A method and system for monitoring fault type in a faulty bearing

Abstract

The present invention provides a method and system for monitoring fault types in a faulty bearing, and relates to the technical field of bearing fault diagnosis. The method comprises collecting bearing vibration signals of various fault types, and constructing a fault diagnosis data set, dividing fault tasks at different stages into initial tasks and incremental tasks, and determining the number of fault categories of the initial tasks and the incremental tasks; constructing an initial fault diagnosis model; based on the initial fault diagnosis model, training the incremental tasks, including model adaptation stage training and model fusion stage training, integrating the new model and the old model through distillation loss to obtain an integrated fault diagnosis model; using the integrated fault diagnosis model to perform fault diagnosis on the vibration signal of the faulty bearing to be detected, and obtaining the fault type of the vibration signal of the faulty bearing to be detected; compared with traditional deep learning methods, the present invention can alleviate the problem of catastrophic forgetting and is more in line with the actual scenarios of industrial applications.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a method and system for monitoring the fault type in a faulty bearing. Background Art

[0002] Bearings are an indispensable key component in the field of machinery, especially in major equipment. They are important basic components in today's mechanical equipment and represent the advanced level of a country's industry. Bearings are widely used in various types of rotating machinery. They are important mechanical foundations and supporting components, and play a vital role in almost any system based on rotating machinery. Many problems encountered by equipment based on rotating machinery during operation are closely related to bearing failures, and whether the bearings are healthy is closely related to whether the unit equipment can operate reliably and stably. Once a rotating machinery fails and causes an accident, it will not only cause direct or indirect losses to the economy, but also cause huge safety hazards and even cause adverse social impacts. Therefore, it is necessary to monitor the health of the bearings.

[0003] Nowadays, equipment based on rotating machinery is constantly developing in the direction of complexity, large-scale, high-speed, intelligent and automated. To diagnose bearing faults, it is first necessary to capture the vibration signal for analysis in the operating mechanical equipment, then extract the useful bearing information in the vibration signal, and identify the fault in combination with the corresponding fault diagnosis method. The relatively traditional fault diagnosis methods such as Fourier transform, short-time Fourier transform, empirical mode decomposition, wavelet transform, etc. that are widely used at present all have more or less dependence on prior knowledge, so they are far from meeting the needs of today's complex systems.

[0004] In recent years, with the continuous maturity of machine learning algorithms, diagnostic technologies based on artificial intelligence have been widely used. For example, fault diagnosis technology based on deep learning has achieved many good results. The fault diagnosis process based on deep learning generally includes three steps: acquiring signals, extracting features, and predicting and identifying. Moreover, the fault diagnosis method based on deep learning can adaptively extract the features of fault data. In the process of establishing a deep learning model, a large amount of fault data is required as support, but rotating mechanical equipment is in a healthy operating state under most conditions. Because if the equipment fails, the equipment will stop working, so the fault data that can be collected is quite limited. Therefore, the lack of fault samples limits the application of fault diagnosis technology based on deep learning. In addition, in actual diagnostic scenarios, the bearing is in a complex working environment, which inevitably leads to the emergence of new fault types, making the fault data in an unbalanced state. At this time, the model needs to be continuously updated to learn new fault types, but most of the current deep learning-based diagnostic models will forget the old knowledge learned before when learning new fault types, that is, the so-called catastrophic forgetting problem occurs, which makes the deep learning-based diagnostic model have low detection accuracy when facing the problem of dealing with faulty bearing types. Therefore, how to alleviate catastrophic forgetting is one of the problems that needs to be solved urgently, and the machine learning paradigm based on continuous learning provides a better solution to this problem. Summary of the invention

[0005] To this end, an embodiment of the present invention provides a method and system for monitoring the fault type in a faulty bearing, which is used to solve the problem that the deep learning model in the prior art suffers from catastrophic forgetting under an unbalanced data set, resulting in low detection accuracy when the current deep learning-based diagnostic model processes the fault bearing type problem.

[0006] In order to solve the above problem, an embodiment of the present invention provides a method for monitoring the fault type in a faulty bearing, the method comprising:

[0007] S1: Collect bearing vibration signals of various fault types and construct a fault diagnosis data set. Divide the fault tasks at different stages into initial tasks and incremental tasks, and determine the number of fault categories for the initial tasks and incremental tasks. The initial task refers to the training task performed using the pre-collected faulty bearing vibration signal data set in the initial stage of fault diagnosis model training. The incremental task refers to the task that requires the fault diagnosis model to learn and adapt when a new fault type or new fault sample appears during the fault diagnosis process. The incremental task consists of several incremental stages. In the training process of each incremental stage, it is divided into a model adaptation stage and a model fusion stage.

[0008] S2: Perform initial task training and extract fault features based on the fault samples and corresponding fault categories of the initial task, and build an initial fault diagnosis model;

[0009] S3: Based on the initial fault diagnosis model, the incremental tasks are trained, including model adaptation stage training and model fusion stage training. The new model and the old model are integrated through distillation loss to obtain the integrated fault diagnosis model, where the new model and the old model refer to the model versions corresponding to different stages in the fault diagnosis model training process, including:

[0010] S31: In the incremental task, according to the fault samples and corresponding fault categories of the incremental task, a branch layer is constructed for each old task respectively, so as to store the old knowledge when the new task arrives, and then the branch layer is merged to solve the model growth problem. The old task refers to the task that has been learned and processed before relative to the incremental task currently being processed, and the new task refers to the new fault type that the model needs to identify or process new fault samples;

[0011] S32: Each incremental stage includes a model adaptation stage and a model fusion stage, wherein the model adaptation stage is trained on the added new fault samples, and the model fusion stage is trained on the new fault samples and some old fault samples selected from all previous fault tasks;

[0012] S33: Use the fusion loss to optimize the model to learn various fault types, and integrate the new model with the old model through distillation loss to obtain the integrated fault diagnosis model;

[0013] S4: Perform fault diagnosis on the vibration signal of the faulty bearing to be detected by using the integrated fault diagnosis model to obtain the fault type of the vibration signal of the faulty bearing to be detected.

[0014] Preferably, the method of training the initial task and extracting fault features according to the fault samples of the initial task and the corresponding fault categories to construct the initial fault diagnosis model is:

[0015] The collected bearing vibration signals are preprocessed by fast Fourier transform, and ResNet-32 is used as the backbone network. The initial task is trained and fault features are extracted according to the fault samples and corresponding fault categories of the initial task to obtain the initial fault diagnosis model.

[0016] Preferably, during the training of the incremental task, the fault diagnosis model includes three parts: a feature extractor F n , branch layer B1,...,B n , and the classifier set C n ;

[0017] The feature extractor F n Used to extract fault feature f, expressed as: f = F n (x i ), x i is the training data;

[0018] The branch layer is set for a specific incremental task. All branch layers are composed of lightweight 1×1 convolutional layers without bias. The output of the branch layer is: It is branch layer B m To make it easier for the model to decode old knowledge, the branch layers B1,...,B n-1 The output of is averaged: b1 is the output of the first branch layer, b n-1 is the output of the n-1th branch layer, is the output of the n-1th branch layer after averaging;

[0019] The classifier set C n ={h1, ..., h n} by a specific classifier h1, ..., h for each task n Composition, where h m It is task T m classifier, m = 1, ..., n, and h m is an output dimension of |J m |The fully connected layer, the classifier is calculated by the following formula:

[0020]

[0021] In the formula, l′ n Represents the Logits output by all old branch layers, l n Represents the Logits output by the new branch layer.

[0022] Preferably, in the incremental task from T n-1 to T n When the branch layer B′ n-2 and B n-1 From T n-1 Get, using the weight matrix and Parameterize it, and the process of branch layer fusion is expressed as:

[0023]

[0024] In the formula, is the weight matrix of the branch layer obtained after the fusion of all previous old tasks in stage n-1, is the weight matrix of the branch layer obtained after the fusion of all previous old tasks in stage n, It is the weight matrix of the branch layer obtained in the n-1 stage.

[0025] Preferably, each incremental stage includes a model adaptation stage and a model fusion stage, wherein the model adaptation stage is trained on the added new fault samples, and the model fusion stage is trained on the new fault samples and some old fault samples selected from all previous fault tasks, specifically including:

[0026] In the model adaptation stage, a trainable model with a branch layer fusion structure In the dataset K which only has new fault samples n The model adaptation stage includes a branch layer for preserving old knowledge and a branch layer for learning new knowledge. The outputs of the two branch layers are In the model adaptation stage, the cross entropy loss is used as the object of learning:

[0027]

[0028] In the formula, is the cross entropy loss, x i is the training data, y i is the label corresponding to the training data, CE represents the cross entropy function;

[0029] In the model fusion stage, the old model φ n-1 And the model obtained in the model adaptation stage Integrated into a single model φ n , B′ n-1 It is obtained by fusion of the branch layers of the old task, B n It is obtained in the model adaptation stage, where the classification loss of the branch layer is:

[0030]

[0031] In the formula, is the classification loss of the branch layer, x i is the training data, y i is the label corresponding to the training data, B′ n-1 is the branch layer obtained by merging the branch layers of all previous old tasks in stage n, B n is the branch layer obtained in stage n, CE represents the cross entropy function, and D n is the total dataset containing old and new classes, M n is the dataset containing the old classes, l n Logits output by the new branch layer, It is the output of the old branch layer obtained by the n-stage fusion through the old classifier set.

[0032] Preferably, the fusion loss is expressed as:

[0033]

[0034] In the formula, is the fusion loss, x i is the training data, y i is the label corresponding to the training data, CE represents the cross entropy function, D n is the total data set containing old and new classes, CE represents the cross entropy function, and D n is the total dataset containing old and new classes, is the total output of the branch layer.

[0035] Preferably, the distillation loss is expressed as:

[0036]

[0037] In the formula, is the distillation loss, x i is the training data, y i is the label corresponding to the training data, D n is the total data set containing old and new classes, KL represents divergence, a mathematical operation, p n-1 is the output probability of the n-1 stage obtained by the Softmax operation, is the output probability of the old class at stage n-1 obtained after the Softmax operation, is the output probability obtained in the n-stage model adaptation phase after the Softmax operation, is the output probability of the new class in stage n obtained after the Softmax operation.

[0038] Preferably, the loss in the model fusion stage is expressed as:

[0039]

[0040] In the formula, is the loss in the model fusion stage, is the fusion loss, is the classification loss of the branch layer, is the distillation loss, δ and λ are hyperparameters between 0 and 1.

[0041] The embodiment of the present invention further provides a system for monitoring the fault type in a faulty bearing, which is used to implement the above-mentioned method for monitoring the fault type in a faulty bearing, and specifically includes:

[0042] The data acquisition module is used to collect bearing vibration signals of various fault types and construct a fault diagnosis data set, divide the fault tasks at different stages into initial tasks and incremental tasks, and determine the number of fault categories of the initial tasks and incremental tasks. The initial task refers to the training task performed using the pre-collected faulty bearing vibration signal data set in the early stage of fault diagnosis model training. The incremental task refers to the task that requires the fault diagnosis model to learn and adapt when a new fault type or new fault sample appears during the fault diagnosis process. The incremental task consists of several incremental stages. In the training process of each incremental stage, it is divided into a model adaptation stage and a model fusion stage.

[0043] The initial task training module is used to train the initial task and extract fault features according to the fault samples of the initial task and the corresponding fault categories, and to build an initial fault diagnosis model;

[0044] The incremental task training module is used to train the incremental tasks based on the initial fault diagnosis model, including model adaptation stage training and model fusion stage training. The new model and the old model are integrated through distillation loss to obtain the integrated fault diagnosis model, where the new model and the old model refer to the model versions corresponding to different stages in the fault diagnosis model training process, including:

[0045] In the incremental task, according to the fault samples and corresponding fault categories of the incremental task, a branch layer is built for each old task to store old knowledge when a new task arrives, and then the branch layers are merged to solve the model growth problem. The old task refers to the task that has been learned and processed before relative to the incremental task currently being processed, and the new task refers to the model needs to identify new fault types or process new fault samples;

[0046] Each incremental phase includes a model adaptation phase and a model fusion phase. The model adaptation phase is trained on the newly added fault samples, and the model fusion phase is trained on the new fault samples and some old fault samples selected from all previous fault tasks.

[0047] The fusion loss optimization model is used to learn various fault types, and the new model and the old model are integrated through the distillation loss to obtain the integrated fault diagnosis model;

[0048] The diagnosis module is used to perform fault diagnosis on the vibration signal of the faulty bearing to be detected by using the integrated fault diagnosis model to obtain the fault type of the vibration signal of the faulty bearing to be detected.

[0049] An embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium stores a computer software product, wherein the computer software product includes a number of instructions for enabling a computer device to execute the above-mentioned method for monitoring the fault type in a faulty bearing.

[0050] It can be seen from the above technical solutions that the present invention has the following beneficial effects:

[0051] The method of the present invention can effectively alleviate catastrophic forgetting and realize the continuous diagnosis of bearing faults with new unexpected faults. The present invention divides the fault tasks of different stages into initial tasks and incremental tasks. In the initial task, the model will use the pre-collected fault bearing vibration signal data set to learn how to identify different fault types and build an initial fault diagnosis model. This initial model provides the basis for subsequent incremental tasks. The incremental task consists of several incremental stages. In each stage of the incremental task, the model will go through an adaptive stage and a fusion stage. The adaptive stage mainly uses the newly added fault samples for training so that the model can identify new fault types; while the fusion stage will be trained on the new fault samples and some old fault samples selected from all previous fault tasks. This can not only alleviate catastrophic forgetting (that is, the model forgets the previously learned knowledge), but also ensure that the model can maintain good performance on both new and old tasks. In the incremental task, a branch layer will also be built for each old task to store old knowledge, and the problem of model growth will be solved by merging branch layers. Compared with traditional deep learning methods, the present invention can alleviate the problem of catastrophic forgetting and is more in line with the actual scenarios of industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the implementation cases of the present invention or the technical solutions in the prior art, the following is a brief description of the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0053] Figure 1 It is a flow chart of a method for monitoring the fault type in a faulty bearing provided in an embodiment;

[0054] Figure 2 A schematic diagram of the structure of a fault diagnosis model in an embodiment;

[0055] Figure 3 A schematic diagram of a confusion matrix obtained after completing an incremental task in the embodiment;

[0056] Figure 4 This is a schematic diagram of feature visualization obtained after completing the incremental task in the embodiment;

[0057] Figure 5 A block diagram of a system for monitoring fault types in a faulty bearing provided in an embodiment. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Embodiment 1

[0060] In order to solve the problem that deep learning models have catastrophic forgetting under unbalanced data sets, the current deep learning-based diagnostic models have low detection accuracy when dealing with faulty bearing types. Figure 1 As shown, an embodiment of the present invention provides a method for monitoring the fault type in a faulty bearing, the method comprising:

[0061] S1: Collect bearing vibration signals of various fault types and construct a fault diagnosis data set. Divide the fault tasks at different stages into initial tasks and incremental tasks, and determine the number of fault categories for the initial tasks and incremental tasks.

[0062] S2: Perform initial task training and extract fault features based on the fault samples and corresponding fault categories of the initial task, and build an initial fault diagnosis model;

[0063] S3: Based on the initial fault diagnosis model, the incremental tasks are trained, including model adaptation phase training and model fusion phase training. The new model and the old model are integrated through distillation loss to obtain the integrated fault diagnosis model.

[0064] S4: Perform fault diagnosis on the vibration signal of the faulty bearing to be detected by using the integrated fault diagnosis model to obtain the fault type of the vibration signal of the faulty bearing to be detected.

[0065] As can be seen from the above technical solution, the present invention proposes a method for monitoring fault types in faulty bearings. The present invention divides fault tasks at different stages into initial tasks and incremental tasks. In the initial task, the model will use the pre-collected fault bearing vibration signal data set to learn how to identify different fault types and build an initial fault diagnosis model. This initial model provides a basis for subsequent incremental tasks. The incremental task consists of several incremental stages. In each stage of the incremental task, the model will go through an adaptive stage and a fusion stage. The adaptive stage mainly uses the newly added fault samples for training so that the model can identify new fault types; while the fusion stage will be trained on the new fault samples and some old fault samples selected from all previous fault tasks. This can not only alleviate catastrophic forgetting (that is, the model forgets the previously learned knowledge), but also ensure that the model can maintain good performance on both new and old tasks. In the incremental task, a branch layer will also be constructed for each old task to store old knowledge, and the problem of model growth will be solved by merging branch layers. Compared with traditional deep learning methods, the present invention can alleviate the problem of catastrophic forgetting and is more in line with the actual scenarios of industrial applications.

[0066] In this embodiment, in step S1, bearing vibration signals of various fault types required for the experiment are collected and a fault diagnosis data set is constructed. In order to obtain a faulty bearing, cracks of different sizes are set on the inner ring, outer ring and roller of the bearing using wire cutting technology to simulate local bearing faults in actual production.

[0067] The present invention divides fault tasks at different stages into initial tasks and incremental tasks, and determines the number of fault categories of the initial tasks and the incremental tasks.

[0068] Specifically, the initial task refers to the training task performed in the early stage of fault diagnosis model training using pre-collected faulty bearing vibration signal datasets. These datasets usually include relatively more faulty bearing vibration signals and corresponding fault categories. In the initial task, the model will use this data to learn how to identify different fault types and build an initial fault diagnosis model. This initial model provides the basis for subsequent incremental tasks, and in the incremental tasks, the model will adapt and integrate according to new fault samples and categories to continuously improve its fault diagnosis accuracy and generalization ability.

[0069] Incremental tasks refer to tasks that require the fault diagnosis model to learn and adapt when new fault types or new fault samples appear during the fault diagnosis process. These tasks are not completed at one time, but are carried out in stages, and each stage corresponds to a specific incremental stage. In each stage of the incremental task, the model will go through an adaptive stage and a fusion stage. The adaptive stage mainly uses the newly added fault samples for training so that the model can recognize new fault types. The fusion stage will be trained on new fault samples and some old fault samples selected from all previous fault tasks. This can not only alleviate catastrophic forgetting (that is, the model forgets the previously learned knowledge), but also ensure that the model maintains good performance on both new and old tasks. In the incremental task, a branch layer is also built for each old task to store old knowledge, and the problem of model growth is solved by merging branch layers.

[0070] In the experimental parameter setting, the motor speed is set to 800r / min and the sampling frequency is set to 12.8KHz. The constructed fault data set contains 26 health states, as shown in Table 1 (bearing health state and fault task setting), NO represents normal state, IF represents inner ring fault state, OF represents outer ring fault state, BF represents roller fault state, IF+BF represents inner ring fault and roller fault state, and IF+OF represents inner ring fault and outer ring fault state. The bearing has two health states: normal and fault, among which the faulty bearing includes inner ring fault, roller fault, outer ring fault and compound fault.

[0071] The fault vibration signal of the bearing is collected using an acceleration sensor, and a data set containing various fault types is constructed. The fault diagnosis task is divided into an initial task and an incremental task: the initial task corresponds to task T0, at which time the model is in the data set Training is performed on is the training data in the initial stage, is the label corresponding to the training data, p 0 Refers to the number of samples of the fault type in the initial task. n In the imbalanced dataset, the model Training is performed on n Finger T n The sum of all fault samples involved in the training (e.g., the number of samples for each new fault type is 100, and the number of samples for each old fault type is 10).

[0072] As shown in Table 1, the initial task is set with 11 fault types (labels 0-10). T1, T2, and T3 are three incremental tasks, and each task contains 5 new fault types (labels 11-15 correspond to T1, labels 16-20 correspond to T2, and labels 21-25 correspond to T3). In this embodiment, the fault types of all previous tasks are retained in the incremental tasks, but the retained samples are relatively small, thus forming an unbalanced data set. Taking T2 as an example, T2 corresponds to labels 16-20, and labels 16-20 each contain 100 samples. The 11 types contained in the initial task and the 5 types contained in T1 only have 5 to 10 samples.

[0073] In real-life scenarios, the amount of data on healthy bearings is much greater than the amount of data on faulty bearings. These healthy bearings are common, so there are many types and they are relatively easier to learn. Therefore, normal bearings and 10 types of faulty bearings are learned in Task T0; to simulate the incremental fault types that occur due to unexpected occurrences in real-life scenarios, in each subsequent incremental task, the fault types learned include inner ring faults, outer ring faults, roller faults, and composite faults. There are 100 training samples and 100 test samples for each fault type in Task T0, and there are also 100 training samples and 100 test samples for each fault type in the incremental task.

[0074] Table 1

[0075]

[0076]

[0077] In this embodiment, in step S2, the present invention uses fast Fourier transform to preprocess the collected bearing vibration signal, and uses ResNet-32 as the backbone network to train the initial task and extract fault features according to the fault samples of the initial task and the corresponding fault categories to obtain the initial fault diagnosis model.

[0078] In this embodiment, in step S3, based on the initial fault diagnosis model, the incremental task is trained, including model adaptation stage training and model fusion stage training, and the new model and the old model are integrated through distillation loss to obtain an integrated fault diagnosis model, which specifically includes:

[0079] S31: In the incremental task, according to the fault samples and corresponding fault categories of the incremental task, a branch layer is constructed for each old task to store old knowledge when a new task arrives, and then the branch layers are merged to solve the model growth problem;

[0080] S32: Each incremental stage includes a model adaptation stage and a model fusion stage, wherein the model adaptation stage is trained on the added new fault samples, and the model fusion stage is trained on the new fault samples and some old fault samples selected from all previous fault tasks;

[0081] S33: Use the fusion loss optimization model to learn various fault types, and integrate the new model with the old model through distillation loss to obtain the integrated fault diagnosis model.

[0082] The new model and the old model refer to the model versions corresponding to different stages in the fault diagnosis model training process. The new model usually refers to the model version obtained after training on new fault samples in the incremental learning task. This model not only contains the knowledge in the old model (that is, the fault types learned before), but also increases the recognition ability of new fault types. In each stage of incremental learning, the new model will update and optimize its performance through the adaptation stage and the fusion stage to adapt to the new task requirements. The old model refers to the model version that has been trained before incremental learning. This model only contains the recognition ability of the fault types in the initial task, but not the recognition of the newly added fault types. As the incremental learning proceeds, the old model will gradually be replaced by the new model, but its role in the incremental process cannot be ignored because it provides the basic knowledge and framework for the new model. In fault diagnosis, the conversion and update between the new model and the old model is a key process. By continuously integrating the newly learned knowledge into the model while retaining the recognition ability of the old task, the model can gradually adapt to more complex fault diagnosis tasks and improve the accuracy and generalization ability of diagnosis.

[0083] An old task refers to a task that has been learned and processed before, relative to the incremental task currently being processed. A new task refers to a new fault type that the model needs to identify or a new fault sample that needs to be processed. During the training of the fault diagnosis model, as time goes by and the equipment runs, fault types that have not been encountered before may appear. In order for the model to accurately identify these new fault types, it is necessary to perform incremental learning on the model. In this process, the new task is to introduce these new fault types or samples, and require the model to learn to identify new fault types while retaining the ability to identify old fault types. The introduction of new tasks usually requires the model to be trained in the adaptive stage and the fusion stage. In the adaptive stage, the model focuses on new fault samples and learns new fault features; in the fusion stage, the model will weigh between new samples and old samples to ensure that good performance is maintained on both new and old tasks. Therefore, the new task not only represents the new knowledge that the model needs to learn, but also reflects the ability of the model to continuously adapt and evolve in practical applications. By continuously processing new tasks, the model can gradually improve its fault diagnosis accuracy and generalization ability to better meet the needs of practical applications.

[0084] Specifically, after completing the initial task, the incremental task learning begins. In the subsequent incremental tasks, the model continuously learns new fault types. During the training of the incremental tasks, the fault diagnosis model consists of three parts: feature extractor F n , branch layer B1,...,B n , and the classifier set C n The fault diagnosis model structure of the present invention is as follows: Figure 2 shown.

[0085] Furthermore, the feature extractor F n Used to extract fault feature f, expressed as: f = F n (x i ), x i is the training data.

[0086] Furthermore, the branch layer is set for a specific incremental task. The branch layer obtained from the old model will be frozen in the new training process to preserve the old knowledge. All branch layers are composed of lightweight 1×1 convolutional layers without bias. This can alleviate the problem of excessive increase in model parameters caused by generating too many branch layers to a certain extent. The output of the branch layer is: It is branch layer B m The weight of .

[0087] Furthermore, in order to make it easier for the model to decode old knowledge, the branch layers B1,...,Bn-1 The output of is averaged:

[0088]

[0089] Where b1 is the output of the first branch layer, b n-1 is the output of the n-1th branch layer, It is the output of the n-1th branch layer after averaging.

[0090] Classifier Set C n ={h1, ..., h n} by a specific classifier h1, ..., h for each task n Composition, where h m It is task T m classifier, m = 1, ..., n, and h m is an output dimension of |J m |The fully connected layer, the classifier is calculated by the following formula:

[0091]

[0092] in

[0093]

[0094] In the formula, l′ n Represents the Logits output by all old branch layers, l n Represents the Logits output by the new branch layer.

[0095] Furthermore, in the incremental task from T n-1 to T n When the branch layer B′ n-2 and B n-1 From T n-1 Get, using the weight matrix and To parameterize it, since the branch layers are composed of lightweight 1×1 convolutional layers without bias, the weight matrix can be used and It is parameterized. The process of branch layer fusion is expressed as:

[0096]

[0097] In the formula, is the weight matrix of the branch layer obtained after the fusion of all previous old tasks in stage n-1, is the weight matrix of the branch layer obtained after the fusion of all previous old tasks in stage n, It is the weight matrix of the branch layer obtained in the n-1 stage.

[0098] Furthermore, each incremental stage includes a model adaptation stage and a model fusion stage, wherein the model adaptation stage is trained on the newly added fault samples, and the model fusion stage is trained on the new fault samples and some old fault samples selected from all previous fault tasks, and the number of old samples selected is much smaller than the number of samples corresponding to the new fault type of the fault task, specifically including:

[0099] In the model adaptation stage, a trainable model with a branch layer fusion structure In the dataset K which only has new fault samples n The model adaptation stage includes a branch layer for preserving old knowledge and a branch layer for learning new knowledge. The outputs of the two branch layers are In the model adaptation stage, the cross entropy loss is used as the object of learning:

[0100]

[0101] In the formula, is the cross entropy loss, x i is the training data, y i is the label corresponding to the training data, and CE represents the cross entropy function.

[0102] In the model fusion stage, the old model φ n-1 And the model obtained in the model adaptation stage Integrated into a single model φ n , B′ n-1 It is obtained by fusion of the branch layers of the old task, B n It is obtained in the model adaptation stage, where the classification loss of the branch layer is:

[0103]

[0104] In the formula, is the classification loss of the branch layer, x i is the training data, y i is the label corresponding to the training data, B′ n-1 B is the branch layer obtained by merging the branch layers of all previous old tasks in stage n. n is the branch layer obtained in stage n, CE represents the cross entropy function, and D n is the total dataset containing old and new classes, M n is the dataset containing the old classes, l n Logits output by the new branch layer, It is the output of the old branch layer obtained by the n-stage fusion through the old classifier set.

[0105] Furthermore, in order to optimize the model's learning of all fault types, the present invention proposes a fusion loss. Finally, the old model φ is transformed into n-1 And the model obtained in the adaptive stage Integrate them together to obtain the integrated fault diagnosis model.

[0106] The fusion loss is expressed as:

[0107]

[0108] In the formula, is the fusion loss, x i is the training data, y i is the label corresponding to the training data, CE represents the cross entropy function, D n is the total data set containing old and new classes, CE represents the cross entropy function, and D n is the total dataset containing old and new classes, is the total output of the branch layer.

[0109] The distillation loss is expressed as:

[0110]

[0111] In the formula, is the distillation loss, x i is the training data, y i is the label corresponding to the training data, D n is the total data set containing old and new classes, KL represents divergence, a mathematical operation, p n-1 is the output probability of the n-1 stage obtained by the Softmax operation, is the output probability of the old class at stage n-1 obtained after the Softmax operation, is the output probability obtained in the n-stage model adaptation phase after the Softmax operation, is the output probability of the new class in stage n obtained after the Softmax operation.

[0112] The loss at the model fusion stage is expressed as:

[0113]

[0114] In the formula, is the loss in the model fusion stage, is the fusion loss, is the classification loss of the branch layer, is the distillation loss, δ and λ are hyperparameters between 0 and 1.

[0115] In this embodiment, in step S4, the integrated fault diagnosis model is used to perform fault diagnosis on the vibration signal of the faulty bearing to be detected, and the fault type of the vibration signal of the faulty bearing to be detected is obtained.

[0116] Specifically, after the incremental task training is completed, that is, after T3 in Table 1 is completed, the obtained diagnostic model needs to be tested. The test data includes all the fault types that have been learned to verify whether the obtained diagnostic model has the ability to alleviate catastrophic forgetting. The test results are presented in the form of a confusion matrix. Figure 3 The displayed confusion matrix shows that this method has a good anti-forgetting effect. Figure 4 The classification effect of the method is further demonstrated through feature visualization, and the results show that the training model obtained by the method can cluster the features well, indicating that the method of the invention can effectively alleviate catastrophic forgetting and realize continuous bearing fault diagnosis of new unexpected faults.

[0117] In summary, compared with traditional deep learning methods, the present invention can alleviate the problem of catastrophic forgetting and is more in line with the actual scenarios of industrial applications.

[0118] Embodiment 2

[0119] like Figure 5 As shown, the present invention provides a system for monitoring the fault type in a faulty bearing, which is used to implement the method for monitoring the fault type in a faulty bearing of the first embodiment, and specifically includes:

[0120] The data acquisition module 100 is used to collect bearing vibration signals of various fault types and construct a fault diagnosis data set, divide the fault tasks at different stages into initial tasks and incremental tasks, and determine the number of fault categories of the initial tasks and incremental tasks, wherein the initial task refers to a training task performed using a pre-collected faulty bearing vibration signal data set in the initial stage of fault diagnosis model training, and the incremental task refers to a task that requires the fault diagnosis model to learn and adapt when a new fault type or new fault sample appears during the fault diagnosis process. The incremental task consists of several incremental stages, and the training process of each incremental stage is divided into a model adaptation stage and a model fusion stage;

[0121] The initial task training module 200 is used to perform initial task training and extract fault features according to the fault samples of the initial task and the corresponding fault categories, and to build an initial fault diagnosis model;

[0122] The incremental task training module 300 is used to train the incremental task based on the initial fault diagnosis model, including model adaptation stage training and model fusion stage training, and integrate the new model and the old model through distillation loss to obtain an integrated fault diagnosis model, wherein the new model and the old model refer to the model versions corresponding to different stages in the fault diagnosis model training process, specifically including:

[0123] In the incremental task, according to the fault samples and corresponding fault categories of the incremental task, a branch layer is built for each old task to store old knowledge when a new task arrives, and then the branch layers are merged to solve the model growth problem. The old task refers to the task that has been learned and processed before relative to the incremental task currently being processed, and the new task refers to the model needs to identify new fault types or process new fault samples;

[0124] Each incremental phase includes a model adaptation phase and a model fusion phase. The model adaptation phase is trained on the newly added fault samples, and the model fusion phase is trained on the new fault samples and some old fault samples selected from all previous fault tasks.

[0125] The fusion loss optimization model is used to learn various fault types, and the new model and the old model are integrated through the distillation loss to obtain the integrated fault diagnosis model;

[0126] The diagnosis module 400 is used to perform fault diagnosis on the vibration signal of the faulty bearing to be detected by using the integrated fault diagnosis model to obtain the fault type of the vibration signal of the faulty bearing to be detected.

[0127] A system for monitoring fault types in a faulty bearing in this embodiment is used to implement the aforementioned method for monitoring fault types in a faulty bearing. Therefore, the specific implementation method of the system for monitoring fault types in a faulty bearing can be found in the embodiment section of the aforementioned method for monitoring fault types in a faulty bearing. For example, the data acquisition module 100, the initial task training module 200, the incremental task training module 300, and the diagnosis module 400 are respectively used to implement steps S1, S2, S3, and S4 in the aforementioned method for monitoring fault types in a faulty bearing. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part. In order to avoid redundancy, it will not be repeated here.

[0128] Embodiment 3

[0129] An embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium stores a computer software product, wherein the computer software product includes a number of instructions for enabling a computer device to execute the above-mentioned method for monitoring the fault type in a faulty bearing.

[0130] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0131] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0133] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. A method for monitoring the fault type in a faulty bearing, characterized in that: include: S1: Collect bearing vibration signals of various fault types and construct a fault diagnosis data set. Divide the fault tasks at different stages into initial tasks and incremental tasks, and determine the number of fault categories for the initial tasks and incremental tasks. The initial task refers to the training task performed using the pre-collected faulty bearing vibration signal data set in the initial stage of fault diagnosis model training. The incremental task refers to the task that requires the fault diagnosis model to learn and adapt when a new fault type or new fault sample appears during the fault diagnosis process. The incremental task consists of several incremental stages. In the training process of each incremental stage, it is divided into a model adaptation stage and a model fusion stage. S2: Perform initial task training and extract fault features based on the fault samples and corresponding fault categories of the initial task, and build an initial fault diagnosis model; S3: Based on the initial fault diagnosis model, the incremental tasks are trained, including model adaptation stage training and model fusion stage training. The new model and the old model are integrated through distillation loss to obtain the integrated fault diagnosis model, where the new model and the old model refer to the model versions corresponding to different stages in the fault diagnosis model training process, including: S31: In the incremental task, according to the fault samples and corresponding fault categories of the incremental task, a branch layer is constructed for each old task respectively, so as to store the old knowledge when the new task arrives, and then the branch layer is merged to solve the model growth problem. The old task refers to the task that has been learned and processed before relative to the incremental task currently being processed, and the new task refers to the new fault type that the model needs to identify or process new fault samples; In the process of training incremental tasks, the fault diagnosis model consists of three parts: feature extractor F n , branch layer B1,...,B n , and the classifier set C n ; The feature extractor F n Used to extract fault feature f, expressed as: f = F n (x i ), x i is the training data; The branch layer is set for incremental tasks. All branch layers are composed of lightweight 1×1 convolutional layers without bias. The output of the branch layer is: It is branch layer B m To make it easier for the model to decode old knowledge, the branch layers B1,...,B n-1 The output of is averaged: b1 is the output of the first branch layer, b n-1 is the output of the n-1th branch layer, is the output of the n-1th branch layer after averaging; The classifier set C n ={h1,...,h n } is composed of a specific classifier h1,...,h n Composition, where h m It is task T m classifier, m = 1, ..., n, and h m is an output dimension of |J m |The fully connected layer, the classifier is calculated by the following formula: In the formula, l′ n Represents the Logits output by all old branch layers, l n Logits representing the output of the new branch layer; In the incremental task from T n-1 to T n When the branch layer B′ n-2 and B n-1 From T n-1 Get, using the weight matrix and Parameterize it, and the process of branch layer fusion is expressed as: In the formula, is the weight matrix of the branch layer obtained after the fusion of all previous old tasks in stage n-1, is the weight matrix of the branch layer obtained after the fusion of all previous old tasks in stage n, is the weight matrix of the branch layer obtained in the n-1 stage; S32: Each incremental stage includes a model adaptation stage and a model fusion stage, wherein the model adaptation stage is trained on the added new fault samples, and the model fusion stage is trained on the new fault samples and some old fault samples selected from all previous fault tasks; S33: Use the fusion loss to optimize the model to learn various fault types, and integrate the new model with the old model through distillation loss to obtain the integrated fault diagnosis model; S4: using the integrated fault diagnosis model to perform fault diagnosis on the vibration signal of the faulty bearing to be detected, and obtaining the fault type of the vibration signal of the faulty bearing to be detected.

2. The method for monitoring the fault type in a faulty bearing according to claim 1, characterized in that: The method of training the initial task and extracting fault features based on the fault samples of the initial task and the corresponding fault categories to construct the initial fault diagnosis model is as follows: The collected bearing vibration signals are preprocessed by fast Fourier transform, and ResNet-32 is used as the backbone network. The initial task is trained and fault features are extracted according to the fault samples and corresponding fault categories of the initial task to obtain the initial fault diagnosis model.

3. The method for monitoring the fault type in a faulty bearing according to claim 1, characterized in that: Each incremental stage includes a model adaptation stage and a model fusion stage, wherein the model adaptation stage is trained on the newly added fault samples, and the model fusion stage is trained on the new fault samples and some old fault samples selected from all previous fault tasks, specifically including: In the model adaptation stage, a trainable model with a branch layer fusion structure In the dataset K which only has new fault samples n The model adaptation stage includes a branch layer for preserving old knowledge and a branch layer for learning new knowledge. The outputs of the two branch layers are In the model adaptation stage, the cross entropy loss is used as the object of learning: In the formula, is the cross entropy loss, x i is the training data, y i is the label corresponding to the training data, CE represents the cross entropy function; In the model fusion stage, the old model φ n-1 And the model obtained in the model adaptation stage Integrated into a single model φ n , B′ n-1 It is obtained by fusion of the branch layers of the old task, B n It is obtained in the model adaptation stage, where the classification loss of the branch layer is: In the formula, is the classification loss of the branch layer, x i is the training data, y i is the label corresponding to the training data, B′ n-1 is the branch layer obtained by merging the branch layers of all previous old tasks in stage n, B n is the branch layer obtained in stage n, CE represents the cross entropy function, and D n is the total dataset containing old and new classes, M n is the dataset containing the old classes, l n Logits output by the new branch layer, It is the output of the old branch layer obtained by the n-stage fusion through the old classifier set.

4. The method for monitoring the fault type in a faulty bearing according to claim 1, characterized in that: The fusion loss is expressed as: In the formula, is the fusion loss, x i is the training data, y i is the label corresponding to the training data, CE represents the cross entropy function, D n is the total data set containing old and new classes, CE represents the cross entropy function, and D n is the total dataset containing old and new classes, is the total output of the branch layer.

5. The method for monitoring the fault type in a faulty bearing according to claim 1, characterized in that: The distillation loss is expressed as: In the formula, is the distillation loss, x i is the training data, y i is the label corresponding to the training data, D n is the total data set containing old and new classes, KL represents divergence, a mathematical operation, p n-1 is the output probability of the n-1 stage obtained by the Softmax operation, is the output probability of the old class at stage n-1 obtained after the Softmax operation, is the output probability obtained in the n-stage model adaptation phase after the Softmax operation, is the output probability of the new class in stage n obtained after the Softmax operation.

6. The method for monitoring the fault type in a faulty bearing according to claim 1, characterized in that: The loss of the model fusion stage is expressed as: In the formula, is the loss in the model fusion stage, is the fusion loss, is the classification loss of the branch layer, is the distillation loss, δ and λ are hyperparameters between 0 and 1.

7. A fault type monitoring system for a faulty bearing, characterized in that: The system is used to implement the method for monitoring the fault type in a faulty bearing according to any one of claims 1 to 6, specifically comprising: The data acquisition module is used to collect bearing vibration signals of various fault types and construct a fault diagnosis data set, divide the fault tasks at different stages into initial tasks and incremental tasks, and determine the number of fault categories of the initial tasks and incremental tasks. The initial task refers to the training task performed using the pre-collected faulty bearing vibration signal data set in the early stage of fault diagnosis model training. The incremental task refers to the task that requires the fault diagnosis model to learn and adapt when a new fault type or new fault sample appears during the fault diagnosis process. The incremental task consists of several incremental stages. In the training process of each incremental stage, it is divided into a model adaptation stage and a model fusion stage. The initial task training module is used to train the initial task and extract fault features according to the fault samples of the initial task and the corresponding fault categories, and to build an initial fault diagnosis model; The incremental task training module is used to train the incremental tasks based on the initial fault diagnosis model, including model adaptation stage training and model fusion stage training. The new model and the old model are integrated through distillation loss to obtain the integrated fault diagnosis model, where the new model and the old model refer to the model versions corresponding to different stages in the fault diagnosis model training process, including: In the incremental task, according to the fault samples and corresponding fault categories of the incremental task, a branch layer is built for each old task to store old knowledge when a new task arrives, and then the branch layers are merged to solve the model growth problem. The old task refers to the task that has been learned and processed before relative to the incremental task currently being processed, and the new task refers to the model needs to identify new fault types or process new fault samples; Each incremental phase includes a model adaptation phase and a model fusion phase. The model adaptation phase is trained on the newly added fault samples, and the model fusion phase is trained on the new fault samples and some old fault samples selected from all previous fault tasks. The fusion loss optimization model is used to learn various fault types, and the new model and the old model are integrated through the distillation loss to obtain the integrated fault diagnosis model; The diagnosis module is used to perform fault diagnosis on the vibration signal of the faulty bearing to be detected by using the integrated fault diagnosis model to obtain the fault type of the vibration signal of the faulty bearing to be detected.

8. A computer storage medium, characterized in that The computer storage medium stores a computer software product, and the computer software product includes a number of instructions for enabling a computer device to execute the method for monitoring the fault type in a faulty bearing according to any one of claims 1 to 6.

Images