A method for diagnosing an imbalance domain increment of a mechanical fault type

Abstract

This invention provides a method for incremental diagnosis of mechanical fault class imbalance domains. The method includes: collecting mechanical vibration signals using sensors, reconstructing dimensions to generate a two-dimensional feature matrix, and dividing the samples into training and test sets; constructing a diagnostic model, including a frequency-sensing network module and a dynamic weight selection module; during prediction, the model extracts features and inputs them into a frequency-sensing library to obtain prediction opinions, and the dynamic weight selection module performs weighted integration and outputs the predictions; after training is completed in each incremental stage, the diagnostic accuracy of the model on known fault categories is calculated; the above process is iteratively executed to complete all incremental stages and construct a sustainable diagnostic model. The incremental diagnostic model for mechanical fault class imbalance domains of this invention, through the prediction of the frequency-sensing module and the training of the dynamic weight selection module, can improve the diagnostic accuracy of the model while avoiding catastrophic forgetting in the case of continuously added faults and severe imbalance of fault categories.

Description

Technical Field

[0001] This invention relates to the field of mechanical intelligent fault diagnosis technology, and in particular to a method for incremental diagnosis of mechanical fault-type unbalanced domains. Background Technology

[0002] In the field of mechanical fault diagnosis, existing technologies mainly revolve around three major directions: signal processing, analytical modeling, and data-driven approaches. Traditional signal processing-based methods collect physical signals such as vibration and noise during equipment operation, and extract characteristic frequencies using techniques such as spectrum analysis and wavelet transform to identify common mechanical fault types. These methods rely on expert experience and prior knowledge, and while they have good interpretability under steady-state conditions, their robustness in feature extraction is often limited under complex dynamic conditions.

[0003] Analytical model-based methods focus on constructing an accurate mathematical model of the system, and then use the residuals between theoretical outputs and actual monitoring data to detect and locate faults. While this method offers high diagnostic accuracy when the model is accurate, real-world industrial systems often exhibit strong nonlinearity and multivariate coupling, making it difficult to establish a complete mathematical model. Furthermore, the high cost of model updates and maintenance limits its application in large-scale, complex equipment.

[0004] With the development of industrial big data and artificial intelligence technologies, data-driven intelligent diagnostic methods are gradually becoming mainstream. These methods utilize algorithms such as machine learning and deep learning to automatically mine fault features and patterns directly from historical monitoring data, avoiding the limitations of manual feature design and significantly improving the accuracy and generalization ability of complex fault identification. However, existing data-driven methods typically rely on large amounts of labeled data, and once the model is trained, it is difficult to adapt to new fault types or changing operating conditions without retraining.

[0005] In recent years, large-scale pre-trained models have demonstrated powerful feature learning and cross-domain transfer capabilities in fields such as natural language processing and vision, providing new technical pathways for industrial fault diagnosis. How to leverage the general representation capabilities of pre-trained models to achieve continuous adaptive learning of new fault categories and dynamic operating conditions without needing to backtrack to the original data and only through minor parameter fine-tuning has become a key challenge and important research direction in the field of incremental industrial fault diagnosis. Summary of the Invention

[0006] To address the above issues, this invention constructs a frequency-aware network module and a dynamic weight selection module on the basis of the pre-trained ResNet model. By using a parallel expert structure to optimize the feature representation of different sample frequency categories, it effectively alleviates intra-domain class imbalance. Furthermore, it designs a pseudo-sample generation mechanism and a dynamic weight selection module based on historical feature statistics to achieve adaptive fusion and transfer of cross-domain knowledge without relying on historical real data, significantly mitigating catastrophic forgetting caused by class distribution drift.

[0007] According to an embodiment of the present invention, a method for incremental diagnosis of unbalanced domains in mechanical faults is provided.

[0008] In a first aspect of the invention, a method for incremental diagnosis of imbalance domains in mechanical fault types is provided. The method includes: Step S01: Use an accelerometer to collect vibration signals of the machine, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into a training set and a test set for the current task; Step S02: Construct a diagnostic model, including a frequency-aware network module and a dynamic weight selection module; Step S03: Train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task. Step S04: Based on the historical task knowledge stored in the frequency-aware library, generate a synthetic pseudo-feature dataset and divide it into a training set and a test set for the historical task. Use the training set of the historical task to train the weight dynamic selection module. Step S05: Input the test set into the frequency sensing library to obtain diverse prediction views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing prediction views in real time, and outputs the final prediction through weighted integration; Step S06: Use the test set of the current task and the test set of the historical tasks to comprehensively evaluate the diagnostic model. Calculate the overall diagnostic accuracy of the diagnostic model on all learned fault categories and quantitatively analyze its knowledge forgetting degree and ability to identify new categories. Step S07: Update the data and model of the current task to the historical database; when new fault type data arrives, repeat steps S03 to S06 until all incremental stages are completed, and finally build a sustainable diagnostic model that can identify all fault types.

[0009] Furthermore, the specific steps of step S01 are as follows: After collecting the vibration signal of the rolling bearing during mechanical operation, it is divided into 4096 equal-length samples. The wave structure and pattern in the vibration signal are mapped into a two-dimensional three-channel image with clear texture features. Finally, the mapped three-channel image of each sample is bilinearly interpolated to generate the final feature fusion map, thus forming a complete dataset.

[0010] Furthermore, the frequency-aware network module described in step S02 consists of a pre-trained model and a frequency-aware module: the pre-trained model adopts the ResNet-50 architecture, which consists of 16 cascaded residual blocks. Each residual block contains a convolutional layer, a batch normalization layer, and an activation function, and achieves feature identity mapping and residual learning through skip connections; after the 3×3 deep convolutional layers in all 16 residual blocks of ResNet-50, a pluggable adapter module is integrated. This adapter module contains a dimension reduction projection layer, a non-linear activation function, and a dimension increase projection layer.

[0011] Furthermore, in step S03, when training the frequency-aware network module using the training set of the current task, the pre-trained parameters of the backbone network ResNet-50 are kept frozen, and only the weight parameters of the inserted adapter module are updated.

[0012] Furthermore, the frequency sensing module includes a majority class perceptron, a balance class perceptron, and a minority class perceptron. The majority class perceptron is the basic classifier. It uses standard cross-entropy loss; the balanced perceptron is a balanced classifier. The class weights are balanced by adding the logarithm of class frequency to the logits; the minority class perceptron, also known as the inverse classifier, gives higher attention to the minority class; the losses of the above three classifiers... , , The calculation formulas are as follows: , , , Where y is the true class label of the current training sample, Y is the label space, j is the class index, and b is the task domain. It is the probability of class X appearing in domain b.

[0013] Furthermore, the specific steps for generating the synthetic pseudo-feature dataset based on the historical task knowledge stored in the frequency-aware library, as described in step S04, are as follows: Step S041: During the domain incremental learning process, each target category in the historical task knowledge is modeled as a Gaussian distribution. Cross-domain feature statistics are continuously collected and updated. After training the frequency-aware network module for domain b, the corresponding set of feature statistics is calculated for each category c. ,in, and These are the mean vector and covariance matrix of category c in the current domain b, respectively; Step S042: Integrate the statistics from step S041 into the global statistical knowledge base. In this process, progressive modeling and maintenance of cross-domain feature space evolution structures are achieved; Step S043: In the optimization phase of the expert selector, to facilitate cross-domain knowledge transfer, for each domain-category pair (b, c) in the global statistical knowledge base G, Gaussian distribution-based feature sampling is performed to generate K synthetic features and construct a balanced synthetic dataset. , in, For indexing a domain, This represents a synthetic sample obtained by sampling from the corresponding distribution. It is a synthesized feature vector.

[0014] Further, the specific steps of step S05 are as follows: extract the feature vector of the test sample through the frozen pre-trained encoder, input it in parallel into the frequency perception library to obtain diverse category score outputs, and input it into the dynamic weight selection module to generate the fusion weights corresponding to each perceptron; finally, obtain the final category prediction of the sample by weighted fusion of the output scores of all perceptrons and Softmax normalization.

[0015] In a second aspect of the invention, an apparatus for incremental diagnosis of imbalance domains in mechanical fault types is provided. The apparatus includes: Current task acquisition module: used to acquire vibration signals of machinery using accelerometers, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into training and test sets for the current task; Diagnostic model building module: Used to build diagnostic models, including a frequency-aware network module and a dynamic weight selection module; The first training module of the diagnostic model is used to train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task. The second training module of the diagnostic model is used to generate a synthetic pseudo-feature dataset based on the historical task knowledge stored in the frequency-aware library and divide it into training and test sets for historical tasks. The training set of historical tasks is used to train the weight dynamic selection module. Predictive view calculation module: It is used to input the test set into the frequency awareness library to obtain diverse predictive views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing predictive views in real time, and outputs the final prediction through weighted integration; Diagnostic Model Evaluation Module: This module is used to comprehensively evaluate the diagnostic model using the test sets of the current task and the test sets of historical tasks. It calculates the overall diagnostic accuracy of the diagnostic model on all learned fault categories and quantitatively analyzes its knowledge forgetting degree and ability to identify new categories. Sustainable diagnostic model building module: used to update the data and model of the current task to the historical database; when new fault type data arrives, the second training module of the diagnostic model is repeatedly executed to the evaluation module of the diagnostic model until all incremental stages are completed, and finally a sustainable diagnostic model that can identify all fault types is built.

[0016] In a third aspect of the invention, an electronic device is provided. The electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the program to implement the method according to a first aspect of the invention.

[0017] In a fourth aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method according to a first aspect of the invention.

[0018] This invention constructs a frequency-aware network module and a dynamic weight selection module on the basis of the pre-trained ResNet model, and optimizes the feature representation of different sample frequency categories through parallel expert structures, effectively alleviating intra-domain class imbalance. It also designs a pseudo-sample generation mechanism and a dynamic weight selection module based on historical feature statistics, which realizes adaptive fusion and transfer of cross-domain knowledge without relying on historical real data, significantly mitigating catastrophic forgetting caused by class distribution drift.

[0019] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0020] The beneficial effects of this invention are: 1. Through modular design, including frequency-aware network modules and parallel expert structures, it naturally adapts to imbalanced data distribution, effectively alleviates the problem of class imbalance within the domain, and promotes the generalization ability of minority classes while preventing forgetting of the majority class, thereby improving the accuracy and robustness of fault diagnosis. 2. By designing a pseudo-sample generation mechanism based on historical feature statistics and a dynamic weight selection module, this method can achieve adaptive fusion and transfer of cross-domain knowledge without relying on historical real data, solve the catastrophic forgetting problem caused by category distribution drift, and ensure the continuity of knowledge in the incremental learning process. 3. The overall framework keeps the pre-trained backbone network frozen, optimizing only a few parameters, such as the frequency sensing module and the weight selection module. This achieves stability in incremental diagnosis while also ensuring lightweight deployment, reducing computational and storage overhead, and facilitating application in real-world industrial scenarios. Attached Figure Description

[0021] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart of a method for incremental diagnosis of mechanical fault type imbalance domain according to an embodiment of the present invention is shown; Figure 2 A diagram illustrating the structure of a pre-trained model according to an embodiment of the present invention is shown. Figure 3 A schematic diagram of a diagnostic model structure according to an embodiment of the present invention is shown; Figure 4 This diagram illustrates a structure for applying a trained diagnostic model to a test sample set for validation, according to an embodiment of the present invention. Figure 5 A block diagram of an apparatus for incremental diagnosis of mechanical fault type imbalance domain according to an embodiment of the present invention is shown; Figure 6 A schematic diagram of an apparatus for incremental diagnosis of mechanical fault type imbalance domain according to an embodiment of the present invention is shown. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] According to an embodiment of the present invention, a method for incremental diagnosis of imbalanced domains of mechanical faults is proposed. By constructing a frequency-aware network module and a dynamic weight selection module on the basis of the pre-trained ResNet model, and optimizing the feature representation of different sample frequency categories through a parallel expert structure, the class imbalance within the domain is effectively alleviated. A pseudo-sample generation mechanism and a dynamic weight selection module based on historical feature statistics are designed to achieve adaptive fusion and transfer of cross-domain knowledge without relying on historical real data, which significantly alleviates the catastrophic forgetting caused by class distribution drift.

[0024] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0025] Figure 1 This is a schematic flowchart of a method for incremental diagnosis of mechanical fault-type imbalance domain according to an embodiment of the present invention. The method includes: Step S01: Use an accelerometer to collect vibration signals of the machine, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into a training set and a test set for the current task; Step S02: Construct a diagnostic model, including a frequency-aware network module and a dynamic weight selection module; Step S03: Train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task. Step S04: Based on the historical task knowledge stored in the frequency-aware library, generate a synthetic pseudo-feature dataset and divide it into a training set and a test set for the historical task. Use the training set of the historical task to train the weight dynamic selection module. Step S05: Input the test set into the frequency sensing library to obtain diverse prediction views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing prediction views in real time, and outputs the final prediction through weighted integration; Step S06: Use the test set of the current task and the test set of the historical tasks to comprehensively evaluate the diagnostic model. Calculate the overall diagnostic accuracy of the diagnostic model on all learned fault categories and quantitatively analyze its knowledge forgetting degree and ability to identify new categories. Step S07: Update the data and model of the current task to the historical database; when new fault type data arrives, repeat steps S03 to S06 until all incremental stages are completed, and finally build a sustainable diagnostic model that can identify all fault types.

[0026] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0027] To provide a clearer explanation of the above-mentioned method for incremental diagnosis of imbalance domain in mechanical faults, a specific embodiment will be used for illustration below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0028] The following specific example will further illustrate the method of incremental diagnosis of imbalance domain in mechanical faults: Step S01: Use an accelerometer to collect vibration signals of the machine, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into a training set and a test set for the current task.

[0029] Specifically, the vibration signal of the rolling bearing during mechanical operation is collected and segmented into 4096 equal-length samples. Then, the wave structure and pattern in the vibration signal are mapped into a two-dimensional three-channel image with clear texture features, preserving the topological relationship and key features of the original vibration signal in the time and frequency domains. Finally, bilinear interpolation is performed on the mapped three-channel image of each sample to generate the final feature fusion map, thus forming a complete dataset for model training and testing.

[0030] It should be noted that the input is a vibration signal sample collected by an accelerometer, with a length of 4096. The size of each sample after feature fusion is 3×32×128.

[0031] In this embodiment, a mechanical vibration signal dataset is used as an example, which includes five types of faults (represented by labels 0-4): healthy, intake valve wear, exhaust valve wear, spring failure, and a combined spring failure and exhaust valve wear fault. These faults are operated at five different speeds: 8Hz, 12Hz, 16Hz, 20Hz, and 8-22Hz. The experiment divides the five operating conditions into one initial task and four incremental tasks. The number of samples for each fault state is unbalanced, ranging from 10 to 300, and each sample's monitored vibration signal contains 4096 sampling points. A description of the dataset used is shown in Table 1.

[0032] Table 1 Step S02: Construct a diagnostic model, including a frequency-aware network module and a dynamic weight selection module.

[0033] Specifically, the frequency-aware network module consists of a pre-trained model and a frequency-sensing module. The pre-trained model adopts the ResNet architecture, which comprises multiple cascaded residual blocks. Each residual block contains convolutional layers, batch normalization layers, and activation functions, and achieves feature identity mapping and residual learning through skip connections. To enhance the model's transfer learning capability and preserve pre-trained parameters, pluggable adapter modules are integrated after some convolutional layers. These adapter modules contain dimensionality reduction projection layers, non-linear activation functions, and dimensionality increase projection layers. During training, the pre-trained parameters of the backbone ResNet are kept frozen, and only the weight parameters of the inserted adapter modules are updated, achieving compact storage of task-specific features and providing a knowledge base for subsequent incremental learning tasks.

[0034] In this embodiment, as Figure 2 As shown, the pre-trained model adopts the ResNet-50 architecture, which consists of 16 cascaded residual blocks. Each residual block contains a convolutional layer, a batch normalization layer, and an activation function. The feature identity mapping and residual learning are achieved through skip connections. After the 3×3 deep convolutional layers in all 16 residual blocks of ResNet-50, a pluggable adapter module is integrated. This adapter module contains a dimension reduction projection layer, a non-linear activation function, and a dimension up projection layer.

[0035] Step S03: Train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task.

[0036] Specifically, the frequency sensing module includes a majority class perceptron, a balanced class perceptron, and a minority class perceptron, which are actually three parallel classifiers. The majority class perceptron is the basic classifier. Using standard cross-entropy loss, it provides basic classification predictions, but does not consider class frequency differences, thus favoring the majority class; the balanced class perceptron is a balanced classifier. The class weights are balanced by adding the logarithm of the class frequency to the logits, resulting in a balanced prediction that considers the differences in the number of samples from each class. The minority class perceptron, also known as the inverse classifier, gives higher attention to the minority class. The losses of the above three classifiers... , , The calculation formulas are as follows: , , , Where y is the true class label of the current training sample. Y It is a tag space.j It is a category index. b For the task domain, It is a domain b Medium X The probability of occurrence.

[0037] The overall goal of the frequency-aware network module is the sum of three losses: .

[0038] Step S04: Based on the historical task knowledge stored in the frequency-aware library, generate a synthetic pseudo-feature dataset and divide it into a training set and a test set for the historical task. Use the training set of the historical task to train the weight dynamic selection module.

[0039] Specifically, the steps for generating a synthetic pseudo-feature dataset based on historical task knowledge stored in the frequency-aware library are as follows: Step S041: During the domain incremental learning process, each target category in the historical task knowledge is modeled as a Gaussian distribution, and cross-domain feature statistics are continuously collected and updated. This process is repeated after completing the domain incremental learning. b After training the frequency-aware network module, for each category... c Calculate the corresponding set of characteristic statistics. ,in, and Categories c In the current domain b The mean vector and covariance matrix in the matrix; Step S042: Integrate the statistics from step S041 into the global statistical knowledge base. In this process, progressive modeling and maintenance of cross-domain feature space evolution structures are achieved; Step S043: In the optimization phase of the expert selector, to promote cross-domain knowledge transfer, the global statistical knowledge base is... G Each group of domain-category pairs ( b , c Perform Gaussian distribution-based feature sampling to generate K 1 synthetic feature, construct a balanced synthetic dataset: , in, For indexing a domain, This represents a synthetic sample obtained by sampling from the corresponding distribution. It is a synthesized feature vector; In this embodiment, the incremental diagnostic model for the unbalanced domain of rotating equipment faults undergoes 30 iterations in the first stage and 10 iterations in the second stage, with a learning rate of 0.01. Training stops when the set number of iterations is reached. The structure of the diagnostic training model is as follows: Figure 3As shown.

[0040] Step S05: Input the test set into the frequency sensing library to obtain diverse prediction views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing prediction views in real time, and outputs the final prediction through weighted integration.

[0041] Specifically, feature vectors of test samples are extracted using a frozen pre-trained encoder, and these vectors are input in parallel into a frequency-aware library to obtain diverse class score outputs. Simultaneously, these vectors are input into a dynamic weight selection module to generate fusion weights for each perceptron. Finally, the output scores of all perceptrons are weighted and fused, and then normalized using Softmax to obtain the final class prediction for the sample. The trained diagnostic model is as follows: Figure 4 As shown.

[0042] The loss function calculation formula for the dynamic weight selection module is as follows: , in, To select the first The probability of an expert in a particular field. For the first An expert is processing the sample. The expert assessment value at that time.

[0043] Step S06: Use the test set of the current task and the test set of the historical tasks to comprehensively evaluate the diagnostic model. Calculate the overall diagnostic accuracy of the diagnostic model on all learned fault categories, and quantitatively analyze its knowledge forgetting degree and ability to identify new categories.

[0044] Specifically, such as Figure 4 As shown, an evaluation-optimization-re-evaluation strategy is adopted to evaluate the trained diagnostic model. The first stage is initial task learning evaluation, which verifies the model's evaluation accuracy for new fault types through real test samples, and outputs the evaluation accuracy for all learned tasks. The second stage is optimization re-evaluation, which optimizes the evaluation accuracy for all learned tasks by training the weight dynamic selection module through the generated synthetic pseudo-feature dataset.

[0045] Step S07: Update the data and model of the current task to the historical database; when new fault type data arrives, repeat steps S03 to S06 until all incremental stages are completed, and finally build a sustainable diagnostic model that can identify all fault types, thereby enhancing the mechanical fault diagnosis capability.

[0046] According to the incremental fault diagnosis test results of the unbalanced domain shown in Table 2, the diagnostic accuracy of the method of the present invention reached 97.92% in the initial fault category. As the fault category gradually increases to the T4 stage, the accuracy remains stable at 82.06%. Experimental data show that: (1) the model accuracy decay rate is controlled at 15.86% during the process of adding three new fault categories; (2) in the environment of severe fault category imbalance and changing operating conditions, the model can make the accuracy decay curve show a convergence trend, which verifies its core advantage of resisting catastrophic forgetting.

[0047] Table 2 Based on the same inventive concept, this invention also proposes a device for incremental diagnosis of imbalance domains in mechanical fault types. The implementation of this device can be found in the implementation of the method described above; repeated details will not be elaborated further. Figure 5 As shown, the device 100 includes: Current task acquisition module 101: used to acquire vibration signals of machinery using an accelerometer, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into training set and test set for the current task; Diagnostic model construction module 102: used to construct diagnostic models, including a frequency-aware network module and a weight dynamic selection module; The first training module 103 of the diagnostic model is used to train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task. The second training module 104 of the diagnostic model is used to generate a synthetic pseudo-feature dataset based on the historical task knowledge stored in the frequency-aware library and divide it into a training set and a test set for the historical task. The training set of the historical task is used to train the weight dynamic selection module. Predictive view calculation module 105: It is used to input the test set into the frequency sensing library to obtain diverse predictive views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing predictive views in real time, and outputs the final prediction through weighted integration; Diagnostic Model Evaluation Module 106: Used to comprehensively evaluate the diagnostic model using the test set of the current task and the test set of the historical task. By calculating the overall diagnostic accuracy of the diagnostic model on all learned fault categories, it quantitatively analyzes the degree of knowledge forgetting and the ability to identify new categories. Sustainable diagnostic model building module 107: used to update the data and model of the current task to the historical database; when new fault type data arrives, the second training module of the diagnostic model to the evaluation module of the diagnostic model are repeatedly executed until all incremental stages are completed, and finally a sustainable diagnostic model that can identify all fault types is built.

[0048] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0049] like Figure 6 As shown, the device includes a central processing unit (CPU), which can perform various appropriate actions and processes based on computer program instructions stored in read-only memory (ROM) or loaded from storage units into random access memory (RAM). The RAM can also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0050] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0051] The processing unit executes the various methods and processes described above, such as method steps S01 to S07. For example, in some embodiments, method steps S01 to S07 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of method steps S01 to S07 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute method steps S01 to S07 by any other suitable means (e.g., by means of firmware).

[0052] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0053] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0054] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0055] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0056] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for incremental diagnosis of imbalance domain in mechanical faults, characterized in that, The method includes: Step S01: Use an accelerometer to collect vibration signals of the machine, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into a training set and a test set for the current task; Step S02: Construct a diagnostic model, including a frequency-aware network module and a dynamic weight selection module; Step S03: Train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task. Step S04: Based on the historical task knowledge stored in the frequency-aware library, generate a synthetic pseudo-feature dataset and divide it into a training set and a test set for the historical task. Use the training set of the historical task to train the weight dynamic selection module. Step S05: Input the test set into the frequency sensing library to obtain diverse prediction views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing prediction views in real time, and outputs the final prediction through weighted integration; Step S06: Use the test set of the current task and the test set of the historical tasks to comprehensively evaluate the diagnostic model. Calculate the overall diagnostic accuracy of the diagnostic model on all learned fault categories and quantitatively analyze its knowledge forgetting degree and ability to identify new categories. Step S07: Update the data and model of the current task to the historical database; when new fault type data arrives, repeat steps S03 to S06 until all incremental stages are completed, and finally build a sustainable diagnostic model that can identify all fault types.

2. The method for incremental diagnosis of mechanical fault-type imbalance domains according to claim 1, characterized in that, The specific steps of step S01 are as follows: After collecting the vibration signal of the rolling bearing during mechanical operation, it is divided into 4096 equal-length samples. The wave structure and pattern in the vibration signal are mapped into a two-dimensional three-channel image with clear texture features. Finally, the mapped three-channel image of each sample is bilinearly interpolated to generate the final feature fusion map, thus forming a complete dataset.

3. The method for incremental diagnosis of mechanical fault-type imbalance domains according to claim 1, characterized in that, The frequency-aware network module described in step S02 consists of a pre-trained model and a frequency-aware module: the pre-trained model adopts the ResNet-50 architecture, which consists of 16 cascaded residual blocks. Each residual block contains a convolutional layer, a batch normalization layer and an activation function, and achieves feature identity mapping and residual learning through skip connections. A pluggable adapter module is integrated after the 3×3 depth convolutional layer inside all 16 residual blocks of ResNet-50. This adapter module contains a dimension reduction projection layer, a non-linear activation function, and an dimension increase projection layer.

4. The method for incremental diagnosis of mechanical fault-type imbalance domains according to claim 1, characterized in that, When training the frequency-aware network module using the training set of the current task as described in step S03, the pre-trained parameters of the backbone network ResNet-50 are kept frozen, and only the weight parameters of the inserted adapter modules are updated.

5. The method for incremental diagnosis of mechanical fault-type imbalance domains according to claim 3, characterized in that, The frequency sensing module includes a majority class perceptron, a balance class perceptron, and a minority class perceptron. The majority class perceptron is the basic classifier. It uses standard cross-entropy loss; the balanced perceptron is a balanced classifier. The class weights are balanced by adding the logarithm of class frequency to the logits; the minority class perceptron, also known as the inverse classifier, gives higher attention to the minority class; the losses of the above three classifiers... , , The calculation formulas are as follows: ， ， ， Where y is the true class label of the current training sample. Y It is a tag space. j It is a category index. b For the task domain, It is a domain b Medium X The probability of occurrence.

6. The method for incremental diagnosis of mechanical fault-type imbalance domains according to claim 1, characterized in that, The specific steps for generating the synthetic pseudo-feature dataset based on the historical task knowledge stored in the frequency-aware library, as described in step S04, are as follows: Step S041: During the domain incremental learning process, each target category in the historical task knowledge is modeled as a Gaussian distribution, and cross-domain feature statistics are continuously collected and updated. This process is repeated after completing the domain incremental learning. b After training the frequency-aware network module, for each category... c Calculate the corresponding set of characteristic statistics. ,in, and Categories c In the current domain b The mean vector and covariance matrix in the matrix; Step S042: Integrate the statistics from step S041 into the global statistical knowledge base. In this process, progressive modeling and maintenance of cross-domain feature space evolution structures are achieved; Step S043: In the optimization phase of the expert selector, to promote cross-domain knowledge transfer, the global statistical knowledge base is... G Each group of domain-category pairs ( b , c Perform Gaussian distribution-based feature sampling to generate K 1 synthetic feature, construct a balanced synthetic dataset: ， in, For indexing a domain, This represents a synthetic sample obtained by sampling from the corresponding distribution. It is a synthesized feature vector.

7. The method for incremental diagnosis of mechanical fault-type imbalance domains according to claim 1, characterized in that, The specific steps of step S05 are as follows: extract the feature vector of the test sample through the frozen pre-trained encoder, input it in parallel into the frequency perception library to obtain diverse category score outputs, and input it into the dynamic weight selection module to generate the fusion weights of each perceptron; finally, obtain the final category prediction of the sample by weighted fusion of the output scores of all perceptrons and Softmax normalization.

8. A device for incremental diagnosis of imbalance domain in mechanical faults, characterized in that, The device implements the method as described in any one of claims 1 to 7, comprising: Current task acquisition module: used to acquire vibration signals of machinery using accelerometers, generate a two-dimensional feature matrix through dimensional reconstruction, and randomly divide it into training and test sets for the current task; Diagnostic model building module: Used to build diagnostic models, including a frequency-aware network module and a dynamic weight selection module; The first training module of the diagnostic model is used to train the frequency-aware network module using the training set of the current task, output the predicted score of the current task, and after training, save the network parameters and extracted features to the frequency-aware library as a knowledge archive for the task. The second training module of the diagnostic model is used to generate a synthetic pseudo-feature dataset based on the historical task knowledge stored in the frequency-aware library and divide it into training and test sets for historical tasks. The training set of historical tasks is used to train the weight dynamic selection module. Predictive view calculation module: It is used to input the test set into the frequency awareness library to obtain diverse predictive views from different tasks; the trained weight dynamic selection module calculates the optimal weight for fusing predictive views in real time, and outputs the final prediction through weighted integration; Diagnostic Model Evaluation Module: This module is used to comprehensively evaluate the diagnostic model using the test sets of the current task and the test sets of historical tasks. It calculates the overall diagnostic accuracy of the diagnostic model on all learned fault categories and quantitatively analyzes its knowledge forgetting degree and ability to identify new categories. Sustainable diagnostic model building module: used to update the data and model of the current task to the historical database; when new fault type data arrives, the second training module of the diagnostic model is repeatedly executed to the evaluation module of the diagnostic model until all incremental stages are completed, and finally a sustainable diagnostic model that can identify all fault types is built.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images