A novel fault identification method and device based on a flow discrimination model

Abstract

The present invention discloses a novel fault identification method and device based on a flow discrimination model. The method includes: preprocessing known bearing monitoring data to obtain a closed data set and class labels; building a fault diagnosis model, which includes a feature extractor and a classifier; training the fault diagnosis model based on the closed data set and class labels, and using the trained fault diagnosis model to extract high-dimensional features of each monitoring data sample in the closed data set; building a flow discrimination model, which includes a flow model and a discriminator; training the flow discrimination model based on the closed data set and its class labels, and performing online fault diagnosis based on the trained fault diagnosis model and flow discrimination model to achieve the diagnosis of bearing fault categories. In the present invention, the flow model can promote the convergence of the discriminator when performing adversarial training with the discriminator, thereby solving the problem of difficult selection of optimal parameters for the discriminator. The proposed solution has a high recognition accuracy for novel faults.

Description

Technical Field

[0001] The present invention relates to the field of intelligent diagnosis of mechanical faults, and particularly to a novel fault identification method and device based on a flow discrimination model. Background Art

[0002] With the development of technology, mechanical structures have become increasingly precise and complex. As the most vulnerable component, the health status of bearings directly affects the stability and safety of mechanical operations. If bearing faults are not detected in a timely manner, it may lead to the functional failure of large-scale machinery, resulting in significant economic losses and even casualties. Therefore, timely monitoring and diagnosis of bearing faults are of great significance for ensuring the stable and safe operation of machinery.

[0003] In recent years, bearing intelligent diagnosis technology based on deep learning has become a research hotspot. Deep neural networks have powerful feature extraction capabilities and can accurately diagnose various bearing faults after sufficient and complete data training. However, it is difficult to obtain sufficient and complete data in engineering practice. To ensure the safety and stability of mechanical systems, bearings are not allowed to operate with faults in engineering. Therefore, bearings are in a normal operating state for a long time, with short operating times in fault states and small amounts of data for different fault types, resulting in an imbalanced bearing monitoring dataset. An imbalanced bearing dataset is prone to overfitting problems for minority-class faults when training a fault diagnosis model. The main solution to imbalanced bearing fault diagnosis is the rebalancing method, including generating sufficient minority-class fault samples through a deep learning model to rebalance the training samples, or giving a larger weight to the minority class in the loss function to alleviate the overfitting phenomenon of the minority class. As a generative model, the generative adversarial network can generate approximate samples of the minority class and expand the dataset, which is a commonly used method to solve the problem of imbalanced bearing fault diagnosis.

[0004] In addition to some typical faults, bearings may occasionally experience some atypical faults. These atypical fault data are even rarer and harder to obtain, so they cannot be used for training the fault diagnosis model but may occur in online fault diagnosis. Existing fault diagnosis models can only identify unparticipated atypical fault categories as known fault categories and cannot identify them as novel fault categories. The rebalancing method for solving imbalanced fault diagnosis also cannot identify novel fault categories when there is no novel fault data for training. How to achieve the identification of novel faults is a pain point in the field of bearing fault diagnosis.

[0005] Currently, the new fault identification methods mainly include the method based on feature threshold and the method based on generative model. The method based on feature threshold first uses a deep neural network to learn sample features, and then divides the sample features by setting a threshold to determine whether the sample belongs to a known class or a new class. The method based on generative model, such as generative adversarial network, first uses a generator to reconstruct an approximate sample of the real sample, and then uses the real sample and the approximate sample to train the discriminator's ability to identify new faults.

[0006] However, the method based on feature threshold needs to determine a threshold to judge whether a sample belongs to a new fault. However, the determination of the threshold is only based on the knowledge learned from existing classes. Since new fault class samples cannot be accessed, it is difficult to determine a suitable threshold for this type of method, and it is easy to make mistakes in identifying new faults. The method based on generative model has problems that the model is difficult to converge and the optimal parameters of the model are difficult to determine. For example, although the output of the discriminator of the generative adversarial network can be directly used as the likelihood function of the new fault class, it continuously confronts the generator during the training process, and the effect of the discriminator is sometimes good and sometimes bad, making it difficult to converge. It is difficult to determine which model parameters to choose for new fault identification. Therefore, the existing new fault identification methods have at least the following disadvantages: (1) The method based on generative model has a difficult-to-converge model and difficult-to-determine optimal parameters; (2) The recognition accuracy of new faults is not high; (3) The method based on feature threshold is difficult to determine a suitable threshold. Summary of the Invention

[0007] Object of the Invention: The object of the present invention is to provide a new fault identification method and device based on a flow discriminant model, which uses a flow model to perform adversarial training with a discriminator. Since the flow model has a strong generative ability, the flow model can promote the convergence of the discriminator when performing adversarial training with the discriminator, thereby solving the problem of difficult selection of the optimal parameters of the discriminator. This solution can also achieve accurate identification of new faults in bearings.

[0008] Technical Solution: The new fault identification method based on a flow discriminant model of the present invention includes the following steps:

[0009] Perform data preprocessing on the known bearing monitoring data to obtain a closed data set and its class labels;

[0010] Combine a feature extractor and a classifier to build a fault diagnosis model. Among them, the feature extractor is used to extract high-dimensional features of data samples, and the classifier is used to classify the extracted high-dimensional features;

[0011] Train the fault diagnosis model based on the closed data set and its class labels to obtain a trained fault diagnosis model, and use the trained fault diagnosis model to extract the high-dimensional features of each monitoring data sample in the closed data set;

[0012] The combined flow model and discriminant model, where the flow model and the discriminator are in confrontation with each other during the training process to form a flow discriminant model. Among them, the flow model is used to reconstruct approximate samples of real samples, and the discriminator is used to judge whether the input high-dimensional features come from real samples or approximate samples, and is also used to judge whether the samples belong to known-class faults or new faults in online fault diagnosis;

[0013] Train the flow discriminant model based on the closed dataset and its class labels to obtain a trained flow discriminant model; and

[0014] Perform online fault diagnosis based on the trained fault diagnosis model and flow discriminant model to realize the diagnosis of bearing fault categories.

[0015] Furthermore, preprocess the known-class bearing monitoring data to obtain a closed dataset and its class labels, including the following steps:

[0016] Divide the known-class bearing monitoring data according to the bearing health status categories to obtain several monitoring data samples and the class labels corresponding to the monitoring data samples, and perform short-time Fourier transform on each monitoring data sample to obtain the time-frequency spectrum of the monitoring data sample. All monitoring data samples form a closed dataset.

[0017] Furthermore, the feature extractor is constructed by using one of a fully connected network, a deep convolutional network, a deep belief network, and a deep residual network;

[0018] The classifier is constructed by using one of a support vector machine, a k-nearest neighbor algorithm, a random forest, a fuzzy system, and a Softmax function.

[0019] Furthermore, during the process of training the fault diagnosis model based on the closed dataset and its class labels, cross-entropy loss is used as the loss function for training the fault diagnosis model, and one of an adaptive moment estimation algorithm, a stochastic gradient descent method, and a root mean square propagation algorithm is used as the optimization algorithm for training.

[0020] Furthermore, the flow model for reconstructing approximate samples of real samples includes the following steps:

[0021] The flow model adopts the structure of an autoencoder and fits the mapping relationship between the actual dataset and a simple prior distribution through a series of reversible transformations; assume that f(·) is the mapping function of the flow model, and f --1 (·) is its inverse function. During inference, randomly sample a monitoring data sample x from the closed dataset and map it to a vector z of a simple known distribution through f --1 (·); during generation, randomly sample a vector z from the known distribution and obtain the generated approximate sample through f(·)

[0022] Further, the discriminator is constructed using a neural network structure, which is one of a fully connected network, a deep convolutional network, a deep belief network, and a deep residual network. A sigmoid function is added to the last layer of the neural network structure. The high-dimensional features of the input monitoring data samples are input, and the output value is constrained between 0 and 1.

[0023] Further, based on the closed dataset and its class labels, the flow discriminant model is trained to obtain a trained flow discriminant model, including the following steps:

[0024] Input the high-dimensional features of the monitoring data samples in the extracted closed dataset into the discriminator, and calculate the loss between the output result and 1;

[0025] Randomly sample a vector z from the known distribution, and obtain an approximate sample through f(z) Input the approximate sample into the trained fault diagnosis model to extract the high-dimensional features of the approximate sample, and then input the high-dimensional features of the approximate sample into the discriminator, and calculate the loss between the output result and 0; and

[0026] Alternately train the discriminator and the flow model, record the loss of the discriminator each time, save the parameters of the flow discriminant model when the loss is the lowest, and obtain a trained flow discriminant model.

[0027] Furthermore, the loss of the discriminator each time includes the binary cross-entropy loss between the output corresponding to the closed dataset sample and 1 and the binary cross-entropy loss between the output corresponding to the randomly known distribution vector and 0.

[0028] Further, based on the trained flow discriminant model, online fault diagnosis is performed to realize the diagnosis of bearing fault categories, including the following steps:

[0029] Input the time-frequency spectrum of the monitoring data sample into the trained fault diagnosis model to extract high-dimensional features, and input the extracted high-dimensional features into the discriminator to determine whether the monitoring data sample belongs to a new fault; if not, obtain the known fault category to which the monitoring data sample belongs according to the output of the fault diagnosis model; if it is determined that the monitoring data sample belongs to a new fault, the detected new fault category does not belong to the closed dataset and belongs to the open dataset in the model performance test.

[0030] The present invention also includes a new fault recognition device based on a flow discriminant model. The new fault recognition device based on a flow discriminant model is used to implement the new fault recognition method based on a flow discriminant model as described in claim 1, including

[0031] A data preprocessing module for preprocessing known bearing monitoring data to obtain a closed data set and its class labels; a fault diagnosis model building module for building a fault diagnosis model, the fault diagnosis model including a feature extractor and a classifier;

[0032] A high-dimensional feature extraction module for training the fault diagnosis model based on the closed data set and its class labels to obtain a trained fault diagnosis model, and using the trained fault diagnosis model to extract the high-dimensional features of each monitoring data sample in the closed data set;

[0033] A flow discriminant model building module for building a flow discriminant model, the flow discriminant model including a flow model and a discriminator;

[0034] A flow discriminant model training module for training the flow discriminant model based on the closed data set and its class labels to obtain a trained flow discriminant model;

[0035] An online fault diagnosis module for performing online fault diagnosis based on the trained flow discriminant model to achieve the diagnosis of bearing fault categories.

[0036] Advantageous effects: Compared with the prior art, the technical solution of the present invention has the following advantageous effects:

[0037] (1) The discriminator can converge, and the optimal parameters of the discriminator are easy to determine: The flow model adopted by the present invention has stronger generation ability than the traditional generation model. When the flow model performs adversarial training with the discriminator, it can promote the convergence of the discriminator, thus solving the problem of difficult selection of the optimal parameters of the discriminator;

[0038] (2) The present invention uses the high-dimensional features of the monitoring data samples of the bearing as the input of the discriminator, rather than the samples themselves as the input of the discriminator, avoiding the problem of possible noise interference in the original data samples, improving the discriminability of the discriminator for real samples and approximate samples, as well as for known class faults and new faults, and achieving accurate identification of new bearing faults.

[0039] (3) The solution of the present invention avoids the problem of setting thresholds compared with the existing feature threshold-based methods. Description of the Drawings

[0040] Figure 1 It is a schematic flow chart of the novel fault identification method based on the flow discriminant model of the present invention;

[0041] Figure 2 It is a loss curve graph of the discriminator in the T1 task of the embodiment of the present invention;

[0042] Figure 3 It is a receiver operating characteristic curve graph in the T1 task of the embodiment of the present invention;

[0043] Figure 4 This is the confusion matrix diagram of the classification results of the T1 task test data in the embodiments of the present invention. Specific embodiments

[0044] The technical solution of the present invention will be introduced in detail below in combination with the specific embodiments and the accompanying drawings of the specification.

[0045] As Figure 1 shown, the novel fault identification method based on the flow discrimination model of the present invention includes the following steps:

[0046] S101: Data preprocessing: Perform data preprocessing on the known bearing monitoring data to obtain a closed data set and its class labels, including the following steps:

[0047] Divide the known bearing monitoring data according to the bearing health status categories to obtain several monitoring data samples and the class labels corresponding to the monitoring data samples, and perform short-time Fourier transform on each monitoring data sample to obtain the time-frequency spectrum of the monitoring data sample. All the monitoring data samples form a closed data set. The closed data set and its class labels all participate in the training of the subsequent model.

[0048] S102: Fault diagnosis model construction: Combine a feature extractor and a classifier to construct a fault diagnosis model. The feature extractor is used to extract high-dimensional features of the data samples, and the classifier is used to classify the extracted high-dimensional features.

[0049] Among them, the feature extractor includes, but is not limited to, using one of a fully connected network, a deep convolutional network, a deep belief network, and a deep residual network to construct; the classifier includes, but is not limited to, using one of a support vector machine, a k-nearest neighbor algorithm, a random forest, a fuzzy system, and a Softmax function to construct.

[0050] S103: High-dimensional feature extraction: Use the closed data set and its class labels to train the fault diagnosis model to obtain a trained fault diagnosis model, and use the trained fault diagnosis model to extract the high-dimensional features of each monitoring data sample in the closed data set.

[0051] During the process of training the fault diagnosis model based on the closed data set and its class labels, cross-entropy loss is used as the loss function for training the fault diagnosis model, and one of the adaptive moment estimation algorithm, the stochastic gradient descent method, and the root mean square propagation algorithm is used as the optimization algorithm for training.

[0052] S104: Flow discriminant model construction: Combine the flow model and the discriminant model. The flow model and the discriminator compete with each other during training to form the flow discriminant model. The flow model is used to reconstruct approximate samples of real samples, and the discriminator is used to determine whether the input high-dimensional features come from real samples or approximate samples, and is also used to determine whether the samples belong to known-class faults or new faults in online fault diagnosis.

[0053] Among them, the flow model is used to reconstruct approximate samples of real samples, including the following steps:

[0054] The flow model adopts the structure of an autoencoder, and fits the mapping relationship between the actual data set and the simple prior distribution through a series of reversible transformations, and has strong generative ability; assume that f(·) is the mapping function of the flow model, and f --1 (·) is its inverse function; during inference, randomly sample the monitoring data sample x from the closed data set, and through f --1 (·) can be mapped to a vector z of a simple known distribution; during generation, randomly sample the vector z from the known distribution, and obtain the generated approximate sample through f(·)

[0055] The flow model searches for a reversible mapping between the known distribution and the unknown distribution, including encoding and decoding. The basic architecture of the flow model is the architecture of an autoencoder. The flow model is constructed using one of Nice, RealNVP, and Glow. Nice, RealNVP, and Glow are three mainstream flow models, all belonging to the autoencoder structure.

[0056] The discriminator is constructed using a neural network structure, which is one of, but not limited to, a fully connected network, a deep convolutional network, a deep belief network, and a deep residual network. A sigmoid function is added to the last layer of the neural network structure. The high-dimensional features of the monitoring data sample of the input bearing are input, and the output value is constrained between 0 and 1.

[0057] S105: Flow discriminant model training: Use the closed data set and its labels to train the flow discriminant model to obtain the trained flow discriminant model, including the following steps:

[0058] Input the high-dimensional features of the monitoring data samples in the extracted closed data set into the discriminator, and the output result passes through the sigmoid function to obtain a value between 0 and 1, and calculate the binary cross-entropy loss with 1;

[0059] Randomly sample the vector z from the known distribution, and obtain the approximate sample through f(z) Extract the high-dimensional features of the approximate samples from the trained fault diagnosis model, then input the high-dimensional features of the approximate samples into the discriminator, and the output result passes through the sigmoid function to obtain a value between 0 and 1, and calculate the binary cross-entropy loss with 0;

[0060] Alternately train the discriminator and the flow model, record the loss of the discriminator each time, save the parameters of the flow discriminator model when the loss is the lowest, and obtain the trained flow discriminator model.

[0061] Among them, the training loss of the flow model is the mean squared sum of the output after the input sample x passes through the flow operation, and the selected optimization algorithm is used to optimize its loss. The loss of the discriminator each time includes the binary cross-entropy loss between the output corresponding to the samples in the closed dataset and 1 and the binary cross-entropy loss between the output corresponding to the randomly known distribution vector and 0.

[0062] The optimization algorithms for training the flow discriminator model include, but are not limited to, one of the adaptive moment estimation algorithm, the stochastic gradient descent method, and the root mean square propagation algorithm.

[0063] S106: Online fault diagnosis: Perform online fault diagnosis based on the trained fault diagnosis model and the flow discriminator model to realize the diagnosis of bearing fault categories, including the following steps:

[0064] Input the time-frequency spectrum of the monitored data sample into the trained fault diagnosis model to extract high-dimensional features, and input the extracted high-dimensional features into the discriminator to determine whether the monitored data sample belongs to a new type of fault; if not, obtain the known fault category to which the monitored data sample belongs according to the output of the fault diagnosis model; if it is determined that the monitored data sample belongs to a new type of fault, the detected new type of fault category does not belong to the closed dataset and belongs to the open dataset in the model performance test.

[0065] To more clearly understand the technical solution and its effect of the present invention, the following combines a specific embodiment to detail the solution.

[0066] Taking the bearing fault diagnosis as an example, the bearing model is SKF 6205-2RS, the inner ring of the bearing is driven to rotate by a motor, the rotation speed is set to 896 rpm, an acceleration sensor is installed on the bearing housing to collect the vibration signal of the bearing, the sampling frequency is 10 kHz, and 7 bearing health states are simulated in the experiment, including: healthy (N), inner ring fault (IR1) of the bearing with a fault size of 0.2 mm, outer ring fault (OR1) of the bearing, rolling element fault (B1) of the bearing, inner ring fault (IR2) of the bearing with a fault size of 0.3 mm, outer ring fault (OR2) of the bearing, rolling element fault (B2) of the bearing. Three tasks are set in the experiment, and each task contains different category types and numbers in the closed set and the open set, and the category division is shown in the following table.

[0067] Bearing health status category division table in different fault diagnosis tasks

[0068]

[0069] The training set only contains closed sets, with the same number of samples in each class, all containing 600 groups of signal samples. The test set is divided into a closed test set and an open test set, and each category contains 100 groups of signal samples. The length of each signal sample is 1024 data points. The fault diagnosis model and the flow discriminant model are trained using the training set, and the model effect is verified using the test set.

[0070] Subsequently, the new fault identification method based on the flow discriminant model of the present invention is used to perform fault diagnosis on three tasks, and the specific details are as follows:

[0071] S101: Data preprocessing

[0072] The known-class bearing monitoring data is divided according to the bearing health status category, and the short-time Fourier transform is performed on each monitoring data sample. The number of signal data points is 1024, the Hamming window is selected, the window length is 128, and it moves 3 points to the right each time to obtain the time-frequency spectrum and uniformly adjust the time-frequency spectrum size to 32*32. According to Figure 2 the three tasks shown, set the number of categories of the data set.

[0073] S102: Fault diagnosis model construction

[0074] The fault diagnosis model consists of a feature extractor and a classifier. The feature extractor is used to extract the high-dimensional features of the data samples, and the classifier is used to classify the extracted high-dimensional features.

[0075] In this embodiment, the Resnet-18 convolutional neural network is used to train the closed bearing data set. Among them, the data features are extracted through multiple residual blocks, and the ReLU activation function layer is connected between the convolutional layers. The classifier uses a fully connected network, followed by a Softmax activation function. According to the different numbers of closed set types in the task, the model finally outputs an n-dimensional vector to represent the category of the input data.

[0076] S103: High-dimensional feature extraction

[0077] The closed data set and its labels are used to train the fault diagnosis model to obtain the trained fault diagnosis model and the high-dimensional features of each data sample.

[0078] The cross-entropy loss is used as the loss function of the fault diagnosis model, and the adaptive moment estimation algorithm is used as the optimization algorithm in the training process.

[0079] S104: Flow discriminant model construction

[0080] The flow discrimination model consists of a flow model and a discriminator. The flow model is used to reconstruct approximate samples of real samples, and the discriminator is used to determine whether the input high-dimensional features come from real samples or approximate samples, and is also used to determine whether the samples belong to known-class faults or new faults in online fault diagnosis. The discriminator adopts a five-layer convolutional neural network structure with a convolution kernel size of 1*1. A batch normalization layer is set behind the convolutional layer to pull the eigenvalue distribution back to the standard normal distribution. A LeakyReLU activation function is connected between the convolutional layers. A sigmoid function is added to the last layer of the convolutional neural network. The high-dimensional features of the input bearing data are input, and the output will be constrained between 0 and 1.

[0081] The flow model uses Glow, which follows the framework of RealNVP and introduces 1*1 invertible convolution to replace the sorting layer. During inference, x is randomly sampled from the closed bearing dataset, and through f -1 (x), it can be mapped to a vector z with a simple known distribution. During generation, a vector z is randomly sampled from the known distribution, and the generated approximate sample can be obtained through f(z).

[0082] The flow model and the discriminator are combined and antagonize each other during training to form a flow discrimination model.

[0083] S105: Training of the flow discrimination model

[0084] The flow discrimination model is trained using a closed dataset and its labels to obtain a trained flow model and discriminator. During training, the high-dimensional features extracted from the closed set are input into the discrimination model, and the binary cross-entropy loss is calculated between the output result and 1. Samples z are randomly drawn from the Gaussian distribution, approximate samples are obtained through f(z), the high-dimensional features are extracted, and then input into the discrimination model, and the binary cross-entropy loss is calculated between the output result and 0. During the backpropagation process, the adaptive moment estimation algorithm (ADAM) is used to optimize the model. The learning rate of the discrimination model is set to 0.001, the learning rate of the flow model is set to 0.0001, and the number of iterations is 20. The discriminator and the flow model are alternately trained, the loss of the discriminator each time is recorded, and the parameters of the flow discrimination model with the lowest loss are saved.

[0085] As Figure 2 shown, the loss curve of the discriminator in Task T1 is given. It can be seen that the loss of the discriminator can converge quickly. The point with the lowest loss is selected, and the model parameters are saved.

[0086] S106: Online fault diagnosis

[0087] First, input the time-frequency spectrum of the sample into the trained fault diagnosis model to extract high-dimensional features, and then input the high-dimensional features into the discriminator to determine whether the sample belongs to a new type of fault. If not, obtain the known fault category to which the sample belongs according to the output of the fault diagnosis model. The detected new fault categories do not belong to the closed dataset and belong to the open dataset in the model performance test. The following table gives the closed-set accuracy, open-set accuracy, and average accuracy of the datasets in different tasks.

[0088] Accuracy of bearing health state recognition in different fault diagnosis tasks

[0089] It can be seen that the method proposed in the present invention can achieve 100% open-set accuracy under the conditions of different numbers of closed-set and open-set categories, which means that new unknown faults can be stably identified. For the closed-set data, the model proposed in the present invention can also accurately classify it into known classes and output its fault categories.

[0090] Taking the T1 task as an example, input 100 test samples of each category into the model and draw its receiver operating characteristic curve (ROC), as Figure 3 shown. It can be seen that the ROC curve is very close to the upper left corner, indicating that the model can accurately judge both real samples and false samples. The area under the ROC curve is 1, and there is at least one threshold that can obtain perfect prediction.

[0091] Taking the T3 task as an example, output the classification labels of the test results. By comparing with the true state labels, obtain the fault diagnosis results of each data in the test set, and represent them with a confusion matrix, as Figure 4 shown. It can be seen that the method of the present invention has a high accuracy for identifying new unknown faults, which can reach 100%, and can also accurately classify known classes, with only one error item.

[0092] In summary, the flow discriminant model of the present invention uses a flow model and a discriminator for adversarial training to promote the convergence of the discriminator and solve the problem of optimal parameter selection of the discriminator. Using the high-dimensional features of bearing monitoring data samples as the input of the discriminator avoids the problem of possible noise interference in the original data samples and improves the discriminant ability of the discriminator for real samples and approximate samples, as well as for known-class faults and new faults. The discriminator of the present invention can converge quickly, has a strong discriminant ability for known-class faults and new faults, and has a high accuracy for identifying new faults.

[0093] The present invention also includes a new fault identification device based on the flow discriminant model. The new fault identification device based on the flow discriminant model is used to implement the new fault identification method based on the flow discriminant model, including

[0094] A data preprocessing module for preprocessing known bearing monitoring data to obtain a closed dataset and its class labels;

[0095] A fault diagnosis model building module for building a fault diagnosis model, where the fault diagnosis model includes a feature extractor and a classifier;

[0096] A high-dimensional feature extraction module for training the fault diagnosis model based on the closed dataset and its class labels to obtain a trained fault diagnosis model, and using the trained fault diagnosis model to extract high-dimensional features of each monitoring data sample in the closed dataset;

[0097] A flow discrimination model building module for building a flow discrimination model, where the flow discrimination model includes a flow model and a discriminator;

[0098] A flow discrimination model training module for training the flow discrimination model based on the closed dataset and its class labels to obtain a trained flow discrimination model;

[0099] An online fault diagnosis module for performing online fault diagnosis based on the trained flow discrimination model to realize the diagnosis of bearing fault categories.

Claims

1. A novel fault identification method based on a flow discrimination model, characterized in that, It includes the following steps: Perform data preprocessing on the known bearing-like monitoring data to obtain a closed data set and its class labels; Combine a feature extractor and a classifier to build a fault diagnosis model. Among them, the feature extractor is used to extract high-dimensional features of data samples, and the classifier is used to classify the extracted high-dimensional features; Train the fault diagnosis model based on the closed data set and its class labels to obtain a trained fault diagnosis model, and use the trained fault diagnosis model to extract the high-dimensional features of each monitoring data sample in the closed data set; Combine a flow model and a discriminative model. The flow model and the discriminator compete with each other during training to form a flow discriminative model. Among them, the flow model is used to reconstruct an approximate sample of a real sample, and the discriminator is used to judge whether the input high-dimensional features come from a real sample or an approximate sample, and is also used to judge whether a sample belongs to a known-class fault or a new fault in online fault diagnosis; Train the flow discriminative model based on the closed data set and its class labels to obtain a trained flow discriminative model; and Perform online fault diagnosis based on the trained fault diagnosis model and flow discriminative model to realize the diagnosis of bearing fault classes; Perform data preprocessing on the known bearing-like monitoring data to obtain a closed data set and its class labels, including the following steps: Divide the known bearing-like monitoring data according to the bearing health status categories to obtain several monitoring data samples and the class labels corresponding to the monitoring data samples, and perform short-time Fourier transform on each monitoring data sample to obtain the time-frequency spectrum of the monitoring data sample. All monitoring data samples form a closed data set; The flow model is used to reconstruct an approximate sample of a real sample, including the following steps: The flow model adopts the structure of an autoencoder and fits the mapping relationship between the actual data set and the simple prior distribution through a series of reversible transformations. Assume that f(·) is the mapping function of the flow model, and f --1 (·) is its inverse function. During inference, monitor data samples x are randomly sampled from the closed data set and mapped to vectors z of a simple known distribution through f --1 (·). During generation, vectors z are randomly sampled from the known distribution, and approximate samples generated are obtained through f(·) Train the flow discriminative model based on the closed data set and its class labels to obtain a trained flow discriminative model, including the following steps: Input the high-dimensional features of the monitoring data samples in the extracted closed data set into the discriminator, and calculate the loss with the output result and 1; Randomly sample a vector z from a known distribution and obtain an approximate sample through f(z). Input the approximate sample into the trained fault diagnosis model to extract the high-dimensional features of the approximate sample, and then input the high-dimensional features of the approximate sample into the discriminator to calculate the loss between the output result and 0; and Alternately train the discriminator and the flow model, record the loss of the discriminator each time, save the parameters of the flow discriminative model when the loss is the lowest, and obtain a trained flow discriminative model.

2. The novel fault identification method based on a flow discrimination model according to claim 1, wherein: The feature extractor is constructed by using one of a fully connected network, a deep convolutional network, a deep belief network, and a deep residual network; The classifier is constructed by using one of a support vector machine, a k-nearest neighbor algorithm, a random forest, a fuzzy system, and a Softmax function.

3. The novel fault identification method based on a flow discrimination model according to claim 1, characterized in that, During the process of training the fault diagnosis model based on the closed data set and its class labels, cross-entropy loss is used as the loss function for training the fault diagnosis model, and one of an adaptive moment estimation algorithm, a stochastic gradient descent algorithm, and a root mean square propagation algorithm is used as the optimization algorithm for training.

4. The novel fault identification method based on a flow discrimination model according to claim 1, characterized in that: The discriminator is constructed by using a neural network structure, which is one of a fully connected network, a deep convolutional network, a deep belief network, and a deep residual network. A sigmoid function is added to the last layer of the neural network structure. Input the high-dimensional features of the monitoring data sample, and the output value is constrained between 0 and 1.

5. The novel fault identification method based on a flow discrimination model according to claim 1, wherein The loss of the discriminator each time includes the binary cross-entropy loss between the output corresponding to the samples in the closed dataset and 1, and the binary cross-entropy loss between the output corresponding to the randomly known distribution vectors and 0.

6. The novel fault identification method based on a flow discrimination model according to claim 1, wherein Based on the trained flow discriminant model, online fault diagnosis is carried out to realize the diagnosis of bearing fault categories, including the following steps: Input the time-frequency spectrum of the monitored data samples into the trained fault diagnosis model to extract high-dimensional features, and input the extracted high-dimensional features into the discriminator to determine whether the monitored data sample belongs to a new type of fault; if not, obtain the known fault category to which the monitored data sample belongs according to the output of the fault diagnosis model; if it is determined that the monitored data sample belongs to a new type of fault, the detected new type of fault category does not belong to the closed dataset and belongs to the open dataset in the model performance test.

7. A novel fault identification device based on a flow discrimination model, the novel fault identification device based on a flow discrimination model is used to implement the novel fault identification method based on a flow discrimination model as described in claim 1, characterized in that: Including A data preprocessing module for preprocessing the known bearing-like monitored data to obtain a closed dataset and its class labels; A fault diagnosis model building module for building a fault diagnosis model, where the fault diagnosis model includes a feature extractor and a classifier; A high-dimensional feature extraction module for training the fault diagnosis model based on the closed dataset and its class labels to obtain a trained fault diagnosis model, and using the trained fault diagnosis model to extract the high-dimensional features of each monitored data sample in the closed dataset; A flow discriminant model building module for building a flow discriminant model, where the flow discriminant model includes a flow model and a discriminator; A flow discriminant model training module for training the flow discriminant model based on the closed dataset and its class labels to obtain a trained flow discriminant model; An online fault diagnosis module for performing online fault diagnosis based on the trained flow discriminant model to realize the diagnosis of bearing fault categories; Preprocessing the known bearing-like monitored data to obtain a closed dataset and its class labels, including the following steps: Dividing the known bearing-like monitored data according to the bearing health status categories to obtain several monitored data samples and the class labels corresponding to the monitored data samples, and performing short-time Fourier transform on each monitored data sample to obtain the time-frequency spectrum of the monitored data sample, and all the monitored data samples form a closed dataset; The flow model is used to reconstruct approximate samples of real samples, including the following steps: The flow model adopts the structure of an autoencoder and fits the mapping relationship between the actual data set and the simple prior distribution through a series of reversible transformations. Assume that f(·) is the mapping function of the flow model, and f --1 (·) is its inverse function. During inference, monitor data samples x are randomly sampled from the closed data set and mapped to vectors z of a simple known distribution through f --1 (·). During generation, vectors z are randomly sampled from the known distribution, and approximate samples are generated through f(·) Training the flow discriminant model based on the closed dataset and its class labels to obtain a trained flow discriminant model, including the following steps: Input the high-dimensional features of the monitored data samples in the extracted closed dataset into the discriminator, and calculate the loss between the output result and 1; Randomly sample a vector z from a known distribution and obtain an approximate sample through f(z). Input the approximate sample into the trained fault diagnosis model to extract the high-dimensional features of the approximate sample, and then input the high-dimensional features of the approximate sample into the discriminator. Calculate the loss by comparing the output result with 0; and Alternately train the discriminator and the flow model, record the loss of the discriminator each time, save the parameters of the flow discriminant model when the loss is the lowest, and obtain a trained flow discriminant model.

Images