Fault diagnosis method and diagnosis system based on deep residual variational autoencoder

Abstract

The present invention discloses a fault diagnosis method and a diagnosis system based on a deep residual variational autoencoder. The diagnosis method comprises the following steps: S10: obtaining a vibration acceleration signal X of a mechanical device under different faults; S20: decomposing and reconstructing the collected vibration acceleration signal X through a wavelet packet, obtaining a wavelet packet reconstructed signal as a feature sample set, and randomly dividing the feature sample set into a training set and a test set according to a certain ratio; S30: constructing a deep residual variational autoencoder network model by stacking and connecting a residual autoencoder module and a variational autoencoder module; S40: inputting the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability; S50: inputting the test set into the trained deep network model to automatically identify the type of mechanical fault.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to a fault diagnosis method and a diagnosis system based on a deep residual variational autoencoder. Background Art

[0002] With the rapid development of modern mechanical equipment, rotating mechanical components (such as bearings, gears, and rotors) are widely used in the industrial production of complex mechanical systems such as wind turbines, steam turbines, and aircraft engines. Since rotating mechanical components are in high-speed rotation for a long time, their key components are prone to aging and damage. Once a rotating mechanical device fails, it will inevitably have a great impact on the normal production of the factory and even cause casualties. Therefore, it is very valuable to carry out research on fault diagnosis of rotating mechanical equipment.

[0003] In rotating machinery fault diagnosis, since modern mechanical equipment is in complex working conditions and there is a lot of interference noise in the working environment, if the noisy original vibration signal is directly used for fault diagnosis, it is easy to cause fault diagnosis deviation or inaccuracy. In addition, the traditional rotating machinery fault diagnosis model has problems such as long training time, weak anti-interference ability and insufficient noise reduction ability. Summary of the invention

[0004] In response to the problems and needs raised above, this solution proposes a fault diagnosis method and diagnosis system based on deep residual variational autoencoder. Due to the adoption of the following technical features, it can achieve the above technical objectives and bring about many other technical effects.

[0005] An object of the present invention is to propose a fault diagnosis method based on a deep residual variational autoencoder, comprising the following steps:

[0006] S10: Obtain the vibration acceleration signal X of the mechanical equipment under different faults;

[0007] S20: decomposing and reconstructing the collected vibration acceleration signal X through wavelet packets, obtaining a wavelet packet reconstructed signal as a feature sample set, and randomly dividing the feature sample set into a training set and a test set according to a certain ratio;

[0008] S30: A deep residual variational autoencoder network model is constructed by stacking the residual autoencoder module and the variational autoencoder module front and back.

[0009] S40: inputting the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability;

[0010] S50: Input the test set into the trained deep network model to automatically identify the type of mechanical failure.

[0011] In one example of the present invention, step S20 includes: performing a three-layer wavelet packet decomposition on the collected acceleration signal X to obtain a series of wavelet packet node signals, further reconstructing each wavelet packet node signal, and using the reconstructed wavelet packet node signal as a data sample set, and finally randomly dividing the data sample set into a training set and a test set according to a certain ratio.

[0012] In an example of the present invention, in the wavelet packet decomposition process, the wavelet function used is the db1 wavelet, and its expression is:

[0013]

[0014] Where t is the time variable, j = 1, 2, ..., J, J is the maximum number of decomposition levels, i = 1, 2, ..., 2 j , p i j (t) is the i-th wavelet packet node signal obtained after the j-th layer wavelet packet decomposition, and h and g are the orthogonal conjugate filter coefficients.

[0015] In one example of the present invention, during the wavelet packet node signal reconstruction process, the reconstruction method used is expressed as follows:

[0016]

[0017] In the formula, t is the time variable, s i j (t) is the reconstructed signal of the i-th wavelet packet node obtained after the j-th wavelet packet decomposition, h * and g * are the dual operators of h and g respectively.

[0018] In one example of the present invention, the deep residual variational autoencoder network model constructed in step S30 includes two parts: a residual autoencoder module and a variational autoencoder module.

[0019] The residual autoencoder module is composed of a convolution layer and two residual submodules connected in series; wherein a residual submodule is composed of a batch normalization layer, two convolution layers, and an activation function layer, and a batch normalization layer, a convolution layer, and an activation function layer are connected in series in sequence, and another convolution layer is connected in series with the activation function layer;

[0020] The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series; wherein the coding module is composed of an input layer and a hidden layer connected in series, and the decoding module is composed of a hidden layer and an output layer connected in series.

[0021] In one example of the present invention, the deep residual variational autoencoder network model constructed in step S30 includes:

[0022] S301: Performing a signal denoising and a feature extraction using a residual self-encoding module;

[0023] S302: Use the variational autoencoder module to perform secondary signal denoising and secondary feature extraction.

[0024] In one example of the present invention, in step S30, the loss function of the residual self-encoding module is the mean absolute error, and its specific expression is as follows:

[0025]

[0026] Where N is the number of data points of the input signal, y i is the i-th input of the residual self-encoding module, is the i-th output of the residual self-encoding module;

[0027] The loss function of the variational autoencoder module is the sum of the cross entropy and the KL divergence, that is, J VAE =J 1 +J 2 , where the specific expressions of cross entropy and KL divergence are as follows:

[0028] Cross Entropy:

[0029] Where N is the number of data points of the input signal, y i is the i-th input of the variational autoencoder module, is the i-th output of the variational autoencoder module, w i is the weight parameter of the i-th sample, λ is the regularization coefficient, is the regularization term;

[0030] KL divergence:

[0031] Where N is the number of data points of the input signal, μ and σ are the mean and variance respectively.

[0032] Another object of the present invention is to propose a fault diagnosis system based on a deep residual variational autoencoder, comprising:

[0033] A signal acquisition module, configured to acquire a vibration acceleration signal X of the mechanical equipment under different faults;

[0034] A signal division module is configured to decompose and reconstruct the collected vibration acceleration signal X through wavelet packets, obtain the wavelet packet reconstructed signal as a feature sample set, and randomly divide the feature sample set into a training set and a test set according to a certain ratio;

[0035] A model building module is configured to construct a deep residual variational autoencoder network model by stacking a residual autoencoder module and a variational autoencoder module in a front-to-back manner;

[0036] A training module is configured to input the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability;

[0037] The fault identification module is configured to input the test set into the trained deep network model to automatically identify the type of mechanical fault.

[0038] In one example of the present invention, the signal division module includes:

[0039] The wavelet packet decomposition unit is configured to perform a three-layer wavelet packet decomposition on the collected acceleration signal X to obtain a series of wavelet packet node signals;

[0040] A wavelet reconstruction unit is configured to reconstruct each wavelet packet node signal and use the reconstructed wavelet packet node signal as a data sample set;

[0041] The data partitioning unit is configured to randomly partition the data sample set into a training set and a test set according to a certain ratio.

[0042] In one example of the present invention, the model building module includes: a residual autoencoder module and a variational autoencoder module,

[0043] The residual autoencoder module is composed of a convolution layer and two residual submodules connected in series; wherein a residual submodule is composed of a batch normalization layer, two convolution layers, and an activation function layer, and a batch normalization layer, a convolution layer, and an activation function layer are connected in series in sequence, and another convolution layer is connected in series with the activation function layer;

[0044] The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series; wherein the coding module is composed of an input layer and a hidden layer connected in series, and the decoding module is composed of a hidden layer and an output layer connected in series.

[0045] The best embodiment for carrying out the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0047] Figure 1 is a flow chart of a mechanical fault diagnosis method according to an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of wavelet packet decomposition according to an embodiment of the present invention;

[0049] Figure 3 Schematic diagram of a residual self-encoding module according to an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of a variational autoencoder module according to an embodiment of the present invention;

[0051] Figure 5 A schematic diagram of a deep residual variational autoencoder network according to an embodiment of the present invention;

[0052] Figure 6 1. A time domain waveform diagram corresponding to different fault signals according to an embodiment of the present invention;

[0053] Figure 7 This is a result diagram of a three-layer wavelet packet decomposition of a bearing inner ring fault signal according to an embodiment of the present invention;

[0054] Figure 8 A signal diagram before and after wavelet packet node reconstruction of a bearing inner race fault signal according to an embodiment of the present invention; DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described 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.

[0056] In the actual operation of mechanical equipment, there is often a lot of noise, and useful label data is not easy to obtain. Traditional mechanical fault diagnosis methods are powerless to solve such problems. Therefore, in order to improve the accuracy and effectiveness of mechanical fault detection, the present invention discloses a fault diagnosis method and diagnosis system based on deep residual variational autoencoder.

[0057] According to a first aspect of the present invention, a fault diagnosis method based on a deep residual variational autoencoder is provided. Figures 1 to 5 As shown, the following steps are included:

[0058] S10: Obtaining a vibration acceleration signal X of the mechanical equipment under different faults; for example, installing an acceleration sensor at a key component of the mechanical equipment, and obtaining the vibration acceleration signal X through the acceleration sensor.

[0059] S20: decomposing and reconstructing the collected vibration acceleration signal X through wavelet packets, obtaining a wavelet packet reconstructed signal as a feature sample set, and randomly dividing the feature sample set into a training set and a test set according to a certain ratio;

[0060] Step S20 includes: performing 3-layer wavelet packet decomposition on the collected acceleration signal X to obtain a series of wavelet packet node signals, further reconstructing each wavelet packet node signal, and using the reconstructed wavelet packet node signal as a data sample set, and finally randomly dividing the data sample set into a training set and a test set according to a certain ratio;

[0061] Among them, in the wavelet packet decomposition process, the wavelet function used is db1 wavelet, and its expression is:

[0062]

[0063] Where t is the time variable, j = 1, 2, ..., J, J is the maximum number of decomposition levels, i = 1, 2, ..., 2 j , p i j (t) is the i-th wavelet packet node signal obtained after the j-th layer wavelet packet decomposition, and h and g are the orthogonal conjugate filter coefficients.

[0064] Among them, in the process of reconstructing the wavelet packet node signal, the expression of the reconstruction method used is:

[0065]

[0066] In the formula, t is the time variable, s i j (t) is the reconstructed signal of the i-th wavelet packet node obtained after the j-th wavelet packet decomposition, h * and g * are the dual operators of h and g respectively.

[0067] S30: A deep residual variational autoencoder network model is constructed by stacking the residual autoencoder module and the variational autoencoder module front and back.

[0068] The deep residual variational autoencoder network model constructed in step S30 includes two parts: a residual autoencoder module and a variational autoencoder module.

[0069] The residual autoencoder module is composed of a convolution layer and two residual submodules connected in series; wherein a residual submodule is composed of a batch normalization layer, two convolution layers, and an activation function layer, and a batch normalization layer, a convolution layer, and an activation function layer are connected in series in sequence, and another convolution layer is connected in series with the activation function layer;

[0070] The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series; wherein the coding module is composed of an input layer and a hidden layer connected in series, and the decoding module is composed of a hidden layer and an output layer connected in series; the sampling module is equivalent to a hidden layer.

[0071] The residual autoencoder module is composed of a wide convolution layer and two residual submodules connected in series. The wide convolution layer uses a wide convolution kernel, and each residual submodule is composed of a batch normalization layer, two convolution layers and an activation function layer, wherein a batch normalization layer, a convolution layer and an activation function layer are connected in series in sequence, and the activation function layer is connected in series with another convolution layer;

[0072] The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series. The coding module is composed of an input layer and a hidden layer, and the input layer and the hidden layer are connected in series; the sampling module is equivalent to a hidden layer, and the decoding module is composed of a hidden layer and an output layer, and the hidden layer and the output layer are connected in series;

[0073] Among them, the deep residual variational autoencoder network model constructed in step S30 includes:

[0074] S301: Performing a signal denoising and a feature extraction using a residual self-encoding module;

[0075] S302: Using the variational autoencoder module for secondary signal denoising and secondary feature extraction can enhance the robustness of the network model.

[0076] Among them, in step S30, the loss function of the residual self-encoding module is the mean absolute error, and its specific expression is as follows:

[0077]

[0078] Where N is the number of data points of the input signal, y i is the i-th input of the residual self-encoding module, is the i-th output of the residual self-encoding module;

[0079] The loss function of the variational autoencoder module is the sum of the cross entropy and the KL divergence, that is, J VAE =J 1 +J2 , where the specific expressions of cross entropy and KL divergence are as follows:

[0080] Cross Entropy:

[0081] Where N is the number of data points of the input signal, y i is the i-th input of the variational autoencoder module, is the i-th output of the variational autoencoder module, w i is the weight parameter of the i-th sample, λ is the regularization coefficient, is the regularization term;

[0082] KL divergence:

[0083] Where N is the number of data points of the input signal, μ and σ are the mean and variance respectively.

[0084] It should be noted that, in the loss function of the variational autoencoder module, the smaller the KL divergence, the better the fault identification performance of the obtained network model. Therefore, the present invention improves the proportion of KL divergence in the loss function of the variational autoencoder module from the original -1 / 2 to -1 / 10.

[0085] S40: inputting the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability;

[0086] Specifically, in the deep residual variational autoencoder network, the learning rate of its residual autoencoder module is set to 0.01, the PReLU activation function is selected, and the training is completed by adjusting the learning rate; for the training of the variational autoencoder module, the Adam optimization algorithm is used to iteratively update its network parameters. The deep residual variational autoencoder can increase the fitting ability of the network model by stacking the residual autoencoder module and the variational autoencoder module front and back, thereby improving the diagnostic performance of the entire network;

[0087] The training sample set obtained by S20 is input into the deep residual variational autoencoder for training. In the training process of the deep network model, since the Adam algorithm does not change with the gradient size of the parameters and has the advantages of small calculation amount and fast network training, the Adam algorithm is used in this step to iteratively update the deep network parameters. After multiple iterations of the parameters, the training of the deep residual variational autoencoder network model can be completed.

[0088] S50: Input the test set into the trained deep network model to automatically identify the type of mechanical failure; specifically, the output layer of the deep residual variational autoencoder network is equivalent to a softmax classifier. After inputting the features, the final classification result can be obtained.

[0089] Specifically, this method first performs three-layer wavelet packet decomposition on the collected vibration acceleration signal to extract the effective feature sample set, and divides it into a training set and a test set; then, a deep residual variational autoencoder network model consisting of a stacked residual autoencoder module and a variational autoencoder module is constructed; finally, the constructed deep network model is trained with the training set, and the training set is input into the trained deep network model for intelligent fault identification.

[0090] According to a fault diagnosis method based on a deep residual variational autoencoder of the present invention, in view of the nonlinear and unstable nature of the vibration signal of the rotating machinery, the present invention first uses wavelet packet decomposition for noise reduction preprocessing, and then uses a deep residual variational autoencoder to further filter the noise influence, thereby realizing fault feature extraction and fault diagnosis. The fault diagnosis method has a short training time and strong anti-interference ability, can improve the accuracy and credibility of fault diagnosis, and reduce equipment downtime.

[0091] It should be noted that in the constructed deep network model, the residual autoencoder module is regarded as a feature extractor, which can not only fully mine fault information but also has good noise resistance performance; the variational autoencoder module is regarded as an unsupervised neural network model, which does not require a large amount of labeled data to realize automatic identification of fault types and has good generalization performance.

[0092] Another object of the present invention is to propose a fault diagnosis system based on a deep residual variational autoencoder, comprising:

[0093] A signal acquisition module, configured to acquire a vibration acceleration signal X of the mechanical equipment under different faults;

[0094] A signal division module is configured to decompose and reconstruct the collected vibration acceleration signal X through wavelet packets, obtain the wavelet packet reconstructed signal as a feature sample set, and randomly divide the feature sample set into a training set and a test set according to a certain ratio;

[0095] A model building module is configured to construct a deep residual variational autoencoder network model by stacking a residual autoencoder module and a variational autoencoder module in a front-to-back manner;

[0096] A training module is configured to input the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability;

[0097] The fault identification module is configured to input the test set into the trained deep network model to automatically identify the type of mechanical fault.

[0098] The diagnostic system uses a signal division module to perform three-layer wavelet packet decomposition on the collected vibration acceleration signal to extract an effective feature sample set, and divides it into a training set and a test set. Then, the model building module is used to construct a deep residual variational autoencoder network model consisting of a residual autoencoder module and a variational autoencoder module stacked together. Finally, the training module is used to train the constructed deep network model, and the fault identification module inputs the training set into the trained deep network model for intelligent fault identification.

[0099] According to a fault diagnosis system based on a deep residual variational autoencoder of the present invention, in view of the nonlinear and unstable nature of the vibration signal of the rotating machinery, the present invention first uses wavelet packet decomposition for noise reduction preprocessing, and then uses a deep residual variational autoencoder to further filter the noise influence, thereby realizing fault feature extraction and fault diagnosis. The fault diagnosis system has a short training time and strong anti-interference ability, can improve the accuracy and credibility of fault diagnosis, and reduce equipment downtime.

[0100] In one example of the present invention, the signal division module includes:

[0101] The wavelet packet decomposition unit is configured to perform a three-layer wavelet packet decomposition on the collected acceleration signal X to obtain a series of wavelet packet node signals;

[0102] A wavelet reconstruction unit is configured to reconstruct each wavelet packet node signal and use the reconstructed wavelet packet node signal as a data sample set;

[0103] A data partitioning unit, configured to randomly partition a data sample set into a training set and a test set according to a certain ratio;

[0104] The wavelet function used by the wavelet packet decomposition unit is the db1 wavelet, and its expression is:

[0105]

[0106] Where t is the time variable, j = 1, 2, ..., J, J is the maximum number of decomposition levels, i = 1, 2, ..., 2 j , p i j (t) is the i-th wavelet packet node signal obtained after the j-th layer wavelet packet decomposition, and h and g are the orthogonal conjugate filter coefficients.

[0107] The wavelet reconstruction unit uses the reconstruction method expressed as follows:

[0108]

[0109] In the formula, t is the time variable, s i j (t) is the reconstructed signal of the i-th wavelet packet node obtained after the j-th wavelet packet decomposition, h * and g * are the dual operators of h and g respectively.

[0110] In one example of the present invention, the model building module includes: a residual autoencoder module and a variational autoencoder module,

[0111] The residual autoencoder module is composed of a convolution layer and two residual submodules connected in series; wherein a residual submodule is composed of a batch normalization layer, two convolution layers, and an activation function layer, and a batch normalization layer, a convolution layer, and an activation function layer are connected in series in sequence, and another convolution layer is connected in series with the activation function layer;

[0112] The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series; wherein the coding module is composed of an input layer and a hidden layer connected in series, the decoding module is composed of a hidden layer and an output layer connected in series, and the sampling module is equivalent to a hidden layer.

[0113] Specific implementation case: This example uses the rolling bearing signal provided by the Bearing Data Center of Case Western Reserve University in the United States to verify the proposed method. The bearing fault is injected using electrospark machining technology, and the injected fault diameters are 0.007, 014, 0.021, and 0.028 inches. The vibration acceleration data of the bearing in normal state, inner ring, outer ring, and rolling element faults are collected as the original data set. Specifically, during the data collection process, the sampling frequency of the signal is set to 12kHz, and the motor speed is set to 1730r / min. For each bearing fault state, a total of 300 sets of bearing vibration acceleration data are collected, and the number of sampling points for each set of data is 1024 points. Figure 6 The time domain waveforms corresponding to different fault signals are shown.

[0114] First, the collected vibration data of different bearings are decomposed by three-layer wavelet packets. The decomposition results are as follows: Figure 7 At the same time, the decomposition result of a bearing inner ring fault signal is reconstructed by wavelet packet, and the obtained wavelet packet node signal before and after reconstruction is shown in Figure 8As shown. Subsequently, the wavelet packet reconstruction signals obtained under different fault samples are used as fault feature sample sets, and are divided into training sample sets and test sample sets in a 1:1 ratio. Specifically, there are 1200 groups of training samples and test samples. Finally, the training samples are input into the constructed deep residual variational autoencoder for training, and the test samples are input into the trained deep residual variational autoencoder for bearing fault type identification. Table 1 lists the network structure parameters of the constructed deep residual variational autoencoder. When training the deep residual variational autoencoder network model, the adam parameter optimization algorithm is used, the learning rate is set to 0.01, and the number of iterations is set to 200 times.

[0115] Table 1 Network structure parameters of deep residual variational autoencoder

[0116]

[0117]

[0118] In order to verify the effectiveness of the method proposed in the present invention, the method of the present invention is compared with other classic mechanical fault identification methods. The selected comparison methods are support vector machine, artificial neural network, long short-term memory network, stacked autoencoder, and convolutional neural network. Each method is tested 5 times, and the average recognition accuracy of the 5 experiments is taken as the final fault diagnosis result. The fault diagnosis results obtained by different methods are shown in Table 2;

[0119] Table 2 Fault diagnosis results obtained by different methods

[0120]

[0121] It can be clearly seen from Table 2 that the average recognition accuracy of the method proposed in the present invention is 96.5%, which is higher than several other classic fault diagnosis methods. In addition, due to the wavelet packet decomposition and feature extraction, the model training time can be greatly shortened, and the output of the first part of the residual autoencoder module is input into the variational autoencoder module for training, so that fault identification can be performed faster. It can also be seen from the results in Table 2 that the calculation time of the proposed method is 73.85s, which is less than the calculation time of the stacked autoencoder and the convolutional neural network. Therefore, the method involved in the present invention has good fault identification performance and operation efficiency.

[0122] The exemplary implementation scheme of the fault diagnosis method and diagnosis system based on the deep residual variational autoencoder proposed in the present invention is described in detail above with reference to the preferred embodiments. However, those skilled in the art will appreciate that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the protection scope of the present invention, which is determined by the attached claims.

Claims

1. A fault diagnosis method based on deep residual variational autoencoder, characterized in that: The steps include: S10: Obtain the vibration acceleration signal X of the mechanical equipment under different faults; S20: decomposing and reconstructing the collected vibration acceleration signal X through wavelet packets, obtaining a wavelet packet reconstructed signal as a feature sample set, and randomly dividing the feature sample set into a training set and a test set according to a certain ratio; S30: A deep residual variational autoencoder network model is constructed by stacking the residual autoencoder module and the variational autoencoder module front and back. Among them, the loss function of the residual self-encoding module is the mean absolute error, and its specific expression is as follows: Where N is the number of data points of the input signal, y i is the i-th input of the residual self-encoding module, is the i-th output of the residual self-encoding module; The loss function of the variational autoencoder module is the sum of the cross entropy and the KL divergence, that is, J VAE =J1+J2, where the specific expressions of cross entropy and KL divergence are as follows: Cross Entropy: Where N is the number of data points of the input signal, y i is the i-th input of the variational autoencoder module, is the i-th output of the variational autoencoder module, w i is the weight parameter of the i-th sample, λ is the regularization coefficient, is the regularization term; KL divergence: Where N is the number of data points of the input signal, μ and σ are the mean and variance respectively; S40: inputting the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability; S50: Input the test set into the trained deep network model to automatically identify the type of mechanical failure.

2. The fault diagnosis method based on deep residual variational autoencoder according to claim 1 is characterized in that: Step S20 includes: performing three-layer wavelet packet decomposition on the collected acceleration signal X to obtain a series of wavelet packet node signals, further reconstructing each wavelet packet node signal, and using the reconstructed wavelet packet node signal as a data sample set, and finally randomly dividing the data sample set into a training set and a test set according to a certain ratio.

3. The fault diagnosis method based on deep residual variational autoencoder according to claim 2 is characterized in that: In the process of wavelet packet decomposition, the wavelet function used is db1 wavelet, and its expression is: Where t is the time variable, j = 1, 2, ..., J, J is the maximum number of decomposition levels, i = 1, 2, ..., 2 j , is the i-th wavelet packet node signal obtained after decomposing the j-th layer wavelet packet, and h and g are the orthogonal conjugate filter coefficients.

4. The fault diagnosis method based on deep residual variational autoencoder according to claim 2 is characterized in that: In the process of reconstructing the wavelet packet node signal, the expression of the reconstruction method used is: In the formula, t is the time variable, is the reconstructed signal of the i-th wavelet packet node obtained after the j-th layer wavelet packet decomposition, h * and g * are the dual operators of h and g respectively.

5. The fault diagnosis method based on deep residual variational autoencoder according to claim 1, characterized in that: The deep residual variational autoencoder network model constructed in step S30 includes two parts: a residual autoencoder module and a variational autoencoder module. The residual autoencoder module is composed of a convolution layer and two residual submodules connected in series; wherein a residual submodule is composed of a batch normalization layer, two convolution layers, and an activation function layer, and a batch normalization layer, a convolution layer, and an activation function layer are connected in series in sequence, and another convolution layer is connected in series with the activation function layer; The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series; wherein the coding module is composed of an input layer and a hidden layer connected in series, and the decoding module is composed of a hidden layer and an output layer connected in series.

6. The fault diagnosis method based on deep residual variational autoencoder according to claim 1, characterized in that: The deep residual variational autoencoder network model constructed in step S30 includes: S301: Performing a signal denoising and a feature extraction using a residual self-encoding module; S302: Use the variational autoencoder module to perform secondary signal denoising and secondary feature extraction.

7. A fault diagnosis system based on deep residual variational autoencoder, characterized in that: include: A signal acquisition module, configured to acquire a vibration acceleration signal X of the mechanical equipment under different faults; A signal division module is configured to decompose and reconstruct the collected vibration acceleration signal X through wavelet packets, obtain the wavelet packet reconstructed signal as a feature sample set, and randomly divide the feature sample set into a training set and a test set according to a certain ratio; A model building module is configured to construct a deep residual variational autoencoder network model by stacking a residual autoencoder module and a variational autoencoder module in a front-to-back manner; Among them, the loss function of the residual self-encoding module is the mean absolute error, and its specific expression is as follows: Where N is the number of data points of the input signal, y i is the i-th input of the residual self-encoding module, is the i-th output of the residual self-encoding module; The loss function of the variational autoencoder module is the sum of the cross entropy and the KL divergence, that is, J VAE =J1+J2, where the specific expressions of cross entropy and KL divergence are as follows: Cross Entropy: Where N is the number of data points of the input signal, y i is the i-th input of the variational autoencoder module, is the i-th output of the variational autoencoder module, w i is the weight parameter of the i-th sample, λ is the regularization coefficient, is the regularization term; KL divergence: Where N is the number of data points of the input signal, μ and σ are the mean and variance respectively; A training module is configured to input the training set into the deep residual variational autoencoder network model for sequential training to obtain a deep network model with fault discrimination capability; The fault identification module is configured to input the test set into the trained deep network model to automatically identify the type of mechanical fault.

8. The fault diagnosis system based on deep residual variational autoencoder according to claim 7, characterized in that: The signal division module comprises: The wavelet packet decomposition unit is configured to perform a three-layer wavelet packet decomposition on the collected acceleration signal X to obtain a series of wavelet packet node signals; A wavelet reconstruction unit is configured to reconstruct each wavelet packet node signal and use the reconstructed wavelet packet node signal as a data sample set; The data partitioning unit is configured to randomly partition the data sample set into a training set and a test set according to a certain ratio.

9. The fault diagnosis system based on deep residual variational autoencoder according to claim 7, characterized in that: The model building module includes: a residual autoencoder module and a variational autoencoder module, The residual autoencoder module is composed of a convolution layer and two residual submodules connected in series; wherein a residual submodule is composed of a batch normalization layer, two convolution layers, and an activation function layer, and a batch normalization layer, a convolution layer, and an activation function layer are connected in series in sequence, and another convolution layer is connected in series with the activation function layer; The variational autoencoder module is composed of a coding module, a sampling module and a decoding module connected in series; Among them, the encoding module is composed of an input layer and a hidden layer in series, and the decoding module is composed of a hidden layer and an output layer in series.

Images