CN114609994B - Fault diagnosis method and device based on multi-granularity regularized rebalancing incremental learning - Google Patents

Abstract

The invention discloses a fault diagnosis method and device based on multi-granularity regularization rebalancing increment learning, wherein the method comprises the following steps: constructing a continuous label with multi-granularity information, and optimizing by utilizing KL divergence loss; the feature extraction layer is used for obtaining feature expression vectors of new and old categories, decision output of the current model is constrained to be identical to output distribution of the model before incremental learning through knowledge distillation, relatively low weight is applied to the new categories with a large number of samples based on a multi-granularity regularization term, gap between new fault categories and gradient update of the old fault categories is balanced, and category imbalance is relieved; and adopting a two-stage training strategy, performing a first-stage training by using the data which can be acquired currently, updating the feature extraction layer, decoupling with the feature extraction layer during a second-stage training, namely freezing parameters of the feature extraction layer, and retraining the classifier by adopting a resampled balance training subset. The device comprises: a processor and a memory. The invention improves the continuous learning capability of the fault diagnosis model, thereby improving the fault recognition capability of the device.

Description

Fault diagnosis method and device based on multi-granularity regularized rebalanced incremental learning

Technical Field

The present invention relates to the field of intelligent fault diagnosis of machine learning, and in particular to a fault diagnosis method and device based on multi-granularity regularized rebalanced incremental learning.

Background Art

With the rapid development of contemporary science and technology, large-scale equipment is constantly being innovated and widely used in various fields such as transportation, energy, electricity, and aerospace. Its safe, stable, reliable and efficient operation is closely related to national development and national defense construction. At the same time, as the operation time goes by, major equipment will inevitably fail, resulting in catastrophic accidents from time to time; and the complexity and precision of large-scale equipment make it difficult to detect its failures through simple detection methods. Therefore, how to design intelligent fault diagnosis research for large-scale equipment has become an urgent problem facing industrial intelligence.

Existing intelligent fault diagnosis methods collect sensor information from equipment at different times, design deep neural networks based on the collected data, and then deploy them after training. They predict the newly collected sensor data to determine whether the equipment is currently in a normal state or a certain fault state. Although this method can achieve good performance, the intelligent fault diagnosis model cannot be updated once deployed. Generally, the fault type of equipment will gradually change with the increase of service life, and faults that have not occurred before will follow one after another. Existing methods are difficult to use in practice because they can only identify old types of faults that have been learned.

Incremental learning is a way to update the model to master new tasks, that is, the model learns new tasks based on previous tasks, so as to achieve the ability to handle new and old tasks at the same time. Generally, while learning a large amount of new task data, the incremental learning model can retain a small amount of representative data of old tasks to preserve the knowledge of old tasks. However, since the amount of data for new tasks is much higher than that for old tasks, the deep neural network model changes important parameters related to the old tasks when learning new tasks, resulting in serious deviations in its recognition of new and old tasks, that is, the parameters are biased towards learning new tasks, which greatly reduces the recognition ability of old tasks.

To address this problem, existing methods have designed some deviation correction methods, such as adding linear correction layers, normalizing the output cosine of new and old categories, etc. However, they rely heavily on the assumption of the deviation relationship between the new and old categories. In practical applications, this assumption is difficult to hold, which will damage the performance of the model and make it difficult to adapt to complex practical application scenarios. Therefore, how to solve the problem of forgetting old tasks in incremental learning is the key to whether the intelligent fault diagnosis model can be effectively applied on a large scale.

Summary of the invention

The present invention provides a fault diagnosis method and device based on multi-granularity regularized rebalancing incremental learning. Considering the problem that it is difficult to obtain a priori assumptions about the deviation distribution of new and old types of faults in incremental learning of fault diagnosis tasks, the present invention designs a multi-granularity regularized rebalancing method. The constraint model can simultaneously learn the balanced data distribution and the correlation between different fault categories, so that the model can reduce the knowledge forgetting of the learned old category faults when accurately learning to identify new faults, thereby improving the continuous learning ability of the fault diagnosis model, and then improving the fault identification ability of large equipment, greatly increasing the safety of the equipment, and reducing the failure rate. See the following description for details:

In a first aspect, a fault diagnosis method based on multi-granularity regularized rebalanced incremental learning is provided, the method comprising:

The divided data set is represented by category word vectors, the word vectors corresponding to its semantic labels are obtained, and the K-means algorithm is used for clustering to obtain a two-layer multi-granularity structure; continuous labels with multi-granularity information are constructed and optimized using KL divergence loss;

The feature representation vectors of new and old categories are obtained through the feature extraction layer. The decision output of the current model is constrained to be the same as the output distribution of the model before incremental learning through knowledge distillation. Based on the multi-granularity regularization term, a relatively low weight is applied to new fault categories with a large number of samples to balance the gap between the gradient updates of new fault categories and old fault categories, thereby reducing category imbalance.

A two-stage training strategy is adopted. The currently available data is used for the first stage training to update the feature extraction layer. In the second stage training, the classifier is decoupled from the feature extraction layer, that is, the parameters of the feature extraction layer are frozen, and the classifier is retrained using a resampled balanced training subset.

Among them, the first stage training process is: mix the new category samples D _new with the old category example set D′ _old , denoted as D _t as the input of the model, and obtain the output of the network through the feature extraction layer and classifier. The optimization goal is to minimize the weighted sum of the distillation loss and the rebalancing module loss of multi-granularity regularization.

Furthermore, the second stage training process is: sampling the data into category-balanced training subsets, freezing the parameters of the feature extraction layer, and retraining the classifier separately; the optimization goal is to minimize the weighted sum of the distillation loss and the multi-granularity regularized rebalancing module loss.

In a second aspect, a fault diagnosis device based on multi-granularity regularized rebalanced incremental learning is provided, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to enable the device to execute any one of the method steps described in the first aspect.

In a third aspect, a computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes any one of the method steps described in the first aspect.

The beneficial effects of the technical solution provided by the present invention are:

1. The present invention simultaneously considers the correlation between rebalance modeling and fault categories, further improves the learning ability of new and old categories, reduces the catastrophic forgetting phenomenon of the model, improves the recognition ability of fault diagnosis, and further improves the fault recognition ability of large equipment, greatly increases the safety of equipment, and reduces the failure rate of equipment;

2. This paper proposes a new multi-granularity regularized rebalancing method for intelligent fault diagnosis tasks of large equipment in industrial scenarios. This method does not need to assume the deviation relationship between new and old categories in advance, and improves the catastrophic forgetting problem of old tasks in incremental learning of intelligent fault diagnosis models;

3. The present invention designs a multi-granularity regularized rebalancing module, which constrains the impact of different fault categories on model updating according to the number of samples, and embeds the class correlation based on the hierarchy into the learning process by constructing a hierarchy of fault classes;

4. The present invention can achieve state-of-the-art performance in incremental learning tasks such as fault diagnosis, with average precision improvements of up to 7.47% and 9.99% compared to the comparative methods.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a fault diagnosis method based on multi-granularity regularized rebalanced incremental learning;

FIG2 is a schematic diagram of a multi-granularity structure construction of a large fault diagnosis data set FASON;

Figure 3 is a comparative schematic diagram;

Among them, this figure is applied to a large fault diagnosis data set. Compared with the rebalancing method alone, the accuracy of all categories is improved by using multi-granularity regularization. It can be seen that the accuracy of most categories is significantly improved after using multi-granularity regularization;

FIG4 is a schematic diagram of the structure of a fault diagnosis device based on multi-granularity regularized rebalanced incremental learning;

Table 1 shows the comparison of the accuracy and average accuracy of the last stage of different methods on the public dataset;

Table 2 shows the comparison of the accuracy and average accuracy of the last stage of different methods on the fault diagnosis dataset collected in real scenarios;

Table 3 shows the ablation experiments in the FARON-22/22 setting, demonstrating the effectiveness of each component of the proposed method.

DETAILED DESCRIPTION

In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

Example 1

Since only a small number of samples of the old fault classes that have been learned are saved, and a large number of samples of the new fault classes are learned, in order to solve the problem existing in the background technology, incremental learning of new fault categories in fault diagnosis tasks will lead to serious degradation of the recognition ability of old fault categories. The embodiment of the present invention proposes a fault diagnosis method based on multi-granularity regularized rebalanced incremental learning, which incrementally learns new faults in fault diagnosis while maintaining the recognition ability of old fault categories in the past, thereby quickly and efficiently updating and continuously learning the fault diagnosis model. Referring to FIG1, the method includes the following steps:

101: rebalancing loss;

Among them, the rebalancing loss learns the updated weights of the categories based on the number of samples of the new and old fault categories, and maintains its recognition ability by increasing the gradient update of the old fault categories with sparse samples.

102: Multi-granularity regularization term;

Among them, since only using rebalancing loss will affect the learning of new types of faults, a multi-granularity regularization term is proposed to use the hierarchical structure information of fault categories to constrain the model to learn the relationship between fault categories, thereby improving the recognition ability of new and old classes.

103: Train the classifier.

Among them, a two-stage training strategy is adopted. The currently available data is used for the first stage training to update the feature extraction layer. In the second stage training, it is decoupled from the feature extraction layer, that is, the parameters of the feature extraction layer are frozen, and the classifier is retrained using a resampled balanced training subset.

In summary, the embodiment of the present invention realizes simultaneous consideration of the correlation between rebalancing modeling and fault categories through the above steps 101 to 103, further improving the learning ability of new and old categories, alleviating the catastrophic forgetting phenomenon of the model, improving the recognition ability of fault diagnosis, greatly increasing the safety of the equipment, and reducing the failure rate of the equipment.

Example 2

The scheme in Example 1 is further introduced below in conjunction with FIG. 1 , calculation formulas, and tables. See the following description for details:

1. Model Framework

The model framework consists of a feature extraction module, knowledge distillation, multi-granularity regularized rebalancing modeling, and a classifier, and is divided into two stages of training, as shown in Figure 1. The following two components are the core of the method:

(1) Rebalanced modeling: Use a rebalancing strategy during training to mitigate the impact of data imbalance;

The above-mentioned rebalancing strategy is well known to those skilled in the art, and will not be elaborated in detail in the embodiment of the present invention.

(2) Multi-granularity regularization: Since only using the rebalancing strategy will lead to a decrease in the classification accuracy of new categories (i.e., underfitting of new categories), the embodiment of the present invention designs a multi-granularity regularization term so that the model takes class correlation into account.

2. Data Partitioning, Sampling, and Building Multi-granularity Structures

1. Dataset division

In the embodiment of the present invention, the categories that have been learned in the current state of the model are called old categories, and the categories that have not been learned and are about to be learned in this stage are called new categories. For incremental learning fault diagnosis tasks, it is necessary to divide different batches of categories on the experimental data set to simulate the arrival of new categories. In order to ensure the fairness of the experiment, the category learning order is uniformly set according to the comparison methods such as iCaRL, BiC, IL2M (well known to those skilled in the art), and the random number seed is set to 1993. The division is marked with n/m, where n represents the number of old categories learned in the initial stage, and m represents the number of new categories that need to be learned in each subsequent round until all categories of the data set are learned. The CIFAR10, CIFAR100, MiniImageNet data sets and FARON fault diagnosis data sets are divided, and the specific details are as follows:

(1) The CIFAR10 dataset has 10 categories, each of which contains 5000 samples for training and 1000 samples for testing. The dataset is divided into two settings: 1/1 and 2/2, that is, the 10 categories are evenly divided into 1-category and 2-category incremental batches.

(2) The CIFAR100 dataset has 100 categories, each of which contains 500 samples for training and 100 samples for testing. 50 of the 100 categories are used as the initial learning categories, and every 5 or 10 categories of the remaining 50 categories are divided into an incremental learning batch, which is recorded as a 50/5 or 50/10 division mode.

(3) The MiniImageNet dataset contains 100 categories, each with 600 samples. The embodiment of the present invention divides the 600 samples of each category into a training set and a test set at a ratio of 9:1. The experiment uses a 10/10 and 20/20 division mode.

(4) FARON fault diagnosis data set. The data set is large-scale equipment fault diagnosis data. There are 121 sensor output dimensions in the data, that is, the current operating status of the system is described from 121 different aspects. In order to carry out effective fault diagnosis modeling, the data used must include normal operating data and fault data under different working conditions. After 5.32 hours of normal operating data, a total of 76,632 signal samples were collected. The data set has 66 categories, including 1 normal category and 55 fault categories. The random seed of the learning order of this data set is set to 36 to ensure that the normal category is learned in the initial stage. The experiment uses 11/11, 22/22 and 33/33 partitioning modes.

2. Sampling strategy

The experiment uses a random sampling strategy and assumes that the maximum storage sample space is fixed, with K representing the maximum number of stored samples. For the public dataset, the maximum number of stored samples K is set to 2000. The maximum number of stored samples K for the fault diagnosis dataset is 8000.

Due to the limited storage space, the number of samples in the training set of the old category is different from that in the training set of the new category, but the number of samples in the validation set of the old category is the same as that in the validation set of the new category, which is used to balance the training phase. The sum of the samples in the training set and the test set of the old category must be less than or equal to the upper limit K of the storage space. Specifically, the embodiment of the present invention sets the sample ratio of the training set of the old category to the validation set to 9:1, and the sample ratio of the training set of the old category is 9:1. Old category validation set And the new category validation set New category training set The sample size is given by the following formula:

3. Construct a multi-granular structure for the divided data set

GloVe (Global Vectors for Word Representation) is a commonly used word vector representation method based on global word frequency statistics, which is achieved by mapping words to a high-dimensional space. Using a pre-trained word vector library, each word can be converted into a 300-dimensional word vector. After obtaining the representation of the words in the feature space, the embodiment of the present invention uses a clustering algorithm to cluster the semantic vectors. This method can be used to build the currently appearing categories into a two-layer hierarchical structure. The embodiment of the present invention uses the K-means algorithm to cluster the word vectors converted from the labels.

Among them, the optimization formula of the K-means algorithm is shown as follows:

Where x = (x ₁ , x ₂ , ..., x _n ), represents the feature representation of each point. μ _i is the mean of all points in cluster S _i . The optimization goal of the algorithm is to find a cluster S _i that satisfies the above formula and divide n points into k clusters so that the sum of square errors of all points in the cluster is minimized. The smaller the parameter k is, the fewer superclasses are clustered and the weaker the correlation between categories under the same superclass is; the larger the parameter k is, the more superclasses are clustered and the stronger the correlation between categories belonging to the same superclass is.

The algorithm flow is to first import the GloVe pre-trained model to obtain the word vector corresponding to the semantic label. Then use the K-means algorithm to cluster to obtain a two-layer hierarchical structure.

3. Feature Extraction Module

The embodiment of the present invention uses a residual network (ResNet) as a feature extractor. For CIFAR and MiniImageNet datasets, ResNet32 and ResNet18 networks are used to extract features respectively. For large fault diagnosis datasets, CNN networks are used to extract features.

Among them, ResNet32, ResNet18 and CNN networks are well known to those skilled in the art, and will not be described in detail in the embodiments of the present invention.

4. Knowledge Distillation Module

The model of the N _old class learned in the previous round is used as the teacher model, and the model of the N _old + N _new class to be learned in this round is called the student model. The input of the student model is the union of the old class samples and the new class samples after sampling, symbolically represented by D _t =D′ _old ∪D _new , and the corresponding output is The same data will be output after passing through the teacher model Before passing through the softmax layer, each probability item output by the teacher model is divided by the temperature coefficient T, and the result is recorded as the soft label π(z′). Each probability item output by the student model is divided by the temperature T, recorded as π(z). Then, the predicted probability of the previous N _old class is intercepted from π(z) and the distillation loss is calculated. The distillation loss function is as follows:

The above “'” indicates the sampled ones, for example, D _o ′ _ld indicates the old category samples after sampling. T is the distillation temperature. The higher the temperature T is set, the more uniform the probability distribution will be. When T is set to 1, the distillation loss function is equivalent to the cross entropy loss function. In the embodiment of the present invention, T is set to 2.

The following Algorithm 1 summarizes the working process of applying knowledge distillation technology to this invention.

5. Multi-granularity regularized rebalancing module

The rebalancing module with multi-granularity regularization consists of two components. The first is rebalancing modeling. A rebalancing strategy is used during training to mitigate the impact of data imbalance. The second is multi-granularity regularization. Since using only the rebalancing strategy will lead to a decrease in the classification accuracy of new categories, a multi-granularity regularization term is designed to make the model consider class correlation.

Rebalancing modeling is performed by applying relatively high weights to categories with sparse samples and relatively low weights to categories with a large number of samples, balancing the gap between the gradient updates of new categories and old categories and reducing category imbalance. The category weights are as follows:

Among them, _zi represents the output of the network, _ni represents the number of samples in each category, and γ represents a hyperparameter, which is defined as:

Where N represents the total number of categories, _zj represents the network output, and log(·) represents the predicted probability that the sample belongs to category j. The rebalancing loss is expressed as:

The embodiment of the present invention uses the rebalancing loss to improve the supervision strength of the old categories, so that the negative impact caused by the significant difference in the number of samples between the new and old categories is further weakened. In addition to using the rebalancing method to constrain the impact of the number of samples of different fault categories on model updates, the embodiment of the present invention further designs a multi-granularity regularization term, which embeds the class correlation based on the hierarchy into the learning process by constructing a hierarchy of fault classes.

First, we use the dataset or semantic clustering to obtain the class hierarchy. We use its semantic information to convert the semantic labels into word vectors through GloVe, and use the K-means algorithm for semantic clustering to obtain a two-level structure. Then, we construct continuous labels with multi-granularity information and optimize them using KL divergence loss. The specific calculation method is as follows:

The calculation method for constructing continuous labels with multi-granularity information is:

Among them, C represents the node of the true label, A represents the current class node, M represents the multi-granularity label, and N represents the leaf node set. β represents a hyperparameter of the function, and its value range is (0, +∞), which controls the distribution of the label. The smaller the value of β, the more uniform the label distribution; the larger the value of β, the sharper the label distribution. The formula d(N _i ,N _j ) measures the distance between two fine-grained categories in the hierarchy. LCS (least common subtree) represents the smallest common subtree between two nodes. Height(B) represents the height of the subtree whose root is node B. The continuous label with multi-granularity information is recorded as

The calculation method of the multi-granularity regularization loss function is:

That is, using KL divergence calculation and the difference between the predicted probability q _i .

The loss function of the rebalancing module with multi-granularity regularization is:

_LM ＝(1-λ) _LCB + _LH (13)

in,

The following Algorithm 2 summarizes the rebalancing module with multi-granularity regularization.

6. Classifier

The classifier is a one-layer fully connected network with the number of output neurons equal to the number of all categories seen in this stage and the previous stage, and the corresponding number of neurons is increased as incremental learning proceeds.

7. Training Strategy

The model is trained in two stages. In the first stage, the training process is to mix the new category samples D _new with the old category example set D _o ′ _ld , recorded as D _t as the input of the model. After the feature extraction layer and the classifier, the output of the network is obtained. The optimization goal is to minimize the weighted sum of the distillation loss and the rebalancing module loss of multi-granularity regularization.

The total loss function formula is shown in (14):

L＝L _M +αλL _D (14)

in, Used to adjust the degree of retention of old knowledge. As incremental learning proceeds, the value of parameter λ gradually increases, and the retention of old category knowledge gradually strengthens. α＝10 ^-x , used to adjust the loss _LD to the same order of magnitude as _LM , and x is generally assigned 0 or 1.

After the training is completed with imbalanced data, the second stage of training is to add an additional step to eliminate the classifier bias, that is, decoupling the CNN layer from the classifier, freezing the parameters of the CNN layer, and retraining the classifier. This is done as an additional step after the training is completed with the imbalanced training set. After the balanced validation set is built, the parameters of the CNN layer are frozen before training. The classifier is retrained using the validation set during training. In particular, this step will be performed in all stages except the initial stage (after training with imbalanced samples) because the samples in the initial stage are balanced.

In summary, the embodiments of the present invention realize the simultaneous consideration of the correlation between rebalancing modeling and fault categories through the above parts, further improve the learning ability of new and old categories, reduce the catastrophic forgetting phenomenon of the model, improve the recognition ability of fault diagnosis, greatly increase the safety of the equipment, and reduce the failure rate of the equipment.

Example 3

In conjunction with FIG. 2 and FIG. 3 , specific experimental data are used to verify the feasibility of the schemes in Examples 1 and 2, as described below:

First, we use a public dataset to verify the method. The hierarchical structure of the dataset is gradually constructed by semantic clustering.

For CIFAR10 and CIFAR100, the learning rate starts from 0.01 and is divided by 10 at the 70th and 140th iterations. For MiniImageNet, the batch size is set to 128. The learning rate starts from 0.01 and is divided by 0.1 after every 30 iterations. The SGD optimizer is used with a weight decay of 2e-4 and a momentum of 0.9. The training epochs for CIFAR10, CIFAR100, and MiniImageNet are set to 200, 200, and 100, respectively. The data batch size is set to 128, and the Adam optimizer is used. For the memory limit K, K = 2,000 is set on all public datasets.

Secondly, the method is verified using a fault diagnosis dataset. Figure 2 is a schematic diagram of the multi-granularity structure construction of a large fault diagnosis dataset. On the right side of Figure 2, it can be seen that the features of the categories numbered 1, 5, 27, and 28 are similar, and they belong to the coarse-grained category of water pump faults. Taking the FARON-22/22 experimental setting as an example, the fine-grained fault category numbered 5 is learned in the initial stage, and the categories numbered 1, 27, and 28 are learned in the last round of incremental learning. Through multi-granularity regularization, the structural information of the data can be added to the incremental learning to maintain the correlation between fault categories.

The random seed of the learning sequence of the data set in the embodiment of the present invention is set to 36 to ensure that the normal category is learned in the initial stage. For this data set, the embodiment of the present invention uses the Adam optimizer and sets the learning rate to 1e-5 for training 50 epochs. The memory limit K is set to 8,000.

The experiments in the embodiments of the present invention follow the standard evaluation system, and evaluate the model's ability to retain old knowledge and learn new knowledge through incremental accuracy and average accuracy. The average accuracy measures the average accuracy of each stage of incremental learning, including the initial stage. The higher the average incremental accuracy and the average accuracy, the stronger the model's ability to learn new category samples and retain old category knowledge. The experimental results are compared with six existing methods, including: iCaRL, LwF, LUCIR, MT, BiC, and PODNet.

Table 1 shows the accuracy and average accuracy of the last incremental stage on the CIFAR10, CIFAR100 and MiniImageNet datasets, which comprehensively reflects the performance of the model provided by the embodiment of the present invention. Table 2 shows the accuracy and average accuracy of each incremental stage on the actual fault diagnosis dataset, which reflects the performance of the model provided by the embodiment of the present invention under the actual complex scenario FARON. The embodiment of the present invention takes into account the imbalanced nature of the categories, utilizes the correlation between the categories, and imposes constraints on the model, thereby improving the model performance.

Table 1 Incremental learning results (accuracy) on three public datasets.

Among them, last represents the accuracy of the last stage, and Avg represents the average incremental accuracy.

Table 2 Incremental learning results (accuracy) of fault diagnosis data tasks

In addition, the embodiment of the present invention verifies the effectiveness of its own components, as shown in Table 3. FIG3 visualizes the accuracy improvement of each category using the rebalancing and multi-granularity regularization method (MGRB) and the rebalancing method (RB) alone.

Table 3 Ablation experiment results of FARON large fault diagnosis dataset

Among them, CE is the abbreviation of cross entropy loss, CB is the class balance classification loss part in the rebalancing module of multi-granularity regularization. MG represents the multi-granularity regularization part in the rebalancing module of multi-granularity regularization. Decoupling indicates whether there is a second stage of training.

Example 4

A device for capturing dynamic key points of a video, see FIG3 , the device comprises: a processor 1 and a memory 2, the memory 2 stores program instructions, the processor 1 calls the program instructions stored in the memory 2 to enable the device to execute the following method steps in Example 1:

A fault diagnosis method based on multi-granularity regularized rebalanced incremental learning, the method comprising:

The divided data set is represented by word vectors, the word vectors corresponding to the semantic labels are obtained, and the K-means algorithm is used for clustering to obtain a two-layer multi-granularity structure; continuous labels with multi-granularity information are constructed and optimized using KL divergence loss;

The feature extraction layer obtains feature representation vectors of new and old categories. Through knowledge distillation, the decision output of the current model is constrained to be the same as the output distribution of the model before incremental learning. Based on multi-granularity regularization terms, relatively low weights are applied to categories with a large number of samples to balance the gap between the gradient updates of new fault categories and old fault categories, thereby reducing category imbalance.

A two-stage training strategy is adopted. The currently available data is used for the first stage training. In the second stage training, it is decoupled from the feature extraction layer, that is, the parameters of the feature extraction layer are frozen, and the classifier is retrained using a resampled balanced training subset.

The construction of continuous labels with multi-granularity information is specifically as follows:

Where C represents the node of the true label, A represents the current class node, M represents the multi-granularity label, N represents the leaf node set, β represents a hyperparameter of the function, the formula d(N _i ,N _j ) measures the distance between two fine-grained categories in the hierarchy, LCS represents the minimum common subtree containing two nodes, Height(B) represents the height of the subtree with the root as node B, and the continuous label with multi-granularity information is recorded as

Furthermore, the loss function of the multi-granularity regularization term is calculated as:

The loss function of the rebalanced multi-granularity regularization is:

_LM ＝(1-λ) _LCB + _LH

The total loss function is:

L＝ _LM + _αλLD

in, α=10 ^-x , is used to adjust the loss _LD to the same order of magnitude as _LM .

The first stage training process is as follows: the new category samples D _new are mixed with the old category example set D _o ′ _ld , denoted as D _t as the input of the model, and the network output is obtained through the feature extraction layer and classifier. The optimization goal is to minimize the weighted sum of the distillation loss and the rebalancing module loss of multi-granularity regularization.

Furthermore, the second stage training process is: sampling the data into category-balanced training subsets, freezing the parameters of the feature extraction layer, and retraining the classifier separately; the optimization goal is to minimize the weighted sum of the distillation loss and the rebalancing module loss of multi-granularity regularization.

It should be pointed out here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

The execution subjects of the above-mentioned processor 1 and memory 2 can be devices with computing functions such as computers, single-chip microcomputers, and microcontrollers. In specific implementation, the embodiment of the present invention does not limit the execution subject and it is selected according to the needs of actual applications.

The data signal is transmitted between the memory 2 and the processor 1 via the bus 3, which will not be described in detail in the embodiment of the present invention.

Example 5

Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, the storage medium includes a stored program, and when the program is running, the device where the storage medium is located is controlled to execute the method steps in the above embodiment.

The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

It should be pointed out here that the description of the readable storage medium in the above embodiment corresponds to the description of the method in the embodiment, and the embodiment of the present invention will not be described in detail here.

In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present invention are generated.

The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions may be stored in a computer-readable storage medium or transmitted via a computer-readable storage medium. The computer-readable storage medium may be any available medium that can be accessed by the computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium or a semiconductor medium, etc.

Unless otherwise specified, the models of the components in the embodiments of the present invention are not limited, and any device that can perform the above functions may be used.

Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (3)

1. A fault diagnosis method based on multi-granularity regularized rebalancing incremental learning, characterized in that the method includes: Perform word vector representation on the divided fault diagnosis data set to obtain the word vectors corresponding to the semantic labels, and use the K-means algorithm to cluster the word vectors converted from the semantic labels to obtain a two-layer multi-granularity structure; Construct continuous labels with multi-granularity information and optimize using KL divergence loss; The feature representation vectors of the new and old fault categories are obtained through the feature extraction layer. The fault categories learned in the current state of the model are called old fault categories, and the fault categories that have not been learned and are about to be learned at this stage are called New fault categories; The decision output of the current model is constrained by knowledge distillation to be the same as the output distribution of the model before incremental learning. Based on multi-granularity regularization terms, relatively low weights are applied to categories with a large number of samples to balance the gradient update of new fault categories and old fault categories. gap; Adopt a two-stage training strategy, use the currently acquired data for the first stage training, update the feature extraction layer, and decouple the classifier from the feature extraction layer during the second stage training, that is, freeze the parameters of the feature extraction layer and use resampling Retrain the classifier on the balanced training subset; The construction of continuous labels with multi-granularity information is specifically: Among them, C represents the real label node, A represents the current class node, M represents the multi-granularity label, N represents the leaf node set, β represents a hyperparameter of the function, and the formula d(N _i ,N _j ) measures the hierarchical structure The distance between two fine-grained categories, LCS represents the minimum common subtree containing two nodes, Height(B) represents the height of the subtree whose root is node B, and continuous labels with multi-granularity information are recorded as The calculation method of the loss function of the multi-granularity regularization term is: The loss function of multi-granularity regularization rebalancing is: L _M =(1-λ)L _CB +L _H The rebalancing loss is expressed as: Among them, w _i represents the category weight, z _i represents the output of the network, N represents the total number of categories, and z _j represents the network output; The total loss function is: L＝L _M +αλL _D in, α=10 ^-x , used to adjust the distillation loss _LD to the same order of magnitude as _LM ; Among them, the model that learned the N _old category in the previous round is used as the teacher model, and the model that needs to learn the N _old + N _new category in this round is called the student model; the input of the student model is the sampled old category sample and the new category sample. The union of , the symbol is expressed as D _t =D _o ′ _ld ∪D _new , and the corresponding output is After the same data passes through the teacher model, the output Before passing through the softmax layer, divide each probability item output by the teacher model by the temperature coefficient T, and record this result as a soft label π(z′). Divide each probability item output by the student model by the temperature T, record as π(z); the first-stage training process is: mix the new fault category sample D _new with the old fault category example set D _o ′ _ld , denoted as D _t as the input of the model, through the feature extraction layer and classifier Obtain the output of the network, and the optimization goal is to minimize the weighted sum of distillation loss and multi-granularity regularization rebalancing module loss; The second-stage training process is: sampling the data into a category-balanced training subset, freezing the parameters of the feature extraction layer, and retraining the classifier separately; the optimization goal is to minimize the distillation loss and the rebalancing module loss of multi-granularity regularization weighted sum. 2. A fault diagnosis device based on multi-granularity regularized rebalancing incremental learning, characterized in that the device includes: a processor and a memory, program instructions are stored in the memory, and the processor calls the memory stored in the memory. program instructions to cause the apparatus to perform the method steps of claim 1. 3. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, the computer program includes program instructions, and when executed by a processor, the program instructions cause the processor to execute the right The method steps described in claim 1.