CN116026590A - Self-supervision bearing fault diagnosis method - Google Patents

Abstract

The invention provides a self-supervision bearing fault diagnosis method, which comprises the following steps: combining contrast learning and clustering algorithm, performing bearing fault classification based on contrast clustering task in self-supervision, and performing pre-training by utilizing a pre-training data set to obtain a pre-training model; retraining the fine tuning of the pre-trained model to obtain a depth model, and putting the fine tuning model into a real industrial environment for use.

Description

Self-supervision bearing fault diagnosis method

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a self-supervision bearing fault diagnosis method.

Background

Conventional intelligent fault diagnosis methods typically take as the end result what type of fault is well-defined and are typically a single type of fault. In an actual industrial scenario, the device operates normally in most cases, and multiple faults may occur simultaneously when an error occurs. Thus, while failure data is difficult to obtain, it is more difficult to define what failure category the final result belongs to.

Based on the self-supervised learning idea, the model is made to learn itself using information provided by the data points themselves in the data set acquired from the vibration signal. After model pre-training, fine tuning is performed on the model according to the real industrial environment data set of the high-quality small sample, so that the final model can have an effective diagnosis capability. Because the model is used for self-learning of the data, the whole model does not need the participation of manual labels in the pre-training stage, and a large amount of labor cost is reduced. At present, the common self-learning scheme aiming at bearing fault diagnosis mainly has the problem that data cannot well reflect actual faults.

Disclosure of Invention

The invention aims to provide a self-supervision bearing fault diagnosis method, which aims to solve the problem that the existing bearing fault diagnosis self-learning data cannot well reflect actual faults.

In order to solve the technical problems, the invention provides a self-supervision bearing fault diagnosis method, which comprises the following steps:

performing bearing fault classification on the self-supervision clustering task based on comparison by using a comparison learning and clustering algorithm, and performing pre-training by using a pre-training data set to obtain a pre-training model; and

retraining the fine tuning of the pre-trained model to obtain a depth model, and putting the depth model into a real industrial environment for use.

Optionally, in the self-supervised bearing fault diagnosis method, the comparison learning includes:

the method comprises the steps of directly classifying source data extraction features through deep learning, changing direct classification into the use of twin network extraction features, comparing different features obtained by the twin network by defining positive and negative samples so as to distinguish similar data from different data, and distinguishing the data of different categories by oneself.

Optionally, in the self-supervised bearing fault diagnosis method, the comparison learning based on the clustering algorithm includes:

the twin network comprises two identical deep neural networks through comparison and learning of the architecture twin network;

the deep neural network selects ResNet50, and gradient explosion and gradient disappearance of model training are avoided by using residual blocks;

the output of the deep neural network is directly provided to a prototype layer, and the output of the prototype layer is used for clustering algorithm calculation and cross entropy loss calculation.

Optionally, in the self-supervised bearing fault diagnosis method, the method further includes:

carrying out different data enhancement processing on vibration signal source data x to obtain data sets x1 and x2 respectively, and providing the data sets x1 and x2 into two models of a twin network respectively;

the output of the ResNet50 classifier is provided to the prototype layer computation and outputs data sets y1 and y2, respectively;

the data sets y1 and y2 are calculated through a clustering algorithm to obtain data sets label1 and label2;

the output of the prototype layer of one model and the pseudo tag calculated by the other model calculate cross entropy losses through an objective function, respectively calculate the cross entropy losses of y1 and label2 and y2 and label1, and optimize the parameters of ResNet50 through the cross entropy losses.

Optionally, in the self-supervised bearing fault diagnosis method, the proxy task of contrast learning using unlabeled data replaces the real label;

in contrast learning, after different data enhancement processing is carried out on vibration signal source data, input with different forms and unchanged semantics is obtained;

the agent task uses a twin network, vibration signal source data is subjected to different data enhancement processes, so that different input forms are provided to the twin network, and different outputs are obtained through the twin network;

defining positive and negative samples through proxy tasks, and calculating loss through an objective function according to the difference between the positive and negative samples so as to optimize model parameters;

using a contrast clustering algorithm as an agent task, and using a clustering center generated by the clustering algorithm to replace a negative sample in contrast learning;

and alternately comparing the output of one model in the twin network with the pseudo labels obtained by the other model through a clustering algorithm to calculate the loss of the whole twin network and optimize the whole model.

Optionally, in the self-monitoring bearing fault diagnosis method,

the depth neural network used in the twin network is a 50-layer residual network, and the main framework of the depth neural network comprises an encoder and a classifier;

introducing a residual block into the encoder, wherein in the final output H (x) of the residual block, the output F (x) obtained by calculating the input x by a convolution layer and the input x themselves participate in the calculation of the final output;

the residual block retains the original input information during output, so that the residual block retains the original characteristics of the input data in a high-dimensional characteristic space;

the prototype layer is added after ResNet50 to learn as a cluster center in the clustering algorithm.

Optionally, in the self-supervised bearing fault diagnosis method, the pre-training stage includes:

collecting 500 sets of signals from each state of the collected vibration signals as a source data set, wherein each set of signals has a length of 2048 data points;

carrying out random data enhancement on the acquired data respectively to obtain data sets X1 and X2 respectively, and inputting the data sets into a twin network;

wherein the data enhancement method comprises adding random gaussian noise, performing random masking, and performing random variation on the signal amplitude.

The overall structure of the ResNet encoder includes:

the first part trains the input x of 1 x 2048 to 64 x 512, and the subsequent parts continue to extract features from the data;

the output of the second fraction is 256 x 512;

the output of the third fraction is 512 x 256;

the output of the fourth part is 1024×128; and

the output of the fifth part is 2048×64;

in the sixth part, the output obtained before is subjected to an average pooling layer, and the ResNet encoder finally outputs the characteristic with the length of 2048;

the characteristics are calculated through a classifier, and the output length of the whole model is 128 finally;

the input dimension of the prototype layer is the output length 128 of ResNet, and the output dimension of the prototype layer is 4 times of the number of estimated fault categories;

the output of ResNet50 is input directly to the prototype layer, outputting y1 and y2, respectively.

Obtaining respective clustering centers of y1 and y2 through a sink horn-Knopp algorithm respectively, and using the clustering centers as pseudo tags label1 and label2;

calculating information entropy by interaction between the output of the model and the generated pseudo tag, respectively calculating cross entropy by y1 and label2, and calculating cross entropy by y2 and label1, and reversely optimizing parameters of the whole model through calculated loss;

the iteration is repeated until the model converges.

Optionally, in the self-supervised bearing fault diagnosis method, the fine tuning stage includes:

collecting 500 groups of signals from each state of the vibration signals as data sets X ', wherein the length of each group of signals X in the X' is 2048 data points, and inputting the collected data into a deep neural network after data enhancement;

reconstructing a deep neural network, using an encoder of the model of the pre-training stage, and freezing all weights and bias values therein;

redesigning new classifier of the deep neural network according to the number of fault categories contained in the data set X', wherein the output length of the classifier is the same as the number of the fault categories;

inputting the data set into a model, extracting the characteristic with the length of 2048 according to the encoder, inputting the characteristic into a classifier for classification, and obtaining a final output Y;

calculating cross entropy loss by using the output Y of the model and label information Labels, and optimizing parameters of a classifier of the model by using the loss;

the iteration is repeated until the model converges.

Optionally, in the self-monitoring bearing fault diagnosis method,

contrast learning is introduced to compare by defining examples in the samples that are semantically similar to it and examples that are semantically different, and clustering is achieved by designing the model structure and contrast loss so that positive samples are closer together and positive and negative samples are farther apart from each other in the feature space.

The inventor of the invention discovers through research and induction summary that the common self-learning scheme aiming at bearing fault diagnosis mainly has the following problems:

first, the simulation data set used in model training usually uses vibration signals in a fixed environment as data sources. However, the vibration signal is affected by a very wide range of factors, especially the rotational speed of the rotary machine. The same bearing component generates a different vibration signal when the rotational speed of the machine changes.

Secondly, it is generally accepted that vibration signals in the data sets in use are collected by replacing them in the machine with a single faulty component, and therefore the data sets collected therewith are all vibration signals generated by a definite single cause of the fault. However, the actual industrial conditions are different, on one hand, when faults occur, such as cracks, abrasion of outer rings and the like, the severity and the position of the faults are distributed, and each instrument is inconsistent, so that the similar fault causes cannot be directly unified into the same fault type. On the other hand, the composite fault causes are simultaneously present. When the bearing parts are worn to cause faults, two or more faults are likely to occur simultaneously, and vibration signals of the two faults are not collected in the data set which is currently used mainly.

In addition, during the pre-training stage of the current model, the adopted data set is as described before, and the vibration signals are all collected by a clear machine with a single fault cause at a fixed rotating speed, so that the data set used by the current model is clean and clear in category. Based on the above, in the pre-training stage, the existing model mostly uses the accurate number of fault categories as output in the classifier, even if the characteristics output by the encoder are clustered by using a clustering algorithm, the number of the clustering centers is the same as the accurate number of fault categories. However, this situation does not exist in a real industrial environment, and it is basically difficult to fully explore the specific conditions of the machinery and the faults represented by the vibration signals as the data sources.

In the self-supervision bearing fault diagnosis method provided by the invention, a self-supervision bearing fault classification method based on a comparison clustering task is provided, so that under the condition that a data set collected for pre-training in a real industrial scene is mixed and fault categories represented by vibration signals are difficult to distinguish, a high-quality model can be still pre-trained by the data set, and the model can be put into the real industrial environment for use through fine adjustment and re-training. The invention combines the ideas of contrast learning and clustering algorithm, and designs a self-supervision bearing fault classification method based on contrast clustering task.

Drawings

FIG. 1 is a schematic diagram of a self-monitoring bearing fault diagnosis method according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a self-monitoring bearing fault diagnosis method transfer learning according to an embodiment of the present invention;

fig. 3 is a schematic diagram of an encoder-induced residual block of a self-supervised bearing failure diagnostic method in accordance with an embodiment of the present invention.

Detailed Description

The invention is further elucidated below in connection with the embodiments with reference to the drawings.

It should be noted that the components in the figures may be shown exaggerated for illustrative purposes and are not necessarily to scale. In the drawings, identical or functionally identical components are provided with the same reference numerals.

In the present invention, unless specifically indicated otherwise, "disposed on …", "disposed over …" and "disposed over …" do not preclude the presence of an intermediate therebetween. Furthermore, "disposed on or above" … merely indicates the relative positional relationship between the two components, but may also be converted to "disposed under or below" …, and vice versa, under certain circumstances, such as after reversing the product direction.

In the present invention, the embodiments are merely intended to illustrate the scheme of the present invention, and should not be construed as limiting.

In the present invention, the adjectives "a" and "an" do not exclude a scenario of a plurality of elements, unless specifically indicated.

It should also be noted herein that in embodiments of the present invention, only a portion of the components or assemblies may be shown for clarity and simplicity, but those of ordinary skill in the art will appreciate that the components or assemblies may be added as needed for a particular scenario under the teachings of the present invention. In addition, features of different embodiments of the invention may be combined with each other, unless otherwise specified. For example, a feature of the second embodiment may be substituted for a corresponding feature of the first embodiment, or may have the same or similar function, and the resulting embodiment would fall within the disclosure or scope of the disclosure.

It should also be noted herein that, within the scope of the present invention, the terms "identical", "equal" and the like do not mean that the two values are absolutely equal, but rather allow for some reasonable error, that is, the terms also encompass "substantially identical", "substantially equal". By analogy, in the present invention, the term "perpendicular", "parallel" and the like in the table direction also covers the meaning of "substantially perpendicular", "substantially parallel".

The numbers of the steps of the respective methods of the present invention are not limited to the order of execution of the steps of the methods. The method steps may be performed in a different order unless otherwise indicated.

The self-monitoring bearing fault diagnosis method provided by the invention is further described in detail below with reference to the attached drawings and the specific embodiments. The advantages and features of the present invention will become more apparent from the following description. It should be noted that the drawings are in a very simplified form and are all to a non-precise scale, merely for convenience and clarity in aiding in the description of embodiments of the invention.

The invention aims to provide a self-supervision bearing fault diagnosis method, which aims to solve the problem that the existing bearing fault diagnosis self-learning data cannot well reflect actual faults.

In order to achieve the above object, the present invention provides a self-monitoring bearing fault diagnosis method, comprising: combining contrast learning and clustering algorithm, performing bearing fault classification based on contrast clustering task in self-supervision, and performing pre-training by utilizing a pre-training data set to obtain a pre-training model; and retraining the fine tuning of the pre-trained model to obtain a depth model, and putting the fine tuning model into a real industrial environment for use.

Fig. 1 to 3 provide embodiments of the present invention, as shown in fig. 1, the present invention provides a self-supervision bearing fault classification method based on a comparison clustering task, which aims to solve the problem that a high quality model can be pre-trained by using a data set for pre-training under the condition that the data set is collected in a real industrial scene and is more mixed and the fault class represented by vibration signals is difficult to distinguish. Then the model can be put into the real industrial environment for use through fine tuning and retraining. The invention combines the ideas of contrast learning and clustering algorithm, and designs a self-supervision bearing fault classification method based on contrast clustering task.

The large sample data set used in the pre-training phase of the present invention does not require a priori knowledge. The contrast learning and clustering algorithm is to extract and utilize the characteristics of the vibration signals through the information carried by the vibration signals, so as to assist model training.

The invention utilizes the idea of contrast learning, not only directly classifies the extracted features of the source data by deep learning, but also adds a twin network on the basis of the extracted features, and makes the features obtained by the twin network mutually contrast so as to distinguish similar data from different data. This enables the invention to better extract features from the data itself and better self-distinguish between different categories of data.

The invention uses the number of the clustering centers calculated by the clustering algorithm to replace the number of the finally output classifications of the deep neural network. Due to the characteristic of a clustering algorithm, namely, in the input feature space, a clustering center of random positions is generated initially, and all input data are self-classified through continuous iteration. This allows the model to self-classify to generate cluster centers by applying a clustering algorithm. The number of the clustering centers can be several times of the number of the fault categories which are specifically contained, so that the problem that the deep neural network needs to clearly know how many fault categories are in the data set when the deep neural network is pre-trained and classified is avoided.

The invention inputs the output of the self-encoder of the depth model into the classifier to classify, and synchronously calculates a clustering center through a clustering algorithm. And taking the clustering center as a pseudo tag to participate in training of the model again, so that the purpose of self-learning of the model is achieved.

According to the invention, the clustering algorithm is combined into contrast learning, the contrast clustering algorithm is used as a proxy task of the contrast learning, namely, after pseudo labels are generated by the output of two encoders in the twin network, the pseudo labels are respectively calculated with the output of the classifier of the other model, so that the model is trained, and the overall prediction precision of the model is effectively improved.

The contrast learning based on the clustering algorithm is shown in fig. 1, the main framework of the contrast learning is a twin network, the twin network is composed of two identical deep neural networks, and different inputs are trained through the identical neural networks to obtain different outputs. The deep neural network selects ResNet50, which uses residual blocks to alleviate the problems of gradient explosion and gradient disappearance of the whole model training, and can better preserve the original data characteristics in the high-dimensional characteristics. The output of the neural network is directly input into a prototype layer, and the output of the prototype layer is used for clustering algorithm calculation and participates in subsequent cross entropy loss calculation.

The specific flow comprises the following steps: and respectively obtaining x1 and x2 from the vibration signal source data x through different data enhancement, and respectively inputting the vibration signal source data x into two models of the twin network. The output of the ResNet50 classifier is directly input to the prototype layer for further computation and outputs y1, y2, respectively. And obtaining label1 and label2 by the clustering algorithm of y1 and y2. The cross entropy loss is calculated by an objective function by using the direct output of the model and the pseudo tag calculated by another model respectively, namely, the cross entropy loss is calculated by y1 and label2, and y2 and label1 respectively, and then the parameters of ResNet50 are optimized by the loss.

As shown in fig. 2, the present invention also includes a transfer learning that divides the training of the entire model into two parts: a pre-training phase and a fine-tuning phase. The pre-training stage uses a large amount of unlabeled data to train the ResNet50 model by utilizing the information of the data; after the pre-training is completed, the encoder part of the model is migrated into a new model in the fine tuning stage, and the classifier is redesigned; in the fine tuning stage, a small amount of data with specific labels is input into a new model, and classification is carried out through a new classifier under the condition that an encoder is frozen; calculating cross entropy loss by using the known label information and model output, and optimizing parameters of a new model by using the loss; after the loss of the model converges, the trained model can be put into a real industrial scene for application. The invention is divided into two phases in total, namely a pre-training phase and a fine tuning retraining phase, wherein the clustering task-based contrast learning model used in the pre-training phase is used in the invention.

Contrast learning is a self-supervised learning, and its typical paradigm is: proxy task + objective function. The agent task is different from the real Label Label of the data set for supervised learning and the output y of the data set after model processing, and the model parameters are optimized by calculating the loss through the objective function. Whereas contrast learning uses proxy tasks instead, since no Label data, i.e. no real Label, is used.

The proxy task is trained using a twin network, i.e., two identical networks simultaneously. The whole structure is as follows: after the source data is enhanced by different data, two inconsistent inputs are obtained, and then different outputs are obtained through two identical neural networks (twin networks).

The main task of the agent task is to define positive and negative samples, and calculate loss through an objective function according to the difference between the positive and negative samples, so as to optimize model parameters. In the invention, a contrast clustering algorithm is used as an agent task, and a clustering center generated by the clustering algorithm replaces the concept of a negative sample in contrast learning; meanwhile, according to the idea that in contrast learning, after the same source data is enhanced by different data, although two different input modes are obtained, the semantics of the input modes are unchanged, the output of one model in the twin network is alternately compared with the pseudo label obtained by the other model through a clustering algorithm, so that the loss of the whole twin network is calculated, and the whole model is optimized.

The selection of the deep neural network comprises the following steps: the deep neural network used in the twin network is a 50-layer residual network (ResNet 50), the main architecture of which is divided into an encoder part and a classifier part. In the encoder, a residual block is introduced, as shown in fig. 3, in the final output H (x) of the residual block, besides the output F (x) calculated on the input x by the convolution layer, the input x itself also participates in the operation, so that the gradient explosion and gradient disappearance problems generated during the training of the whole training model are effectively relieved, the model can be deep enough, and the residual block retains the original input information during the output, so that the original characteristics of the input data can be still effectively retained in the high-dimensional characteristic space.

The invention adds a Prototype Layer (Prototype Layer) after ResNet50, which is used as a clustering center in a clustering algorithm for learning, and the output of the Prototype Layer is used for the subsequent clustering algorithm and the calculation of the final model loss. Through experiments, when the number of the clustering centers is 3-5 times of the number of specific fault categories, the training effect is good, so that the number of the prototype layer output nodes is 4 times of the number of the estimated fault categories.

The invention uses a sink horn-Knopp algorithm as a clustering algorithm. Because of the limitation of hardware, the video memory capacity of the video card is small, so the model usually splits the data set into a plurality of small batches (min-batch) when learning through the data set. Different from the traditional K-means algorithm, the traditional K-means algorithm needs to traverse all min-latches to obtain all characteristic values, and then calculate a clustering center according to the difference between the characteristic values, so that the traditional clustering algorithm needs to collect the characteristics obtained by extracting all data in the data set through a model, and then perform clustering calculation to obtain the clustering center as a pseudo tag. Then traversing all min-latches again, and calculating the loss by using the output of the model classifier and the pseudo tag. This results in the whole calculation being repeatedly performed twice.

After the encoder of the model, the model is continuously trained by using a Prototype Layer (Prototype Layer), and the Prototype Layer can penetrate through different min-latches of the whole data set. The invention converts the problem that all characteristic values are needed to be clustered into an optimal transmission problem by utilizing a sink horn-Knopp algorithm. This makes it possible to directly input a min-batch into the model, and to perform clustering calculation on the output of the prototype layer by using the sink horn-Knopp, thereby calculating the final loss. Therefore, the whole model is trained by adopting an online clustering algorithm, repeated calculation is not needed, and compared with the traditional clustering algorithm, the whole calculation amount is directly halved.

The specific process of the invention comprises a pre-training stage: step one: from each state of the collected vibration signals, 500 sets of signals were acquired as a source dataset, each set of signals having a length of 2048 data points. And respectively carrying out random data enhancement on the acquired data to respectively obtain two data sets X1 and X2, and then inputting the two data sets into a twin network. The data enhancement application method comprises the following steps: adding random Gaussian Noise (Gaussian Noise), performing random masking (Mask Noise), performing random variation on signal Amplitude (Amplitude Scale), and the like. Step two: the overall structure of the res net encoder is divided into 5 parts, part 1 trains the input x of 1 x 2048 to 64 x 512, and the subsequent parts continue to extract features from the data. The output of part 2 is 256×512, the output of part 3 is 512×256, the output of part 4 is 1024×128, and the output of part 5 is 2048×64. Eventually, through the Average Pooling layer (Average Pooling), the encoder of ResNet outputs a feature of 2048 in length. feature continues to be computed by the classifier, and the final overall model output length is 128. Step three: the input dimension of the prototype layer is the output length 128 of ResNet, and the output dimension of the prototype layer is 4 times the number of estimated fault categories. The output of ResNet50 is input directly to the prototype layer, outputting y1 and y2, respectively. Step four: and respectively obtaining respective clustering centers of y1 and y2 through a sink horn-Knopp algorithm, and taking the clustering centers as pseudo tags label1 and label2 for subsequent use. Step five: and (3) calculating information entropy by interaction between the output of the model and the generated pseudo tag, namely, calculating cross entropy by y1 and label2 and calculating cross entropy by y2 and label1 respectively, and reversely optimizing parameters of the whole model through calculated loss. Repeating the iterative steps two to five until the model converges. In the present invention, the number of iterations is set to 400.

The specific flow of the invention comprises a fine tuning stage: after the pre-training phase is completed, fine tuning is performed using the trained model. The vibration signals adopted in the stage need to be accurately collected, the real fault types of the collected vibration signals are determined, and the data size can be small, namely, a small sample with high precision. Step one: and (3) according to the pre-training stage, acquiring 500 groups of signals from each state of the vibration signals as data sets X ', wherein the length of each group of signals X in the X' is 2048 data points, and inputting the acquired data into the deep neural network after data enhancement. Step two: the deep neural network is reconstructed, the encoders of the models of the pre-training stage are used, and all weights and bias values therein are frozen. And redesigning a new classifier of the deep neural network according to the number of fault categories contained in the data set X', wherein the output length of the classifier is the same as the number of the fault categories. Step three: inputting the data set into a model, extracting the characteristic with the length of 2048 according to the encoder, inputting the characteristic into a classifier for classification, and obtaining a final output Y. Step four: and calculating cross entropy loss by using the output Y of the model and label information Labels, and optimizing parameters of a classifier of the model by using the loss. And repeating the iteration step three and the iteration step four until the model converges. In the present invention, the number of iterations is set to 20. The model completed in the fine tuning stage can be put into a real industrial environment for use, and fault types contained in the data set X' collected in the fine tuning stage can be efficiently diagnosed through the model trained by the method.

The invention introduces the idea of contrast learning, and the contrast learning is performed by defining an example (positive sample) similar to the semantics of the sample and an example (negative sample) different from the semantics of the sample, and the positive samples are closer to each other and the positive samples and the negative samples are farther away from each other in the feature space by designing the model structure and the contrast loss, so that the clustering-like effect is achieved. Compared with the method that the data extraction characteristics are clustered directly through a deep neural network, the contrast learning can enable the difference between different samples to be larger, and therefore the method has better classification performance.

The invention is different from the traditional contrast learning, and positive and negative samples are defined in the agent task so as to train a model. The method introduces the idea of a clustering algorithm, replaces the concept of a negative sample by a clustering center, firstly ensures that the positive sample does not need to be different from all negative samples in calculation, but directly calculates the difference from the clustering center, and reduces the total calculated amount of the model; and in addition, unlike the clustering algorithm for accurately classifying faults, the clustering algorithm provided by the invention uses the clustering centers which are several times the number of the fault categories as a prototype layer for training, so that the problem that the number of the fault categories in the data set is difficult to calculate in a real industrial environment is solved.

The invention introduces a sink horn-Knopp algorithm, and converts a K-Means algorithm which needs to collect all characteristic values and then cluster according to the distribution of data into an optimal transmission problem. This allows the data of each min-batch to be directly passed through the sink horn-Knopp algorithm to calculate the pseudo tag and directly calculate the loss optimization model after the output is obtained through the prototype layer. Thus, an online clustering algorithm is achieved. The method and the system avoid the need of traversing the complete data set, calculating the clustering center, and then traversing the data set again to train the model. The clustering center and the training model are synchronously calculated when the data set is traversed, and the whole training time is saved by half compared with the traditional clustering algorithm; when the data set is too large, even if the feature set is obtained after the data set is extracted by the model, a large amount of storage space may still be occupied if the number of samples is too large. The overlarge feature set is difficult to calculate and takes time, and if the size of the video memory of the video card is exceeded, the feature set is more difficult to calculate. Whereas the online clustering algorithm of the present invention does not have this concern.

The pre-training of the present invention generally divides the overall model into two major parts, a pre-trained model and a later re-trained fine-tuning model. This allows the two-part model to be included, although the objective is to failure classify the vibration signal of the final industrial environment, the way it achieves this objective, the training set used, the method of use, etc. may all be somewhat different. In the invention, aiming at the situations that the number of fault data is less, the cost of manually marking the data is high, and the fault type is difficult to clearly and completely distinguish, a large amount of unlabeled data is adopted in the pre-training stage, and self-supervision learning is applied, so that the whole model is better trained.

The clustering algorithm is applied, and the reason for generating the vibration signals is more complex and various, so that the labeling of all data to be completely and correctly performed is an extremely labor-consuming and even impossible task. According to the invention, the self-classification of the input data is completed by extracting the information carried by the input data and then using the clustering thought that the characteristics of similar data are similar and the difference of different data characteristics is larger. And finally, taking the generated clustering center as a pseudo tag, and training the classification capability of the model by applying a model so that the whole model is changed from unsupervised learning of a clustering algorithm to self-supervised learning, thereby effectively improving the accuracy of model prediction.

The invention provides a fault diagnosis algorithm with good diagnosis effect aiming at the scene that the fault data with the classification labels is insufficient or the fault categories of all the collected data are difficult to distinguish.

According to the method, the clustering algorithm is applied to self-supervision learning, and the clustering center of the clustering algorithm is used for replacing the number of fault categories in traditional classification learning, so that in a model pre-training stage, the strict requirement on data sources is reduced, and the number of data in the unlabeled data set is expanded.

The invention uses the sink horn-Knopp algorithm to convert the traditional clustering algorithm into an optimal transmission problem, so that the method becomes an online clustering algorithm. The operation efficiency of the model is greatly improved, and the occupied video memory required by model calculation is greatly reduced.

The whole scheme is divided into two parts, namely a pre-training model and a later re-training fine-tuning model, so that a data set, a training method and the like adopted by the two parts are divided into two parts. The model migrated in the retraining stage can be trained by using the data set with wider collection range, and the training precision is effectively improved.

The embodiment of the invention comprises the steps that the complete data set of the label is migrated under different working conditions, and bearing vibration signals collected under different working conditions to a certain extent are different, such as influence of air temperature and pressure, air humidity and the like. This is extremely common in real industrial environments, but it is still calculated as data of the same tag as compared to vibration signal variations due to different rotational speeds of the machine, etc. When the data collected in the pre-training model is related to the data used in the fine-tuning training, the invention can extract the high-dimensional characteristics in the vibration signal through the deep learning model, and the characteristics can be basically unchanged under the conditions. Therefore, a large amount of existing data can be used for pretraining, and in the actual working condition, a small amount of homologous data sets are used for fine tuning, so that a good diagnosis effect can be obtained

Embodiments of the present invention include migration of a data set with incomplete labels under different conditions, where the machine is at different speeds or contains an undetected fault, the resulting vibration signal may differ significantly from the expected type of vibration signal. In this case the collected data set will contain more fault categories than expected and vibration signals considered to be classified as a same type of fault are actually different fault categories. If a large amount of data needs to be collected manually, and fault categories of all data are accurately distinguished, the huge manpower resources are consumed, and the method is unreasonable. The invention utilizes the information of the data itself to learn and classify itself through cluster learning. And thereafter retraining with a carefully sampled small data set to avoid degradation of model training accuracy due to the inclusion of "erroneous" data in the training set.

In summary, the foregoing embodiments describe in detail different configurations of the self-supervised bearing fault diagnosis method, and of course, the present invention includes, but is not limited to, the configurations listed in the foregoing embodiments, and any matters of transformation based on the configurations provided in the foregoing embodiments fall within the scope of the present invention. One skilled in the art can recognize that the above embodiments are illustrative.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, the description is relatively simple because of corresponding to the method disclosed in the embodiment, and the relevant points refer to the description of the method section.

The above description is only illustrative of the preferred embodiments of the present invention and is not intended to limit the scope of the present invention, and any alterations and modifications made by those skilled in the art based on the above disclosure shall fall within the scope of the appended claims.

Claims (9)

1. A self-monitoring bearing fault diagnosis method, comprising:

performing bearing fault classification on the self-supervision clustering task based on comparison by using a comparison learning and clustering algorithm, and performing pre-training by using a pre-training data set to obtain a pre-training model; and

retraining the fine tuning of the pre-trained model to obtain a depth model, and putting the depth model into a real industrial environment for use.

2. The self-supervised bearing failure diagnosis method of claim 1, wherein the contrast learning includes:

the method comprises the steps of directly classifying source data extraction features through deep learning, changing direct classification into the use of twin network extraction features, comparing different features obtained by the twin network by defining positive and negative samples so as to distinguish similar data from different data, and distinguishing the data of different categories by oneself.

3. The self-supervised bearing fault diagnosis method as recited in claim 1, wherein the clustering algorithm-based contrast learning comprises:

the twin network comprises two identical deep neural networks through comparison and learning of the architecture twin network;

the deep neural network selects ResNet50, and gradient explosion and gradient disappearance of model training are avoided by using residual blocks;

the output of the deep neural network is directly provided to a prototype layer, and the output of the prototype layer is used for clustering algorithm calculation and cross entropy loss calculation.

4. The self-monitoring bearing fault diagnosis method as claimed in claim 1, further comprising:

carrying out different data enhancement processing on vibration signal source data x to obtain data sets x1 and x2 respectively, and providing the data sets x1 and x2 into two models of a twin network respectively;

the output of the ResNet50 classifier is provided to the prototype layer computation and outputs data sets y1 and y2, respectively;

the data sets y1 and y2 are calculated through a clustering algorithm to obtain data sets label1 and label2;

the output of the prototype layer of one model and the pseudo tag calculated by the other model calculate cross entropy losses through an objective function, respectively calculate the cross entropy losses of y1 and label2 and y2 and label1, and optimize the parameters of ResNet50 through the cross entropy losses.

5. The self-supervised bearing failure diagnostic method of claim 1, wherein contrast learning uses proxy tasks of unlabeled data instead of real labels;

in contrast learning, after different data enhancement processing is carried out on vibration signal source data, input with different forms and unchanged semantics is obtained;

the agent task uses a twin network, vibration signal source data is subjected to different data enhancement processes, so that different input forms are provided to the twin network, and different outputs are obtained through the twin network;

defining positive and negative samples through proxy tasks, and calculating loss through an objective function according to the difference between the positive and negative samples so as to optimize model parameters;

using a contrast clustering algorithm as an agent task, and using a clustering center generated by the clustering algorithm to replace a negative sample in contrast learning;

and alternately comparing the output of one model in the twin network with the pseudo labels obtained by the other model through a clustering algorithm to calculate the loss of the whole twin network and optimize the whole model.

6. A self-monitoring bearing failure diagnosis method according to claim 1, characterized in that,

the depth neural network used in the twin network is a 50-layer residual network, and the main framework of the depth neural network comprises an encoder and a classifier;

introducing a residual block into the encoder, wherein in the final output H (x) of the residual block, the output F (x) obtained by calculating the input x by a convolution layer and the input x themselves participate in the calculation of the final output;

the residual block retains the original input information during output, so that the residual block retains the original characteristics of the input data in a high-dimensional characteristic space;

the prototype layer is added after ResNet50 to learn as a cluster center in the clustering algorithm.

7. The self-supervised bearing failure diagnosis method of claim 1, wherein the pre-training phase includes:

collecting 500 sets of signals from each state of the collected vibration signals as a source data set, wherein each set of signals has a length of 2048 data points;

carrying out random data enhancement on the acquired data respectively to obtain data sets X1 and X2 respectively, and inputting the data sets into a twin network;

wherein the data enhancement method comprises adding random gaussian noise, performing random masking, and performing random variation on the signal amplitude.

The overall structure of the ResNet encoder includes:

the first part trains the input x of 1 x 2048 to 64 x 512, and the subsequent parts continue to extract features from the data;

the output of the second fraction is 256 x 512;

the output of the third fraction is 512 x 256;

the output of the fourth part is 1024×128; and

the output of the fifth part is 2048×64;

in the sixth part, the output obtained before is subjected to an average pooling layer, and the ResNet encoder finally outputs the characteristic with the length of 2048;

the characteristics are calculated through a classifier, and the output length of the whole model is 128 finally;

the input dimension of the prototype layer is the output length 128 of ResNet, and the output dimension of the prototype layer is 4 times of the number of estimated fault categories;

the output of ResNet50 is input directly to the prototype layer, outputting y1 and y2, respectively.

Obtaining respective clustering centers of y1 and y2 through a sink horn-Knopp algorithm respectively, and using the clustering centers as pseudo tags label1 and label2;

calculating information entropy by interaction between the output of the model and the generated pseudo tag, respectively calculating cross entropy by y1 and label2, and calculating cross entropy by y2 and label1, and reversely optimizing parameters of the whole model through calculated loss;

the iteration is repeated until the model converges.

8. The self-monitoring bearing fault diagnosis method as claimed in claim 1, wherein the fine tuning stage comprises:

collecting 500 groups of signals from each state of the vibration signals as data sets X ', wherein the length of each group of signals X in the X' is 2048 data points, and inputting the collected data into a deep neural network after data enhancement;

reconstructing a deep neural network, using an encoder of the model of the pre-training stage, and freezing all weights and bias values therein;

redesigning new classifier of the deep neural network according to the number of fault categories contained in the data set X', wherein the output length of the classifier is the same as the number of the fault categories;

inputting the data set into a model, extracting the characteristic with the length of 2048 according to the encoder, inputting the characteristic into a classifier for classification, and obtaining a final output Y;

calculating cross entropy loss by using the output Y of the model and label information Labels, and optimizing parameters of a classifier of the model by using the loss;

the iteration is repeated until the model converges.

9. A self-monitoring bearing failure diagnosis method according to claim 1, characterized in that,

contrast learning is introduced to compare by defining examples in the samples that are semantically similar to it and examples that are semantically different, and clustering is achieved by designing the model structure and contrast loss so that positive samples are closer together and positive and negative samples are farther apart from each other in the feature space.

Images