained in an embodiment of the present invention without noise. Among them, Figure 4 -(a) is the t-SNE visualization diagram after feature extraction by the extended kernel convolutional layer, Figure 4 -(b) is the t-SNE diagram after feature extraction by GRDAU.

[0048] Figure 5 It is a graph of the comparison result between the true label and the predicted label in an embodiment of the present invention.

[0049] Figure 6 It is a graph of the classification result under the noise interference level of 4 dB in an embodiment of the present invention.

[0050] Figure 7 It is a graph of the comparison result between the true label and the predicted label under the noise interference level of 4 dB in an embodiment of the present invention. Detailed implementation mode

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0052] As Figure 1 shown, a method for diagnosing faults of a gear with time-varying rotational speed and limited samples based on interference suppression and kernel expansion GRDAU includes the following steps:

[0053] S1. Collect vibration data of the gear with time-varying rotational speed to be measured.

[0054] S2. Input the vibration data of the gear with time-varying rotational speed to be measured into the trained fault diagnosis model; the fault diagnosis model is composed of a kernel expansion GRDAU model and a classifier; the fault diagnosis result of the gear with time-varying rotational speed to be measured is obtained by extracting the features of the vibration data through the kernel expansion GRDAU model, and the classification result of the fault of the gear with time-varying rotational speed to be measured is obtained by classifying the fault diagnosis result through the classifier, so as to realize the fault diagnosis and classification of the gear with time-varying rotational speed under strong noise interference.

[0055] The kernel expansion GRDAU model is composed of a kernel expansion convolutional layer and a gated recurrent dropout attention unit (GRDAU) connected in sequence.

[0056] Specifically, the preliminary feature extraction of the input vibration data of the gear with time-varying rotational speed to be measured is completed through the kernel expansion convolutional layer, and most of the high-frequency noise is suppressed. Then, the gated recurrent dropout attention unit (GRDAU) is used to realize deeper-level potential sensitive feature extraction, further improving the deep mining of vibration data and the capture of potential typical features.

[0057] Among them, the kernel expansion convolutional layer is obtained by constructing a kernel expansion parameter in the traditional convolution. The traditional convolution is a mathematical operation widely used in the fields of signal processing, image processing, neural networks, etc. The traditional convolution here refers to the convolution with only the convolution kernel and stride operations and no other operations.

[0058] The gated recurrent dropout attention unit (GRDAU) is obtained by adding an attention gate in the gated recurrent unit (GRU) and adding recurrent dropout (RD) to the input and hidden states of the gated recurrent unit (GRU).

[0059] Construct an extended kernel parameter in traditional convolution to suppress high-frequency noise; improve the robustness of the fault diagnosis model in diagnosing under conditions of limited samples, noise interference, and rotational speed changes by constructing a gated recurrent dropout attention unit (GRDAU).

[0060] Specifically, the dilated convolution layer is expressed as:

[0061] Assume that the input signal is \(x\in T\) n , the filter is \(w\in T\) m and the dilation rate of the extended kernel parameter is \(Dilation Rate = d\). The dilation rate parameter defines the number of intervals between values when the convolutional kernel processes data. When the dilation rate parameter \(Dilation Rate\) is 1, the convolutional kernel will work like a normal convolution; when the dilation rate parameter \(Dilation Rate\) is greater than 1, the convolutional kernel will insert dilations between its elements, thereby expanding the receptive field and enabling each convolutional output to contain a larger range of information.

[0062] Then the calculation process of the dilated convolution layer is:

[0063]

[0064] In the formula, \(b\) l is the bias value, \(z\) l is the linear activation vector of the \(l\)-th layer, \(\sigma\) is a non-linear activation function, such as the sigmoid function; \(x\) l is the output feature of the \(l\)-th layer. \(conv\) represents the convolution operation, and \(same\) represents zero-padding. \(Y\) l refers to the feature vector after convolution processing, \(x\) l-1 is the output feature of the \((l - 1)\)-th layer, \(w\) l refers to the size of the convolutional kernel of the \(l\)-th layer; \((y\) l (t),…,y l (n - m + 1))\in T n-m+1 represents the convolution process, represents the convolution process \(y\) l (t)\) of the calculation process, \(w\) l (i)\) represents the kernel of the \(l\)-th layer, \(x\) l-1 (t + i - 1)\) is the signal of the \((l - 1)\)-th layer. \(T\) represents the time dimension of the signal or filter, \(n\) represents the length of the time series of the input signal, \(m\) represents the length of the time series of the filter, \(t\) represents the time index, which is used to describe the values of the signal at different time points; \(i\) represents the index of the filter, which is used to traverse each time point of the filter.

[0065] Specifically, such as Figure 2As shown, by introducing an attention gate and a recurrent dropout strategy into the gated recurrent unit (GRU), the gated recurrent dropout attention unit (GRDAU) obtained can more effectively capture important information in the input data and significantly enhance the memory and generalization capabilities of the dilated-kernel GRDAU model.

[0066] The detailed calculation process of the gated recurrent dropout attention unit (GRDAU) is as follows:

[0067] x t = RD(input)

[0068] h t-1 = RD(t - 1)

[0069] r t = σ(U r x t + W r h t-1 + b r )

[0070] z t = σ(U z x t + W z h t-1 + b z )

[0071]

[0072] s(x t , h t-1 ) = V T tanh(W s x t + U s h t-1 )

[0073] α t = ELU(s(x t , h t-1 ))

[0074]

[0075] Among them, RD represents the recurrent dropout operation on the input data input. x t represents the input data after being processed by recurrent dropout RD; h t-1represents the hidden state at time t-1, and performs recurrent dropout (RD) on the hidden state. The hidden state is a key variable in the recurrent neural network for capturing the temporal dependence of sequential data. It contains information from all previous time steps and is updated at each time step. By performing the recurrent dropout (RD) operation on the hidden state, the regularization effect of the dilated kernel GRDAU model is further enhanced, preventing the hidden state from passing overly strong specific patterns in the time series and causing overfitting. tanh refers to the hyperbolic tangent function. r t and z t represent the outputs of the reset gate and the update gate, U r , W r , U z , W z , U h and W h are all weight matrices, b r , b z and b h are all bias matrices, represents the temporary hidden state at instant t, and W represents the dot product. s(x t , h t-1 ) is the attention scoring function for the hidden state h t-1 . V is the additive parameter in the attention mechanism, and W s and U s are the weight matrices related to the attention scoring function, and are the weight matrices related to the output of the temporary hidden state at instant t, and α t represents the attention distribution vector.

[0076] ELU is the abbreviation of Exponential Linear Unit. It is an activation function used to introduce non-linearity into the neural network and helps the dilated kernel GRDAU model learn complex patterns and features. The ELU function behaves similarly to the linear function in the positive region, while in the negative region, it smoothly transitions to a negative value through exponential operations. This property enables the ELU function to alleviate the "dying neuron" problem, accelerate the training convergence speed to a certain extent, and maintain good representational ability and generalization performance;

[0077] h t is the output of the hidden state at the current time step t, which is calculated by combining the information of the reset gate, the update gate, and the temporary hidden state, etc. The hidden state h t contains the comprehensive information of all input data from the initial time step to the current time step, and is transmitted and updated in the time series through the structure of the recurrent neural network, providing an important feature representation for subsequent fault diagnosis and classification.

[0078] The classifier is a SoftMax classifier, and the expression for obtaining the classification result of the fault through the SoftMax classifier is as follows:

[0079]

[0080] Among them, C refers to the classification category, f(x) refers to the output of the classification, exp is the exponential function, N is the total number of classification categories, and in the context of gear fault diagnosis, it represents the total number of different fault types that the gear may have, such as normal state, different types or degrees of faults, etc. is the output feature or score of the classifier for the nth category at the lth layer (which may be the last layer). In a neural network, it is usually obtained through a series of linear transformations and activation functions, and is used to measure the correlation or matching degree between the input data and the nth category. p represents the probability value, specifically referring to the probability that the classifier predicts that the input belongs to the Cth category given the input data x. The output features of the classifier are converted into a probability distribution through the SoftMax function, so that the sum of the probabilities of all categories is 1, thereby obtaining the prediction probability of each category.

[0081] Furthermore, before the step S2, it also includes training the fault diagnosis model, and the implementation method is as follows:

[0082] Use an acceleration sensor to collect the vibration data of a time-varying speed gear with different fault types under time-varying working conditions, establish a dataset Channel_1 suitable for input to the extended kernel GRDAU model composed of the vibration data of the time-varying speed gear under each fault type under time-varying working conditions, and divide the dataset Channel_1 into a training set and a test set.

[0083] Use the training set to train the extended kernel GRDAU model to obtain the trained extended kernel GRDAU model, and use the test set to adjust the parameters of the trained extended kernel GRDAU model to obtain the trained extended kernel GRDAU model.

[0084] Specifically, the method for obtaining the trained extended kernel GRDAU model is as follows: Input the data in the test set into the trained extended kernel GRDAU model, use the trained extended kernel GRDAU model to perform fault diagnosis on the data in the test set, observe the output results of the trained extended kernel GRDAU model, evaluate the accuracy and performance of the trained extended kernel GRDAU model by comparing the predicted results and the true labels of the trained extended kernel GRDAU model, and adjust the trained extended kernel GRDAU model according to the evaluation results to obtain the trained extended kernel GRDAU model.

[0085] Add a SoftMax classifier to the trained kernel-expanded GRDAU model to obtain a trained fault diagnosis model.

[0086] After obtaining the trained kernel-expanded GRDAU model, the introduction of the classifier is a crucial step in constructing the fault diagnosis model. Although the kernel-expanded GRDAU model can already effectively extract features from vibration data, these features need to be finally mapped to specific fault categories through the classifier. Specifically, after adding the SoftMax classifier to the trained kernel-expanded GRDAU model, they jointly form a complete fault diagnosis model. In practical applications, when new vibration data is input into the fault diagnosis model, the kernel-expanded GRDAU model first extracts features from the vibration data, and the obtained feature vectors are then passed to the SoftMax classifier. The classifier classifies the input feature vectors according to the mapping relationship between the features learned during training and the fault categories, and finally outputs the fault category to which the vibration data belongs. Therefore, the classifier plays a crucial role in the fault diagnosis model. It transforms the results of feature extraction into diagnostically meaningful conclusions, achieving the leap from features to fault categories, thus completing the entire fault diagnosis task.

[0087] During the training process, a new learning strategy is adopted to improve the learning efficiency of the kernel-expanded GRDAU model. The new learning strategy is to fuse a learning rate decay strategy into the global cyclic learning strategy, that is, to design a dynamically changing exponential function:

[0088]

[0089] In the formula, are respectively the maximum and minimum values of the learning rate in the m-th cycle; α t is the learning rate.

[0090] When the parameter b ∈ [0, 1], the calculation formula for the parameter b is:

[0091] where, △T is the cycle period length, m is the number of cycle periods, and K is a parameter related to the size of the data set.

[0092] Through the new learning strategy, a more flexible learning rate adjustment mechanism is provided for the kernel-expanded GRDAU model, enabling the learning rate to be dynamically adjusted as the training process progresses to adapt to the training needs at different stages, as Figure 3 shown.

[0093] The time-varying speed limited sample gear fault diagnosis method based on interference suppression and kernel expansion GRDAU of the present invention will be applied to a specific example below to illustrate the effectiveness of the method of the present invention. During the test, there were vibration data of 5 different fault degrees (the root crack fault degrees included: 0mm, 0.2mm, 0.6mm, 1mm, 1.4mm). The rotational speed changed first increased and then decreased, and the number of training samples was 360 and the number of test samples was 40 in all cases.

[0094] As Figure 4 shown, it is the diagnosis result diagram of the vibration data of the time-varying speed gear obtained by the present invention without noise. Among them Figure 4 -(a) is the t-SNE visualization diagram after feature extraction by the kernel expansion convolutional layer, Figure 4 -(b) is the t-SNE diagram after feature extraction by GRDAU. It can be seen by observation that the features of the first layer show a state of dispersion and overlap, and there is no obvious clustering rule for the feature points of different fault categories, reflecting that the feature extraction effect in the initial stage is weak. However, as the kernel expansion GRDAU model gradually performs feature extraction and learning, the final output layer shows obvious classification results, and each fault category is effectively distinguished, indicating that the fault diagnosis model proposed by the present invention can effectively extract time-varying speed features and achieve excellent fault classification.

[0095] As Figure 5 shown, it shows the comparison result of the true label and the predicted label without interference. It can be clearly seen from the result that the true label and the predicted label completely coincide, which indicates that the present invention has high effectiveness and accuracy.

[0096] As Figure 6 shown, it shows adding Gaussian white noise to the original vibration signal. That is, the t-SNE result diagram when the signal-to-noise ratio is 4dB. It can be seen from the result that even under strong noise interference, the fault diagnosis model proposed by the present invention can still clearly identify various fault types under limited samples of time-varying speed.

[0097] As Figure 7 shown, it shows the comparison result diagram of the true label and the predicted label under the interference degree of 4dB noise. It can be found from the observation result that even under the interference of strong noise, the fault diagnosis method proposed by the present invention can still effectively distinguish the true label and the predicted label, and has extremely strong robustness when dealing with time-varying speed limited sample data under strong noise.

[0098] In summary, the present invention improves the deep learning algorithm, enhances the ability of the fault diagnosis model to automatically extract features from the original signal, and thus improves the classification accuracy and stability.

[0099] The present invention pays more attention to and can more effectively eliminate the impact of relevant interferences on diagnosis, which is of crucial significance for the accurate fault diagnosis of rotating machinery to ensure its safe and stable operation.

[0100] The present invention is specifically optimized for the case of limited samples, enabling effective feature extraction and fault diagnosis even in the case of limited samples. Finally, for the lack of adaptability to time-varying speed conditions, the present invention mainly focuses on improving the diagnostic accuracy, but insufficiently considers the adaptability of feature extraction and fault diagnosis under time-varying speed conditions. The present invention can better handle the problem of effective feature extraction under time-varying speed conditions with noise interference and limited samples.

[0101] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for diagnosing gear faults with limited samples of time-varying rotational speed of an extended core GRDAU based on interference suppression, characterized in that, It includes the following steps: S1. Collect the vibration data of the time-varying speed gear to be measured; S2. Input the vibration data of the time-varying speed gear to be measured into the trained fault diagnosis model, and the fault diagnosis model is composed of an extended kernel GRDAU model and a classifier; extract the features of the vibration data through the extended kernel GRDAU model to obtain the fault diagnosis result of the time-varying speed gear to be measured, and classify the fault diagnosis result through the classifier to obtain the classification result of the fault of the time-varying speed gear to be measured, so as to realize the fault diagnosis and classification of the time-varying speed gear under strong noise interference; The extended kernel GRDAU model is composed of an extended kernel convolutional layer and a gated recurrent dropout attention unit GRDAU connected in sequence.

2. The method for diagnosing gear faults with time-varying rotational speed and limited samples based on interference suppression and extended kernel GRDAU according to claim 1, wherein, The extended kernel convolutional layer is obtained by constructing an extended kernel parameter in the traditional convolution.

3. The method for diagnosing gear faults with time-varying rotational speed and limited samples of the expanded core GRDAU based on interference suppression according to claim 1, wherein, The gated recurrent dropout attention unit GRDAU is obtained by adding an attention gate in the gated recurrent unit GRU and adding recurrent dropout RD to the input and hidden state of the gated recurrent unit GRU.

4. The method for diagnosing gear faults with time-varying rotational speed and limited samples of the extended-core GRDAU based on interference suppression according to claim 2 or 3, characterized in that, The specific process of extracting the features of the vibration data through the extended kernel GRDAU model is as follows: complete the preliminary feature extraction of the input vibration data of the time-varying speed gear to be measured through the extended kernel convolutional layer and suppress the high-frequency noise, and then realize the extraction of potential sensitive features through the gated recurrent dropout attention unit GRDAU to complete the feature extraction.

5. The method for diagnosing gear faults with limited samples of variable rotational speed of the kernel-expanded GRDAU based on interference suppression according to claim 4, wherein, Assume that the input signal is \(x\in T\) n and the filter is \(w\in T\) m and the dilation rate of the dilation kernel parameter is \(Dilation Rate = d\). The calculation process of the dilated convolution layer is as follows: where b l is the bias value, z l is the linear activation vector of the l-th layer, σ is the non-linear activation function, and x l is the output feature of the l-th layer; conv represents the convolution operation, and same represents zero-padding; Y l refers to the feature vector after convolution processing, x l-1 is the output feature of the l-1 layer, wl refers to the size of the convolution kernel of the l-th layer; (y l (t),…,y l (n - m + 1)) ∈ T n-m+1 represents the convolution process, represents the convolution process y l (t)'s calculation process, w l (i) represents the kernel of the l-th layer, x l-1 (t + i - 1) is the signal of the l-1 layer; T represents the time dimension of the signal or filter, n represents the time series length of the input signal, m represents the time series length of the filter, t represents the time index; i represents the index of the filter.

6. The method for diagnosing gear faults with time-varying rotational speed and limited samples of the kernel-expanded GRDAU based on interference suppression according to claim 5, wherein The calculation process of the gated recurrent dropout attention unit GRDAU is as follows: x t = RD(input) h t-1 = RD(t - 1) r t = σ(U r x t + W r h t-1 + b r ) z t = σ(U z x t + W z h t-1 + b z ) s(x t ,h t-1 ) = V T tanh(W s x t +U s h t-1 ) α t = ELU(s(x t ,h t-1 )) Among them, RD represents the cyclic discard operation on the input data input; x t represents the input data after being processed by the cyclic discard RD; h t-1 represents the hidden state at time t-1, and the hidden state is subjected to the cyclic discard RD; tanh refers to the hyperbolic tangent function; r t and z t represent the outputs of the reset gate and the update gate, U r , W r , U z , W z , U h and W h are all weight matrices, b r , bz and bh are all bias matrices, represents the temporary hidden state at instant t, W represents the dot product; s(x t , h t-1 ) is the attention scoring function of the hidden state ht-1, V is the additive parameter in the attention mechanism, W s and U s are the weight matrices related to the attention scoring function, and are the weight matrices related to the output of the temporary hidden state at instant t, α t represents the attention distribution vector; ELU represents the ELU function; h t is the output of the hidden state at the current time step t.

7. The method for diagnosing gear faults with time-varying rotational speed and limited samples of the extended-core GRDAU based on interference suppression according to claim 1 or 6, characterized in that, The classifier is a SoftMax classifier.

8. The method for diagnosing gear faults with limited samples of time-varying rotational speed of the extended core GRDAU based on interference suppression according to claim 7, wherein The expression for classifying the fault diagnosis result through the classifier to obtain the classification result of the fault of the time-varying speed gear to be measured is: Where C refers to the classification category, f(x) refers to the classification output, exp is the exponential function, N is the total number of classification categories, is the output feature or score of the classifier for the nth category at the l-th layer, and p is the probability value that the input data x belongs to the C-th category.

9. The method for diagnosing gear faults with time-varying rotational speed and limited samples based on interference suppression and core-expanded GRDAU according to claim 8, wherein Before step S2, it also includes training the fault diagnosis model, and the implementation method is: Collect the vibration data of the time-varying speed gears of different fault types under time-varying working conditions, establish a data set suitable for input into the extended kernel GRDAU model composed of the vibration data of the time-varying speed gears of each fault type under time-varying working conditions, and divide the data set into a training set and a test set; Use the training set to train the extended kernel GRDAU model to obtain the trained extended kernel GRDAU model, and use the test set to adjust the parameters of the trained extended kernel GRDAU model to obtain the trained extended kernel GRDAU model; Add a SoftMax classifier to the trained extended kernel GRDAU model to obtain the trained fault diagnosis model.

10. The method for diagnosing gear faults with limited samples of time-varying rotational speed of an expanded core GRDAU based on interference suppression according to claim 9, characterized in that, During the process of training the extended kernel GRDAU model, the learning rate changes dynamically, and the calculation formula of the learning rate is: wherein, are respectively the maximum and minimum values of the learning rate in the m-th cycle; When the parameter b ∈ [0, 1], the calculation formula of the parameter b is: Among them, △T is the cycle period length, m is the number of cycle periods, and K is the parameter of the size of the data set.

Images