Incremental fault diagnosis method for rotating machinery in data-restricted scenarios

Abstract

The application discloses a kind of rotating machinery incremental fault diagnosis methods in data limited scene, belong to the technical field of fault diagnosis.This method is trained rotating machinery incremental fault diagnosis basic backbone network using rotating machinery incremental fault basic training set to obtain basic classifier;Data sampling is carried out in rotating machinery incremental fault basic training set to obtain support set and query set, support set and query set form pseudo incremental mode, support set obtains fault prototype by sharing rotating machinery incremental fault diagnosis basic backbone network parameter, and basic classifier is expanded using fault prototype, query set obtains query embedding by sharing rotating machinery incremental fault diagnosis basic backbone network parameter, and calibration module driven by knowledge feature optimization data knowledge is calculated by support set and query set and constructs meta-training task.It is corrected that the deviation of new fault characterization in incremental process model, pseudo incremental task is dynamically simulated in basic conversation to simulate real incremental task, model learns more general feature embedding, and the generalization ability of model to new fault is improved.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to an incremental fault diagnosis method for rotating machinery in data-constrained scenarios. Background Technology

[0002] Rotating machinery often operates under high speed and heavy load conditions, in harsh environments, and for extended periods, making it prone to failure. This can lead to the downtime of the entire machine, and even cause safety accidents and significant economic losses. Rolling bearings, as crucial mechanical components, play a vital role in almost all rotating machinery systems. Therefore, early, accurate, and efficient diagnosis and monitoring of potential bearing failures are of great significance for improving equipment reliability and maintenance economy. Currently, most deep learning-based intelligent diagnostic methods follow the paradigm of constructing static datasets for model training, meaning all fault modes are available during the training phase. In real-world industrial scenarios, fault data exists within the continuous information stream generated during the operation of mechanical systems. In complex operating environments, rotating machinery may continuously generate new fault categories, but the sample set of new faults is limited, making it difficult to collect data on all possible fault modes of bearings or other mechanical equipment in advance. When a new fault type appears, the original diagnostic model will become ineffective; therefore, it is necessary to update the model using new fault type data. However, training the model solely with new fault data leads to poor performance of older data types—a phenomenon known as catastrophic forgetting. One of the most direct methods to prevent catastrophic forgetting is to integrate data from all fault types to retrain the model. Due to memory and privacy constraints, storing all monitoring data is impractical in engineering. Furthermore, retraining the model with large amounts of data consumes significant time and computational resources. Incremental learning strategies enable diagnostic models to adapt to continuously emerging fault types. Existing incremental fault diagnosis methods assume that data is sufficient in both the initial and incremental phases. However, rotating equipment operates in good working order most of the time. If a malfunction occurs, it will shut down immediately, resulting in a limited amount of collected fault data. The scarcity of new fault data leads to overfitting, making it difficult for the model to effectively infer new categories and resulting in poor incremental learning performance. Secondly, the overfitting problem caused by the scarcity of fault data exacerbates the catastrophic forgetting of few fault types as faults continue to increase. Summary of the Invention

[0003] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing an incremental fault diagnosis method for rotating machinery in data-constrained scenarios. The method uses a data-knowledge jointly driven calibration module to adjust the classifiers for old fault types and the class prototypes for new fault types, correcting the deviation of the fault representation in the incremental process model. In the basic session, a pseudo-incremental task is constructed to dynamically simulate the real incremental task, enabling the model to learn more generalized feature embeddings, improving the model's stability and plasticity, and facilitating its widespread use.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an incremental fault diagnosis method for rotating machinery in data-constrained scenarios, characterized in that the method includes the following steps: Step 1: Train the basic backbone network for incremental fault diagnosis of rotating machinery using the basic training set of incremental faults of rotating machinery to obtain the basic classifier; Step 2: Sample data from the incremental fault training set of rotating machinery to obtain the support set and query set; Step 3: Construct the meta-training task: Form a pseudo-incremental pattern by sharing the support set and query set. The support set obtains fault prototypes by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The basic classifier is then extended using the fault prototypes. The query set obtains query embeddings by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The meta-training task is constructed using the support set and query set, and knowledge features are calculated. Step 4: Optimize the data-knowledge joint-driven calibration module: Input the expanded base classifier, query embedding, and knowledge features into the data-knowledge joint-driven calibration module to obtain the calibrated classifier and the calibrated fault representation. The data-knowledge joint-driven calibration module includes an attention calibration model and a feedforward neural network with fine-tuning by using prior knowledge features as a prefix. The attention calibration model introduces a scaled dot product attention mechanism to calculate the similarity between the calibrated classifier output and the fault representation features, generating a fault prediction result. Subsequently, the prediction result is compared with the true label to calculate the cross-entropy loss value, and the loss function is iteratively optimized through the backpropagation algorithm. Step 5: Construct a support set for incremental faults from the actual newly added limited incremental fault data of rotating machinery. The test data of all fault types constitute the query set for incremental faults of rotating machinery. The support set for incremental faults of rotating machinery obtains new fault prototypes by sharing the basic backbone network parameters of incremental fault diagnosis of rotating machinery. The new fault prototypes are used to expand the calibrated classifier. The query set for incremental faults of rotating machinery obtains new query embeddings by sharing the basic backbone network parameters of incremental fault diagnosis of rotating machinery. A meta-training task is constructed using the support set and the query set for incremental faults of rotating machinery, and new knowledge features are calculated. The expanded calibration classifier, the new query embeddings, and the new knowledge features are input into the optimized data knowledge-driven calibration module to obtain a new calibration classifier and a new calibration fault representation, thereby realizing the diagnosis of incremental faults of rotating machinery.

[0005] The above-mentioned incremental fault diagnosis method for rotating machinery in data-constrained scenarios is characterized in that: the basic backbone network for incremental fault diagnosis of rotating machinery includes ResNet20.

[0006] The aforementioned incremental fault diagnosis method for rotating machinery in data-constrained scenarios is characterized by: in step one, using the basic training set of incremental faults in rotating machinery to train the basic backbone network for incremental fault diagnosis of rotating machinery, thereby obtaining the basic classifier. .

[0007] The aforementioned incremental fault diagnosis method for rotating machinery in data-constrained scenarios is characterized by: in step two, multiple random data samplings are performed on the basic training set of incremental faults in rotating machinery to obtain a support set. and the corresponding query set, where, For sample number and I represents the total number of samples.

[0008] The aforementioned incremental fault diagnosis method for rotating machinery in data-constrained scenarios is characterized by: supporting collections This forms a pseudo-incremental pattern with the corresponding query set, supporting set... Fault prototypes are obtained by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The basic classifier is then expanded using these fault prototypes to obtain the first expanded classifier. The corresponding query set obtains the first query embedding through the shared parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the first knowledge feature is calculated. The first extended classifier... A calibration module jointly driven by the first query embedding and the first knowledge feature input data knowledge obtains the first calibration classifier. and the first calibration fault characterization; the first calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results; then the cross-entropy loss value is calculated between the prediction results and the real labels, and the loss function is iteratively optimized through the backpropagation algorithm. Support set Fault prototypes are obtained by sharing the basic backbone network parameters for incremental fault diagnosis of rotating machinery, and the first calibration classifier is extended using these fault prototypes. The second extended classifier is obtained. The corresponding query set obtains the second query embedding through the shared parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the second knowledge features are calculated. The second extended classifier... The calibration module, jointly driven by the second query embedding and the second knowledge feature input data knowledge, obtains the second calibration classifier. Second calibration fault characterization; second calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results; then the cross-entropy loss value is calculated between the prediction results and the real labels, and the loss function is iteratively optimized through the backpropagation algorithm. And so on, supporting sets Fault prototypes are obtained by sharing the basic backbone network parameters for incremental fault diagnosis of rotating machinery. The fault prototypes are then used to expand the (I-1)th calibration classifier to obtain the I-th expanded classifier. The corresponding query set obtains the I-th query embedding by sharing the parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the I-th knowledge feature is calculated. This is then used to extend the I-th classifier. The calibration module, jointly driven by the I-th query embedding and the I-th knowledge feature input data, obtains the I-th calibration classifier. And the I-th calibration fault characterization, the I-th calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results. Then, the cross-entropy loss value is calculated between the prediction results and the real labels. The loss function is iteratively optimized through the backpropagation algorithm, thereby realizing the optimization of the calibration module driven by data knowledge.

[0009] The above-mentioned method for diagnosing incremental faults in rotating machinery under data-constrained scenarios is characterized in that: the sample knowledge in the basic training set of incremental faults of rotating machinery calculates multiple time-frequency domain statistical features of this type of incremental rotating machinery based on the prior statistical features in the time domain and frequency domain, and concatenates them into a multi-dimensional feature vector. The concatenated vector is used as the prior knowledge feature vector to guide the calibration module jointly driven by data knowledge to establish adaptive calibration between the base class and the new class.

[0010] The above-mentioned incremental fault diagnosis method for rotating machinery in data-constrained scenarios is characterized in that: the types of incremental faults of rotating machinery include inner ring pitting faults, inner ring indentation faults, inner ring defects, outer ring pitting faults, outer ring indentation faults, outer ring defects, rolling element pitting faults, rolling element indentation faults, and rolling element defects.

[0011] Compared with the prior art, the present invention has the following advantages: 1. This invention provides prior knowledge that can complement fault data, enabling performance improvements in fault diagnosis models even with limited fault data. The data-knowledge jointly driven few-shot incremental adaptive calibration fault diagnosis addresses the challenge of limited fault samples in scenarios with multi-stage incremental fault types in rotating machinery components. By introducing prior knowledge prompts, it assists the model in learning more discriminative features, reducing the data requirements for learning optimization. Secondly, considering that few-shot incremental learning is a multi-stage process, we cannot predict future faults in the basic stage. Therefore, a pseudo-incremental learning task with knowledge embedding guides the optimization task, learning invariant information between the basic and incremental stages, and preparing the model for upcoming real few-shot incremental sessions. While continuously adapting to new fault categories, it maintains good diagnostic performance for older categories, facilitating widespread use.

[0012] 2. This invention completes the adjustment of the classifier for old types of faults and the class prototype for new types of faults through a data-knowledge jointly driven calibration module, correcting the deviation of the fault representation in the incremental process model; and constructs pseudo-incremental tasks in the basic session to dynamically simulate real incremental tasks, enabling the model to learn more generalized feature embeddings and improving the model's stability and plasticity.

[0013] 3. The method of this invention has simple steps, embeds knowledge into a few-sample incremental fault diagnosis framework, continuously guides the attention mechanism to calibrate new and old fault types, helps the model learn a more discriminative classifier; enhances the model's adaptability to incremental faults, reduces the risk of overfitting, and is easy to promote and use.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2 This is a schematic diagram illustrating the incremental fault diagnosis method for few-sample classes in this invention.

[0017] Figure 3 This is a schematic diagram of the incremental fault diagnosis process for rotating machinery in data-constrained scenarios according to the present invention.

[0018] Figure 4 This is a schematic diagram of the calibration module driven by data and knowledge in this invention. Detailed Implementation

[0019] like Figures 1 to 4 As shown, the incremental fault diagnosis method for rotating machinery in data-constrained scenarios of the present invention includes the following steps: Step 1: Train the basic backbone network for incremental fault diagnosis of rotating machinery using the basic training set of incremental faults of rotating machinery to obtain the basic classifier; Step 2: Sample data from the incremental fault training set of rotating machinery to obtain the support set and query set; Step 3: Construct the meta-training task: Form a pseudo-incremental pattern by sharing the support set and query set. The support set obtains fault prototypes by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The basic classifier is then extended using the fault prototypes. The query set obtains query embeddings by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The meta-training task is constructed using the support set and query set, and knowledge features are calculated. Step 4: Optimize the data-knowledge joint-driven calibration module: Input the expanded base classifier, query embedding, and knowledge features into the data-knowledge joint-driven calibration module to obtain the calibrated classifier and the calibrated fault representation. The data-knowledge joint-driven calibration module includes an attention calibration model and a feedforward neural network with fine-tuning by using prior knowledge features as a prefix. The attention calibration model introduces a scaled dot product attention mechanism to calculate the similarity between the calibrated classifier output and the fault representation features, generating a fault prediction result. Subsequently, the prediction result is compared with the true label to calculate the cross-entropy loss value, and the loss function is iteratively optimized through the backpropagation algorithm. Step 5: Construct a support set for incremental faults from the actual newly added limited incremental fault data of rotating machinery. The test data of all fault types constitute the query set for incremental faults of rotating machinery. The support set for incremental faults of rotating machinery obtains new fault prototypes by sharing the basic backbone network parameters of incremental fault diagnosis of rotating machinery. The new fault prototypes are used to expand the calibrated classifier. The query set for incremental faults of rotating machinery obtains new query embeddings by sharing the basic backbone network parameters of incremental fault diagnosis of rotating machinery. A meta-training task is constructed using the support set and the query set for incremental faults of rotating machinery, and new knowledge features are calculated. The expanded calibration classifier, the new query embeddings, and the new knowledge features are input into the optimized data knowledge-driven calibration module to obtain a new calibration classifier and a new calibration fault representation, thereby realizing the diagnosis of incremental faults of rotating machinery.

[0020] In this embodiment, the basic backbone network for incremental fault diagnosis of rotating machinery includes ResNet20.

[0021] In this embodiment, in step one, the basic backbone network for incremental fault diagnosis of rotating machinery is trained using the basic training set of incremental faults in rotating machinery to obtain the basic classifier. .

[0022] In this embodiment, in step two, multiple random data samplings are performed on the basic training set of incremental faults in rotating machinery to obtain the support set. and the corresponding query set, where, For sample number and I represents the total number of samples.

[0023] In practice, a decoupling approach is adopted between representation and classifier. In the basic stage, a good base classifier is trained by minimizing the loss through cross-entropy using sufficient base class data. .

[0024] In this embodiment, the support set This forms a pseudo-incremental pattern with the corresponding query set, supporting set... Fault prototypes are obtained by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The basic classifier is then expanded using these fault prototypes to obtain the first expanded classifier. The corresponding query set obtains the first query embedding through the shared parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the first knowledge feature is calculated. The first extended classifier... A calibration module jointly driven by the first query embedding and the first knowledge feature input data knowledge obtains the first calibration classifier. and the first calibration fault characterization; the first calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results; then the cross-entropy loss value is calculated between the prediction results and the real labels, and the loss function is iteratively optimized through the backpropagation algorithm. Support set Fault prototypes are obtained by sharing the basic backbone network parameters for incremental fault diagnosis of rotating machinery, and the first calibration classifier is extended using these fault prototypes. The second extended classifier is obtained. The corresponding query set obtains the second query embedding through the shared parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the second knowledge features are calculated. The second extended classifier... The calibration module, jointly driven by the second query embedding and the second knowledge feature input data knowledge, obtains the second calibration classifier. Second calibration fault characterization; second calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results; then the cross-entropy loss value is calculated between the prediction results and the real labels, and the loss function is iteratively optimized through the backpropagation algorithm. And so on, supporting sets Fault prototypes are obtained by sharing the basic backbone network parameters for incremental fault diagnosis of rotating machinery. The fault prototypes are then used to expand the (I-1)th calibration classifier to obtain the I-th expanded classifier. The corresponding query set obtains the I-th query embedding by sharing the parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the I-th knowledge feature is calculated. This is then used to extend the I-th classifier. The calibration module, jointly driven by the I-th query embedding and the I-th knowledge feature input data, obtains the I-th calibration classifier. And the I-th calibration fault characterization, the I-th calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results. Then, the cross-entropy loss value is calculated between the prediction results and the real labels. The loss function is iteratively optimized through the backpropagation algorithm, thereby realizing the optimization of the calibration module driven by data knowledge.

[0025] In practice, the classifier is continuously expanded. A data-knowledge jointly driven calibration module adjusts the classifiers for old fault types and the class prototypes for new fault types, correcting the model's deviation in fault representation during the incremental process. This allows for the simulation of incremental fault scenarios to cultivate the model's generalization ability. When real fault data arrives, knowledge prompts help the model recall relevant information, guiding it to output correct diagnostic results. In the basic stage, pseudo-incremental tasks are sampled to simulate scenarios where fault types are constantly increasing—that is, constructing a dynamic scenario with multiple incremental tasks in the static basic stage. If the model performs well on these pseudo-incremental tasks, this ability can be extended to real incremental tasks. In other words, the model learns more generalized representations through pseudo-incremental tasks and can handle newly occurring fault types well.

[0026] In this embodiment, the sample knowledge in the basic training set of rotating machinery incremental faults calculates multiple time-frequency domain statistical features of this type of rotating machinery increment based on the time-domain and frequency-domain prior statistical features, and concatenates them into a multi-dimensional feature vector. The concatenated vector serves as the prior knowledge feature vector, guiding the calibration module driven by data knowledge to establish adaptive calibration between the base class and the new class.

[0027] In this embodiment, the types of incremental faults in rotating machinery include inner ring pitting faults, inner ring indentation faults, inner ring defects, outer ring pitting faults, outer ring indentation faults, outer ring defects, rolling element pitting faults, rolling element indentation faults, and rolling element defects.

[0028] It should be noted that prior knowledge can provide features that complement fault data, thereby improving the performance of fault diagnosis models when fault types are increasing and fault data is limited. Data-knowledge joint-driven few-sample incremental adaptive calibration fault diagnosis addresses the challenge of limited fault samples in scenarios with multi-stage incremental fault types in rotating machinery components. By introducing prior knowledge prompts, the model is assisted in learning more discriminative features, reducing the data requirements for learning optimization. Secondly, considering that few-sample incremental learning is a multi-stage process, we cannot predict future faults in the basic stage. To address this, a pseudo-incremental learning task using knowledge embedding is employed to guide task optimization, learning invariant information between the basic and incremental stages, and preparing the model for upcoming real few-shot incremental sessions. This allows the model to continuously adapt to new fault categories while maintaining good diagnostic performance for older categories. A data-knowledge jointly driven calibration module adjusts the classifiers for older fault categories and the class prototypes for newer fault categories, correcting the model's deviation in fault representation during the incremental process. A pseudo-incremental task is constructed in the basic session to dynamically simulate real incremental tasks, enabling the model to learn more generalized feature embeddings and improving its stability and adaptability. The knowledge embedding few-shot incremental fault diagnosis framework continuously guides the attention mechanism to calibrate new and old fault types, helping the model learn a more discriminative classifier. This enhances the model's adaptability to incremental faults and reduces the risk of overfitting.

[0029] In this embodiment, the scaling dot product attention mechanism uses Softmax as the activation function.

[0030] In this embodiment, 14 commonly used time-frequency domain statistical features for rotating machinery fault diagnosis are selected as carriers of prior knowledge. This prior knowledge has been proven effective and readily available; their combination can describe relevant rotating machinery faults. For example, the root mean square (RMS) feature is generally related to the overall vibration energy of the machine, and an increase may indicate mechanical wear. The names and expressions of these prior statistical features are shown in Table 1.

[0031] Table 1

[0032] in, The signal sequence representing the faulty sample, , Indicates the length of the signal sequence, i.e., contains Data points. express The spectrum, Indicates the number of spectral lines. Indicates the first The frequency values ​​of each spectral line. Based on the fault samples collected during the incremental process, we can calculate 14 time-frequency domain statistical features for each type according to Table 1, and concatenate them into a 14-dimensional feature vector. The concatenated vector is used... express: .we will As a prior knowledge vector, it guides the calibration module to establish a suitable calibration between the base class and the new class. The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for incremental fault diagnosis of rotating machinery in data-constrained scenarios, characterized in that, The method includes the following steps: Step 1: Train the basic backbone network for incremental fault diagnosis of rotating machinery using the basic training set of incremental faults of rotating machinery to obtain the basic classifier; Step 2: Sample data from the incremental fault training set of rotating machinery to obtain the support set and query set; Step 3: Construct the meta-training task: Form a pseudo-incremental pattern by sharing the support set and query set. The support set obtains fault prototypes by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The basic classifier is then extended using the fault prototypes. The query set obtains query embeddings by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The meta-training task is constructed using the support set and query set, and knowledge features are calculated. Step 4: Optimize the data-knowledge joint-driven calibration module: Input the expanded base classifier, query embedding, and knowledge features into the data-knowledge joint-driven calibration module to obtain the calibrated classifier and the calibrated fault representation. The data-knowledge joint-driven calibration module includes an attention calibration model and a feedforward neural network with fine-tuning by using prior knowledge features as a prefix. The attention calibration model introduces a scaled dot product attention mechanism to calculate the similarity between the calibrated classifier output and the fault representation features, generating a fault prediction result. Subsequently, the prediction result is compared with the true label to calculate the cross-entropy loss value, and the loss function is iteratively optimized through the backpropagation algorithm. Step 5: Construct a support set for incremental faults from the actual newly added limited incremental fault data of rotating machinery. The test data of all fault types constitute the query set for incremental faults of rotating machinery. The support set for incremental faults of rotating machinery obtains new fault prototypes by sharing the basic backbone network parameters of incremental fault diagnosis of rotating machinery. The new fault prototypes are used to expand the calibrated classifier. The query set for incremental faults of rotating machinery obtains new query embeddings by sharing the basic backbone network parameters of incremental fault diagnosis of rotating machinery. A meta-training task is constructed using the support set and the query set for incremental faults of rotating machinery, and new knowledge features are calculated. The expanded calibration classifier, the new query embeddings, and the new knowledge features are input into the optimized data knowledge-driven calibration module to obtain a new calibration classifier and a new calibration fault representation, thereby realizing the diagnosis of incremental faults of rotating machinery.

2. The incremental fault diagnosis method for rotating machinery in data-constrained scenarios according to claim 1, characterized in that: The basic backbone network for incremental fault diagnosis of rotating machinery includes Resnet20.

3. The incremental fault diagnosis method for rotating machinery in data-constrained scenarios according to claim 1, characterized in that: In step one, the basic backbone network for incremental fault diagnosis of rotating machinery is trained using the basic training set of incremental faults in rotating machinery to obtain the basic classifier. .

4. The incremental fault diagnosis method for rotating machinery in data-constrained scenarios according to claim 3, characterized in that: In step two, multiple random data samplings are performed on the incremental fault training set of rotating machinery to obtain the support set. and the corresponding query set, where, For sample number and I represents the total number of samples.

5. The incremental fault diagnosis method for rotating machinery in data-constrained scenarios according to claim 3, characterized in that: Support set This forms a pseudo-incremental pattern with the corresponding query set, supporting set... Fault prototypes are obtained by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The basic classifier is then expanded using these fault prototypes to obtain the first expanded classifier. The corresponding query set obtains the first query embedding by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The meta-training task is constructed by the support set and the corresponding query set, and the first knowledge feature is calculated. The first extended classifier A calibration module jointly driven by the first query embedding and the first knowledge feature input data knowledge obtains the first calibration classifier. and the first calibration fault characterization; the first calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results; then the cross-entropy loss value is calculated between the prediction results and the real labels, and the loss function is iteratively optimized through the backpropagation algorithm. Support set Fault prototypes are obtained by sharing the basic backbone network parameters for incremental fault diagnosis of rotating machinery, and the first calibration classifier is extended using these fault prototypes. The second extended classifier is obtained. The corresponding query set obtains the second query embedding by sharing the parameters of the basic backbone network for incremental fault diagnosis of rotating machinery. The meta-training task is constructed by the support set and the corresponding query set, and the second knowledge features are calculated. The second extended classifier The calibration module, jointly driven by the second query embedding and the second knowledge feature input data knowledge, obtains the second calibration classifier. Second calibration fault characterization; second calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results; then the cross-entropy loss value is calculated between the prediction results and the real labels, and the loss function is iteratively optimized through the backpropagation algorithm. And so on, supporting sets Fault prototypes are obtained by sharing the basic backbone network parameters for incremental fault diagnosis of rotating machinery. The fault prototypes are then used to expand the (I-1)th calibration classifier to obtain the I-th expanded classifier. The corresponding query set obtains the I-th query embedding by sharing the parameters of the rotating machinery incremental fault diagnosis backbone network. A meta-training task is constructed using the support set and the corresponding query set, and the I-th knowledge feature is calculated. This is then used to extend the I-th classifier. The calibration module, jointly driven by the I-th query embedding and the I-th knowledge feature input data, obtains the I-th calibration classifier. And the I-th calibration fault characterization, the I-th calibration classifier The similarity between the output and the fault characterization features is calculated to generate fault prediction results. Then, the cross-entropy loss value is calculated between the prediction results and the real labels. The loss function is iteratively optimized through the backpropagation algorithm, thereby realizing the optimization of the calibration module driven by data knowledge.

6. The incremental fault diagnosis method for rotating machinery in data-constrained scenarios according to claim 1, characterized in that: The sample knowledge in the basic training set of incremental faults of rotating machinery calculates multiple time-frequency domain statistical features of this type of incremental rotating machinery based on the prior statistical features in the time and frequency domains, and concatenates them into a multi-dimensional feature vector. The concatenated vector serves as the prior knowledge feature vector, guiding the calibration module driven by data knowledge to establish adaptive calibration between the base class and the new class.

7. The incremental fault diagnosis method for rotating machinery in data-constrained scenarios according to claim 1, characterized in that: The types of incremental faults in rotating machinery include inner ring pitting faults, inner ring indentation faults, inner ring defects, outer ring pitting faults, outer ring indentation faults, outer ring defects, rolling element pitting faults, rolling element indentation faults, and rolling element defects.

Images