A bearing fault new class recognition and diagnosis method based on hybrid deep auto-encoding

Abstract

This invention belongs to the technical field of intelligent fault diagnosis, and more specifically, relates to a method for identifying and diagnosing novel bearing fault types based on a hybrid deep autoencoder. The method includes: S1, collecting bearing vibration signal data under different fault conditions as samples; S2, preprocessing the samples collected in S1 and dividing them into training and testing sets; S3, constructing a hybrid deep autoencoder network, including a first encoder, a decoder, a second encoder, and a Softmax classification layer; S4, inputting the training set into the hybrid deep autoencoder network for pre-training to obtain a new fault detection threshold, thus achieving fault classification; and S5, inputting the testing set into the trained hybrid deep autoencoder network for diagnosis. This invention solves the problems in existing technologies where deep learning models require a large amount of labeled data for training, and where a single deep learning model easily leads to misclassification of faults and overfitting when extracting similar features.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent fault diagnosis, and more specifically, relates to a novel bearing fault identification and diagnosis method based on hybrid deep autoencoding. Background Technology

[0002] As a critical component in mechanical equipment, bearings account for approximately 30% of failures in rotating machinery. Their proper functioning directly impacts the overall performance of the equipment. Research into real-time bearing condition monitoring and fault diagnosis technologies is crucial not only for ensuring safe and reliable equipment operation but also for extending equipment lifespan and saving economic costs.

[0003] Existing novel class identification methods are often based on autoencoder network models, which can encode inputs into a more compact representation while retaining important discriminative features, and can also decode specific codes back to their original form. Zhe Yang et al. designed a deep neural network based on a sparse autoencoder, using an encoder network to extract features from a fault dataset. Based on this, a decoder network and a classification module are used for unsupervised reconstruction and supervised classification, respectively. Novel class identification is achieved by comparing the reconstruction error with a predefined threshold. Liu Huihai established an autoencoder network model, continuously analyzing the characteristics of reconstruction error changes, setting different thresholds to detect their trends, and using extreme value theory to set adaptive thresholds to monitor the trend changes of reconstruction error. This threshold is ultimately used as the criterion for abnormal state warnings, reducing false alarms and the impact of extreme reconstruction error values. Liu Zhaorui et al. proposed a user electricity theft detection method based on a deep autoencoder Gaussian mixture model, identifying users whose sample energy exceeds an abnormal threshold as electricity thieves, thereby achieving user electricity theft detection.

[0004] Chinese patent document CN114199570A discloses a method for identifying bearing faults in wind turbine gearboxes based on improved variational mode decomposition. The method involves: first, optimizing the number of intrinsic mode functions (EMFs) and the initial mode center constraint strength in variational mode decomposition using power spectral entropy; then, dividing the EMFs into high-noise and low-noise components using sample entropy thresholding, followed by denoising and reconstructing the high-noise EMF components using wavelet thresholding; next, extracting and reducing the fault features using a deep autoencoder; and finally, identifying bearing faults using a support vector machine.

[0005] In summary, both domestic and international scholars have conducted relevant research in the fields of bearing fault classification and new fault identification. However, research combining the two is still in its early stages. The current shortcomings are mainly reflected in the following aspects: (1) Deep learning models often require a large amount of labeled data for training. In actual industrial applications, the mechanical signals collected by the labels require professional domain knowledge and experience, and also require a lot of time. (2) Existing new fault identification models are often based on autoencoder network models. However, a single deep learning model often extracts similar features from data samples of the same category, which can easily lead to misclassification of faults and overfitting of the model. Summary of the Invention

[0006] The present invention aims to overcome at least one of the defects of the prior art and provide a new type of bearing fault identification and diagnosis method based on hybrid depth autoencoding.

[0007] The detailed technical solution of this invention is as follows:

[0008] To address the aforementioned technical problems, this invention provides a novel bearing fault identification and diagnosis method based on hybrid deep autoencoders. This method reduces the dependence of fault diagnosis on labeled data, resolves the contradiction between small sample fault data and the need for a large number of samples for deep learning model training, and sets a threshold by combining the reconstructed data of the hybrid deep autoencoder network to achieve the identification of novel bearing faults. This solves the problem of difficulty in identifying new faults in mechanical equipment fault diagnosis and greatly improves the accuracy of bearing fault identification.

[0009] To achieve the above objectives, this invention provides a novel bearing fault identification and diagnosis method based on hybrid deep autoencoder, comprising the following steps:

[0010] Step S1: Collect bearing vibration signal data under different fault conditions as samples, including samples of known fault categories and samples of unknown fault categories;

[0011] The horizontal axis of the vibration signal is time, and the vertical axis is acceleration, i.e., the acceleration of the drive end collected when the bearing is operating.

[0012] Step S2: Preprocess the samples collected in S1 to obtain the original image. Take 70% of the samples of known fault categories as the training set and assign class labels. The remaining 30% of the samples of known fault categories (without class labels) and the samples of unknown fault categories form the test set.

[0013] The preprocessing involves wavelet transforming one-dimensional data into a two-dimensional time-frequency graph, and using one-hot encoding to assign the encoding as a class label.

[0014] Step S3: Construct a hybrid deep autoencoder network, which includes convolutional layers, fully connected layers, deconvolutional layers, and a Softmax classification layer, forming a dual encoder structure consisting of a first encoder, a decoder, a second encoder, and a Softmax classification layer.

[0015] The first encoder includes four convolutional layers and two fully connected layers;

[0016] The decoder includes four deconvolutional layers and two fully connected layers;

[0017] The second encoder includes four convolutional layers and two fully connected layers;

[0018] Step S4: Input the training set into the hybrid deep autoencoder network, combine the minimum batch stochastic gradient descent method and the unsupervised learning method to pre-train the network to obtain the new class of fault detection thresholds, and realize fault classification based on class labels and hidden features;

[0019] Step S5: Input the test set into the trained hybrid deep autoencoder network for diagnosis. If it is a new type of fault, the diagnosis is completed; if it is not a new type of fault, the fault type is determined by matching the class label, and the diagnosis is completed.

[0020] Furthermore, the specific processing steps of step S3 are as follows:

[0021] Step S31: The original image x is processed by the first encoder to obtain the hidden layer feature representation z;

[0022] Step S32: The hidden layer feature representation z is processed by the decoder to generate the reconstructed image x′;

[0023] Step S33: The reconstructed image x′ is processed by the second encoder to obtain the feature representation z′ of the reconstructed image;

[0024] Step S34: The Softmax classification layer performs classification processing based on S31-S33.

[0025] Furthermore, the output feature map after convolutional layer processing is processed by ReLU activation function and batch normalized in the second convolutional layer of the first encoder and the second encoder respectively; each convolutional layer will produce a corresponding output feature map.

[0026] When the absolute value of the input value is large, the derivatives of both the Sigmoid and Tanh functions are close to 0. This can cause the error to fail to propagate downwards as the number of neural network layers increases when using backpropagation to update weights, resulting in incomplete training of the lower-level networks. This is known as the vanishing gradient phenomenon. The ReLU function, on the other hand, always has a derivative of 1 when the input value is greater than 0, which effectively overcomes the vanishing gradient phenomenon. Therefore, the ReLU function is used as the activation function of the convolutional neural network, and batch normalization (BN) is performed in the second convolutional layer to prevent overfitting.

[0027] Furthermore, the hybrid deep autoencoder network also includes a flattening layer added after the four convolutional layers as a connecting layer between the convolutional layers and the fully connected layers. The connecting layer will convert the two-dimensional feature map into one-dimensional data, and input the one-dimensional data into the fully connected layer to achieve further feature extraction.

[0028] Furthermore, the classification layer is a Softmax classification layer. Softmax is often used to solve multi-class classification problems. It maps the hidden features output by the encoder to the range [0,1], and all values ​​are added together to equal 1, directly outputting the probability of classification, thereby realizing the classification of fault types.

[0029] The input to the Softmax classification layer is the output z of the first encoder; after passing through the Softmax classification layer with C neurons, a probability vector P is obtained, P = {P1, P2, ..., P...}. C}, when there exists a maximum value P in P k If the hidden layer feature representation z represents the k-th class label, k = 1, 2, ..., C;

[0030] Furthermore, dropout is used in the Softmax classification layer to randomly discard neuron nodes, thereby reducing the co-adaptation relationship between neurons and the correlation between nodes, thus preventing the network from overfitting during training.

[0031] Furthermore, since the ultimate goals of the hybrid deep autoencoder network are twofold: first, to use the network for unsupervised training to obtain the hidden features of the input data, and then to perform supervised classification using a classifier to achieve the fault classification task; second, to determine whether a fault is of a new type by comparing the error between the original data and the reconstructed data, thus achieving the new fault identification task; therefore, the network needs to define three loss functions to achieve iterative optimization. These three loss functions are used to formulate the objective functions required for different training stages, and each loss function optimizes various parts.

[0032] The following three loss functions are defined to achieve iterative optimization of the hybrid deep autoencoder network:

[0033] Define reconstruction loss L con The reconstruction loss between the original input image x and the reconstructed image x′ is measured. The loss function is as follows:

[0034]

[0035] Define encoder loss L enc To minimize the Euclidean distance between the hidden layer feature representation z of the original data and the feature representation z′ of the reconstructed image, the loss function is expressed as follows:

[0036] L enc =||zz′||2 (4);

[0037] Define the classifier supervised training loss L st To measure the distance between the Softmax classification layer's prediction and the true label of the labeled data, the loss function is as follows:

[0038]

[0039] Where i represents the number of neurons in the hidden layer, i = 1, 2, ..., N, k is the fault type, k = 1, 2, ..., C, y i For the original data labels, These are the predicted labels after mapping through the Softmax classification layer.

[0040] Furthermore, step S4 specifically includes:

[0041] Step S41, Unsupervised Training Phase of the Hybrid Deep Autoencoder Network: Input all known fault class data into the above network for unsupervised training, obtain the hidden features of these data, and use the minimum batch gradient descent method to minimize the following objective function:

[0042] Loss=αL con +βL enc (6)

[0043] Where α and β are weights that adjust the impact of a single loss on the overall objective function;

[0044] Step S42: Labeled training of the hybrid deep autoencoder network. Labeled data is input into the hybrid deep autoencoder network for supervised training. The weights and bias vectors in the Softmax classification layer are continuously updated using gradient descent. Based on labels and hidden features, the network acquires fault classification capabilities. The objective function for the supervised training phase is as follows:

[0045] Loss=αL non +βL enc +γL st (7)

[0046] Where α, β, and γ are weights that adjust the impact of individual losses on the overall objective function;

[0047] Step S43: Record the error between the original image and the reconstructed image for known fault classes, and set a new fault detection threshold based on this error. The hybrid deep autoencoder network uses mean squared error loss to calculate the reconstruction error.

[0048]

[0049] The new fault detection threshold δ is calculated using the above formula, and the expression for δ is as follows:

[0050] δ=Q3+1.5(Q3-Q1) (9)

[0051] Where Q1 is the lower quartile of the reconstruction error of the training set, i.e., the 25th percentile; and Q3 is the upper quartile of the reconstruction error, i.e., the 75th percentile.

[0052] Furthermore, step S5 specifically includes:

[0053] S51. Input the test set into the trained hybrid deep autoencoder network to obtain the hidden layer feature representation z of the test set. TEST Reconstructed Image and the feature representation z of the reconstructed image TEST ′;

[0054] S52, Calculate the original image x of the test set. TEST and reconstructed image The reconstruction error between the two is compared with the new type of fault detection threshold δ described in step 4. If the reconstruction error MSE TEST If the value is greater than the new fault detection threshold δ, it is detected as a new fault, and the diagnosis is completed, as shown in the following formula:

[0055]

[0056] If the reconstruction error MSE TEST If the value is less than the new fault detection threshold δ, it does not belong to the new fault category. Then, by matching the class label, the fault type is obtained, and the diagnosis is completed.

[0057] The stride of each of the above-mentioned convolutional and deconvolutional layers is set to 2, and the padding is set to 1.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] (1) The present invention provides a new type of bearing fault identification and diagnosis method based on hybrid depth autoencoder, which realizes the detection function of new type faults and the classification function of known type faults. The new type detection threshold is set according to the reconstruction error, and the faults are automatically judged to be of unknown type. This solves the problem that unknown type faults will be incorrectly classified into known types.

[0060] (2) The present invention provides a new type of bearing fault identification and diagnosis method based on hybrid deep autoencoder. This method uses known type data to train the network in an unsupervised manner, obtains the reconstructed data of the samples and extracts low-dimensional features, and then uses a small amount of labeled data to fine-tune the network in a supervised manner, which effectively solves the contradiction between small sample fault data and the need for a large number of samples to train the deep learning model.

[0061] (3) The present invention provides a new type of bearing fault identification and diagnosis method based on hybrid deep autoencoder, which combines the advantages of convolutional neural network and autoencoder network. Three different loss functions are designed to optimize the network, which can effectively improve the feature extraction capability of the network, avoid extracting a single feature under small sample dataset, reduce overfitting, and thus improve the accuracy of fault classification. Attached Figure Description

[0062] Figure 1 This is a flowchart of the fault identification and diagnosis process of the present invention.

[0063] Figure 2 This is the time-frequency diagram generated by wavelet transform in Embodiment 1 of the present invention.

[0064] Figure 3 This is a diagram of the hybrid deep autoencoder network structure in Embodiment 1 of the present invention.

[0065] Figure 4 This is a flowchart of the test sample diagnosis process in Embodiment 1 of the present invention.

[0066] Figure 5 This is a graph showing the running results when the number of class-labeled samples is 264 in Embodiment 1 of the present invention. Detailed Implementation

[0067] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0068] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0069] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0070] Where there is no conflict, the embodiments and features described herein can be combined with each other.

[0071] Example 1

[0072] This embodiment provides a novel bearing fault identification and diagnosis method based on hybrid deep autoencoder, such as... Figure 1 As shown:

[0073] Step S1: Collect bearing vibration signal data under different fault conditions as samples, including samples of known fault categories and samples of unknown fault categories;

[0074] The horizontal axis of the vibration signal is time, and the vertical axis is acceleration, i.e., the acceleration of the drive end collected when the bearing is operating.

[0075] Step S2: Preprocess the samples collected in S1 to obtain the original image. Take 70% of the samples of known fault categories as the training set and assign class labels. The remaining 30% of the samples of known fault categories (without class labels) and the samples of unknown fault categories form the test set.

[0076] The preprocessing involves wavelet transforming the one-dimensional data into a two-dimensional time-frequency graph. Figure 2 As shown, one-hot encoding is used to treat the encoding as a class label.

[0077] Step S3: Construct a hybrid deep autoencoder network, which includes convolutional layers, fully connected layers, deconvolutional layers, and a softmax classification layer, forming a dual encoder structure consisting of a first encoder, a decoder, a second encoder, and a softmax classification layer, as shown below. Figure 3 As shown;

[0078] The first encoder includes four convolutional layers and two fully connected layers;

[0079] The decoder includes four deconvolutional layers and two fully connected layers;

[0080] The first encoder includes four convolutional layers and two fully connected layers;

[0081] Preferably, a flattening layer is added after the first encoder and its four convolutional layers as a connecting layer between the convolutional layers and the fully connected layer. This connecting layer transforms the two-dimensional feature map into one-dimensional data, which is then input into the fully connected layer for further feature extraction.

[0082] The specific processing steps of step S3 are as follows:

[0083] Step S31: The original image x is processed by the first encoder to obtain the hidden layer feature representation z;

[0084] Step S32: The hidden layer feature representation z is processed by the decoder to generate the reconstructed image x′;

[0085] Step S33: The reconstructed image x′ is processed by the second encoder to obtain the feature representation z′ of the reconstructed image;

[0086] Step S34: The Softmax classification layer performs classification processing based on S31-S33;

[0087] Convolutional layers use convolution kernels to perform convolution operations on local regions of the input signal (or features), generating corresponding features. The most important characteristic of convolutional layers is weight sharing, meaning the same convolution kernel traverses the input once with a fixed stride. Weight sharing reduces the number of network parameters in the convolutional layer, avoiding overfitting caused by too many parameters, and also reduces the system's memory requirements. The specific operation of a convolutional layer is shown in the following formula:

[0088]

[0089] in, Let be the output feature map of the l-th layer, and fl be the activation function of the l-th layer. is the convolution kernel of the l-th layer, bl is the bias of the l-th layer, and * is the convolution operation;

[0090] Each convolutional layer uses a 3x3 kernel with a stride of 2 and padding of 1. The size of the feature map is halved in each convolutional layer, and the feature map gradually decreases as the network goes deeper. After four consecutive convolutional layers, the original 64x64 time-frequency map will be converted into a 4x4 feature map.

[0091] The output feature maps after convolutional layer processing are processed by the ReLU activation function; that is, a ReLU activation function is added after each convolutional layer. Batch normalization is performed in the second convolutional layer of the first encoder and the second encoder, respectively. Each convolutional layer produces a corresponding output feature map. When the absolute value of the input value is large, the derivatives of the Sigmoid and Tanh functions are close to 0. This can cause the error to fail to propagate downwards as the number of neural network layers increases when updating weights using backpropagation, resulting in incomplete training of the lower-level networks. This is also known as the gradient vanishing phenomenon. The derivative of the ReLU function is always 1 when the input value is greater than 0, which effectively overcomes the gradient vanishing phenomenon. Therefore, the ReLU function is used as the activation function of the convolutional neural network, and batch normalization (BN) is performed in the second convolutional layer to prevent overfitting.

[0092] After obtaining the 4x4 feature map, a flattening layer is added after the last convolutional layer as a connector between the convolutional and fully connected layers. This connector layer transforms the two-dimensional feature map into one-dimensional data, which is then input into the fully connected layer for further feature extraction. The specific operations of the fully connected layer are as follows:

[0093] a=f(W l x+b l (2)

[0094] Where a = [a1, a2, ..., al] is the hidden layer data, f is the activation function, Wl is the weight matrix of the l-th hidden layer, and bl is the bias vector of the l-th hidden layer; after compression operations of two fully connected layers, the features extracted by the first encoder are obtained;

[0095] The input image x is processed by convolutional and fully connected layers for feature extraction, and then deconvolutional layers are used for image upsampling. The operation is similar to that of convolutional layers. The difference is that convolutional layers use parameters such as stride to reduce the size of the image, while deconvolutional layers use parameters such as stride to increase the size of the image. In the network structure proposed in this invention, the number of deconvolutional layers is still 4, the stride of each layer is set to 2, and the padding is 1. This ensures that the output of each layer is twice the size of the input, which corresponds to the convolutional layer and makes it easier to obtain a reconstructed image with the same size as the network input.

[0096] The reconstructed image is then input into the second encoder for a second pass of feature extraction. The second encoder is identical to the first encoder, consisting of four convolutional layers and two fully connected layers. The features extracted by the second encoder and the features extracted by the first encoder have the same dimensionality, allowing for comparison and calculation of the loss between the two. Subsequently, the network is optimized by continuously reducing the loss, thereby enhancing its feature extraction capabilities.

[0097] The classification layer is a Softmax classification layer. Softmax is often used to solve multi-class classification problems. It maps the hidden features output by the encoder to the range [0,1] and adds all the values ​​to 1. It can directly output the probability of classification, thereby realizing the classification of fault types.

[0098] The input to the Softmax classification layer is the output z of the first encoder; after passing through the Softmax classification layer with C neurons, a probability vector P is obtained, P = {P1, P2, ..., P...}. C}, when there exists a maximum value P in P k If the hidden layer feature representation z represents the k-th class label, k = 1, 2, ..., C;

[0099] Furthermore, dropout is used in the Softmax classification layer to randomly discard neuron nodes, thereby reducing the co-adaptation relationship between neurons and the correlation between nodes, thus preventing the network from overfitting during training.

[0100] Furthermore, since the ultimate goals of the hybrid deep autoencoder network are twofold: first, to use the network for unsupervised training to obtain the hidden features of the input data, and then to perform supervised classification using a classifier to achieve the fault classification task; second, to determine whether a fault is of a new type by comparing the error between the original data and the reconstructed data, thus achieving the new fault identification task; therefore, the network needs to define three loss functions to achieve iterative optimization. These three loss functions are used to formulate the objective functions required for different training stages, and each loss function optimizes various parts.

[0101] The following three loss functions are defined to achieve iterative optimization of the hybrid deep autoencoder network:

[0102] Define reconstruction loss L con The reconstruction loss between the original input image x and the reconstructed image x′ is measured. The loss function is as follows:

[0103]

[0104] Define encoder loss L enc To minimize the Euclidean distance between the hidden layer feature representation z of the original data and the feature representation z′ of the reconstructed image, the loss function is expressed as follows:

[0105] L enc =||zz′||2 (4);

[0106] Define the classifier supervised training loss L st To measure the distance between the Softmax classification layer's prediction and the true label of the labeled data, the loss function is as follows:

[0107]

[0108] Where i represents the number of neurons in the hidden layer, i = 1, 2, ..., N, k is the fault type, k = 1, 2, ..., C, y i For the original data labels, These are the predicted labels after mapping through the Softmax classification layer.

[0109] Step S4: Input the training set into a hybrid deep autoencoder network, and pre-train the network using a combination of minimum batch stochastic gradient descent and unsupervised learning methods to obtain new fault detection thresholds. Fault classification is then achieved based on class labels and hidden features, such as... Figure 4 As shown;

[0110] Step S4 specifically includes:

[0111] Step S41, Unsupervised Training Phase of the Hybrid Deep Autoencoder Network: Input all known fault class data into the above network for unsupervised training, obtain the hidden features of these data, and use the minimum batch gradient descent method to minimize the following objective function:

[0112] Loss=αL con +βL enc (6)

[0113] Where α and β are weights that adjust the impact of a single loss on the overall objective function;

[0114] Step S42: Labeled training of the hybrid deep autoencoder network. Labeled data is input into the hybrid deep autoencoder network for supervised training. The weights and bias vectors in the Softmax classification layer are continuously updated using gradient descent. Based on labels and hidden features, the network acquires fault classification capabilities. The objective function for the supervised training phase is as follows:

[0115] Loss=αL con +βL enc +γL st (7)

[0116] Here, α, β, and γ are weights that adjust the impact of a single loss on the overall objective function; these weights are network parameters and need to be adjusted and set to their optimal values ​​based on experimental results.

[0117] Step S43: Since only known fault class data is input for learning during the entire training process, the network can reconstruct known fault classes well, but it cannot reconstruct new fault classes well. This results in the reconstruction error of new fault classes being much larger than that of known fault classes. Therefore, it is necessary to record the error between the original and reconstructed images of known fault classes, and set a new fault detection threshold based on this error. The hybrid deep autoencoder network uses mean squared error loss to calculate the reconstruction error.

[0118]

[0119] The new fault detection threshold δ is calculated using the above formula, and the expression for δ is as follows:

[0120] δ=Q3+1.5(Q3-Q1) (9)

[0121] Where Q1 is the lower quartile of the reconstruction error of the training set, i.e., the 25th percentile; and Q3 is the upper quartile of the reconstruction error, i.e., the 75th percentile.

[0122] Step S5: Input the test set into the trained hybrid deep autoencoder network for diagnosis. If it is a new type of fault, the diagnosis is completed; if it is not a new type of fault, the fault type is determined by matching the class label, and the diagnosis is completed.

[0123] Specifically, step S5 includes:

[0124] S51. Input the test set into the trained hybrid deep autoencoder network to obtain the hidden layer feature representation z of the test set. TEST Reconstructed Image and the feature representation z of the reconstructed image TEST ′;

[0125] S52, Calculate the original image x of the test set. TEST and reconstructed image The reconstruction error between the two is compared with the new type of fault detection threshold δ described in step 4. If the reconstruction error MSE TEST If the value is greater than the new fault detection threshold δ, it is detected as a new fault, and the diagnosis is completed, as shown in the following formula:

[0126]

[0127] If the reconstruction error MSE TEST If the value is less than the new fault detection threshold δ, it does not belong to the new fault category. Then, by matching the class label, the fault type is obtained and the diagnosis is completed.

[0128] When a known fault is input into the network, the reconstructed image generated by the network after training in step 4 will be very similar to the original image, and its reconstruction error will be less than δ. This indicates that the data corresponds to a known fault. When a known fault is detected, the hidden features obtained by the network are then input into the Softmax classifier, which determines the specific type of fault, thus realizing the fault identification and classification task of the network.

[0129] However, when a new type of fault is input into the network, since the new type of fault was not learned in step 4, the network cannot reconstruct it well. Therefore, the reconstruction error calculated from the reconstructed image and the original image will be much greater than the threshold δ set in step 4. Thus, it is concluded that this fault type is a new type of fault, and the function of identifying new type of fault is realized.

[0130] To evaluate the effectiveness and accuracy of the hybrid deep autoencoder network for fault identification and diagnosis, experimental verification was conducted using a bearing dataset. The experimental data for the training and test sample datasets in this specific embodiment of the invention were obtained from bearing vibration data acquired at the Bearing Experiment Center of Case Western Reserve University. The experimental setup consisted of a 2-horsepower electric motor, a torque sensor and decoder, a power meter, and an electronic controller. The data used in this embodiment of the invention consisted of drive bearing fault data and bearing data under normal conditions at a sampling frequency of 12 kHz.

[0131] This bearing mainly exhibits three types of failures: inner ring failure, outer ring failure, and rolling element failure, with failure diameters of 0.007 inches, 0.014 inches, and 0.021 inches. The failure types used in the dataset are shown in Table 1.

[0132]

[0133]

[0134] This experiment was conducted on a Windows macOS system using the PyCharm platform and the PyTorch deep learning framework.

[0135] First, wavelet transform is performed on the above data to convert the one-dimensional data into a 64*64 two-dimensional time-frequency graph. The number of samples after the transformation is shown in the table.

[0136] Split the dataset. Assume that normal condition, inner race fault, and rolling element fault are known fault types (classes 0-6), and outer race fault (classes 7-9) are unknown faults. One-hot encoding is performed on the known fault categories of the bearing, which serve as class labels. 70% of the known fault samples are used as the training set, and the remaining 30% and all data of the unknown fault categories form the test set.

[0137]

[0138] The network training consists of two phases: unsupervised training and supervised fine-tuning. To verify the effectiveness of the proposed method, in the unsupervised training phase, unlabeled data is input into the network, and the Adam optimizer is used to continuously update the network training parameters and minimize the objective function using the minimum batch gradient descent method. The unsupervised training ends when the error is less than an acceptable value.

[0139] Then, a small amount of labeled data is input for supervised training, and the weights and bias vectors of the network are fine-tuned. Based on the class labels and hidden features, the network is made capable of fault classification. For example, when there are 200 labeled samples, the rest are all trained in unsupervised mode using unlabeled samples.

[0140] After training, a test set containing 10 types of faults is input into the network. The reconstruction error is used to determine whether it belongs to a new type of unknown fault, and the fault type is diagnosed based on its hidden features.

[0141] Multiple experiments have demonstrated that the more labeled samples in the training set, the higher the accuracy of fault classification and new class identification. When half of the training set consists of labeled data, the accuracy of both fault classification and new class identification reaches over 90%, which basically meets the requirements for fault identification and classification. (Appendix) Figure 5 The table shows the results when the number of labeled samples is 264. The horizontal axis represents the number of iterations, and the vertical axis represents the accuracy and loss value. Curve 1 represents the reconstruction loss of the test set, curve 2 represents the accuracy of the test set, and curve 3 represents the loss of the training set. As can be seen from the table, the loss of the training set and the test set decreases with the increase of the number of iterations, while the overall accuracy of the test set increases with the increase of the number of iterations and stabilizes at around 98%, indicating high recognition and classification accuracy.

[0142]

[0143] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A novel bearing fault identification and diagnosis method based on hybrid deep autoencoder, characterized in that, include; Step S1: Collect bearing vibration signal data under different fault conditions as samples, including samples of known fault categories and samples of unknown fault categories; Step S2: Preprocess the samples collected in S1 to obtain the original image. Take 70% of the samples of known fault categories as the training set and assign class labels. The remaining 30% of the samples of known fault categories and unknown fault categories form the test set. Step S3: Construct a hybrid deep autoencoder network, which includes a first encoder, a decoder, a second encoder, and a Softmax classification layer; The first encoder includes four convolutional layers and two fully connected layers; The decoder includes four deconvolutional layers and two fully connected layers; The second encoder includes four convolutional layers and two fully connected layers; Step S4: Input the training set into the hybrid deep autoencoder network, combine the minimum batch stochastic gradient descent method and the unsupervised learning method to pre-train the network to obtain the new class of fault detection thresholds, and realize fault classification based on class labels and hidden features; Step S5: Input the test set into the hybrid deep autoencoder network trained on the training set for diagnosis; If it is a new type of fault, then the diagnosis is complete; If it is not a new type of fault, then the fault type is determined by matching the class tags to complete the diagnosis. The following three loss functions are defined to achieve iterative optimization of the hybrid deep autoencoder network: Define reconstruction loss The reconstruction loss is measured between the original input image x and the reconstructed image x′, and the loss function is as follows: （3）； Define encoder loss To minimize the Euclidean distance between the hidden layer feature representation z of the original data and the feature representation z′ of the reconstructed image, the loss function is expressed as follows: （4）； Define the classifier supervised training loss To measure the distance between the Softmax classification layer's prediction and the true class label of the labeled data, the loss function is as follows: （5）； Where i represents the number of neurons in the hidden layer, i=1,2,…,N, and k represents the fault type, k=1,2,…,C. For raw data type labels, , The predicted class label after mapping through the Softmax classification layer; Step S4 specifically includes: Step S41, Unsupervised Training Phase of the Hybrid Deep Autoencoder Network: Input all known fault class data into the above network for unsupervised training, obtain the hidden features of these data, and use the minimum batch gradient descent method to minimize the following objective function: （6）； in, and It adjusts the weights of the impact of individual losses on the overall objective function; Step S42: Class-labeled training of the hybrid deep autoencoder network. Class-labeled data is input into the hybrid deep autoencoder network for supervised training. The weights and bias vectors in the Softmax classification layer are continuously updated using gradient descent. The objective function for the supervised training phase is as follows: （7）； in, , and It adjusts the weights of the impact of individual losses on the overall objective function; Step S43: Record the error between the original image and the reconstructed image for known fault classes, and set a new fault detection threshold based on this error. The hybrid deep autoencoder network uses mean squared error loss to calculate the reconstruction error. （8）； The threshold δ for detecting new types of faults is calculated using the above formula. The expression for δ is as follows: （9）； Where Q1 is the lower quartile of the reconstruction error of the training set, i.e., the 25th percentile; and Q3 is the upper quartile of the reconstruction error, i.e., the 75th percentile.

2. The method for identifying and diagnosing novel bearing fault types based on hybrid deep autoencoder as described in claim 1, characterized in that, The specific processing steps of step S3 are as follows: Step S31: The original image x is processed by the first encoder to obtain the hidden layer feature representation z; Step S32: The hidden layer feature representation z is processed by the decoder to generate the reconstructed image x′; Step S33: The reconstructed image x′ is processed by the second encoder to obtain the feature representation z′ of the reconstructed image; Step S34: The Softmax classification layer performs classification processing based on S31-S33.

3. A novel bearing fault identification and diagnosis method based on hybrid deep autoencoder according to claim 1 or 2, characterized in that, The output feature map after being processed by the convolutional layer is processed by the ReLU activation function and then batch normalized in the second convolutional layer of the first encoder and the second encoder, respectively.

4. The method for identifying and diagnosing novel bearing fault types based on hybrid deep autoencoder as described in claim 3, characterized in that, The hybrid deep autoencoder network also includes a flattening layer added after the four convolutional layers as a connecting layer between the convolutional layers and the fully connected layers.

5. The method for identifying and diagnosing novel bearing fault types based on hybrid deep autoencoder according to claim 2, characterized in that, The input to the Softmax classification layer is the output of the first encoder. ; After passing through a Softmax classification layer with C neurons, a probability vector P is obtained. When there exists a maximum value in P Then it is assumed that the hidden layer feature representation z represents the label of the k-th class. ; In the Softmax classification layer, dropout is used to randomly discard neuron nodes.

6. The method for identifying and diagnosing novel bearing fault types based on hybrid deep autoencoder according to claim 5, characterized in that, Step S5 specifically includes: S51. Input the test set into the trained hybrid deep autoencoder network to obtain the hidden layer feature representation of the test set. Reconstructed Image and feature representation of reconstructed images ′; S52, Calculate the original images of the test set. and reconstructed image The reconstruction error between them is compared with the new type of fault detection threshold δ described in step 4. If the reconstruction error is... If the value is greater than the new fault detection threshold δ, it is detected as a new fault, and the diagnosis is completed, as shown in the following formula: （10）； If reconstruction error If the value is less than the new fault detection threshold δ, it does not belong to the new fault category. Then, by matching the class label, the fault type is obtained, and the diagnosis is completed.

7. The method for identifying and diagnosing novel bearing fault types based on hybrid deep autoencoder according to claim 1, characterized in that, The horizontal axis of the vibration signal represents time, and the vertical axis represents acceleration, specifically the acceleration of the drive end collected during bearing operation.

8. The method for identifying and diagnosing novel bearing fault types based on hybrid deep autoencoder according to claim 1, characterized in that, The preprocessing described in step S2 consists of wavelet transform and one-hot encoding.

Images