Generalized incremental lifelong migration intelligent fault diagnosis method and device for time-varying working condition, medium and product

Abstract

The invention discloses a generalized incremental lifelong migration intelligent fault diagnosis method and device for a time-varying working condition, a medium and a product, and relates to the field of fault diagnosis. The method comprises the following steps: constructing an information enhancement source domain, an information enhancement target domain and a corresponding increment task based on a vibration sensor signal, and generating an initial task, an increment task training set and a test set; performing initial training on the class increment transfer learning model by adopting the initial task training set; continuously training by adopting an incremental task training set, generating a dynamic network initial training parameter through a dynamic network expansion strategy, obtaining an optimal training parameter by relying on a multi-stage training strategy, and updating a double-layer memory buffer area; and evaluating the updated class increment migration learning model by adopting a test set, and outputting a generalized class increment migration diagnosis result. The incremental fault diagnosis performance of the rotating machine under the time-varying working condition can be improved.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis, and in particular to a generalized incremental lifetime migration intelligent fault diagnosis method, device, medium and product for time-varying operating conditions. Background Technology

[0002] With the advancement of artificial intelligence technology, the rise of the Fourth Industrial Revolution is reshaping the modern Prognostics and Health Management (PHM) system. As a core component of PHM technology, intelligent fault diagnosis systems effectively support intelligent maintenance decisions for high-end equipment, ensuring the safety, availability, and operational efficiency of machinery. With the development of the Industrial Internet of Things (IIoT) and sensor technology, deep learning-based fault diagnosis methods can directly map raw sensor data to fault modes, achieving automatic fault diagnosis without relying on complex fault mechanisms or manual feature extraction.

[0003] While deep learning-based fault diagnosis methods have achieved significant results in the field, their application remains limited to closed and static fault mode scenarios, a limitation known as isolated learning. Isolated learning focuses only on fault mode information within the task itself, neglecting knowledge of new fault modes outside the task. Therefore, isolated learning is limited in engineering applications because the fault mode library in real-world industrial scenarios is inherently open and dynamic; monitoring data for various fault modes is continuously generated throughout the entire lifecycle of rotating machinery. Directly updating deep learning models with fixed fault mode data often leads to catastrophic forgetting, where newly acquired fault knowledge overwrites previously learned knowledge. Consequently, directly updated models may quickly lose their ability to recognize previous fault modes, resulting in a significant decline in diagnostic capabilities. Furthermore, collecting complete data on both new and old fault modes for model training is highly resource-intensive, severely limiting the application of current deep learning-based fault diagnosis methods in real-world industrial scenarios. Ideally, deep learning-based fault diagnosis methods should be able to continuously learn new fault modes from the data stream while minimizing redundant computation.

[0004] While preserving historical data and retraining deep learning models can address catastrophic forgetting to some extent, this approach is costly and difficult to implement in real-world industrial deployments due to limitations in storage capacity, computing resources, and data security. Therefore, developing lifelong intelligent diagnostic models with anti-forgetting capabilities and continuous learning abilities has become a core challenge urgently needing to be overcome in the field of fault diagnosis.

[0005] In recent years, incremental learning has offered new possibilities for intelligent diagnostic models. This method allows deep learning models to continuously learn new fault mode knowledge from data streams without forgetting previously learned fault mode classes. Existing research on intelligent fault diagnosis driven by incremental learning can be broadly categorized into parameter-based methods, regularization-based methods, and replay-based methods. Parameter-based methods enhance the model's incremental diagnostic capabilities by introducing new network structures to learn new tasks while retaining the original parameters to maintain the performance of previous fault mode classes. Regularization-based methods mitigate the catastrophic forgetting problem by adding additional constraints to the training objective to penalize the parameter update process. Replay-based methods integrate new and old fault mode classes to adapt to the new fault class and effectively retain learned prior fault knowledge by replaying a small amount of past data during the learning process. Although incremental learning methods for fault diagnosis have shown certain superior performance, many key challenges remain in real-world industrial scenarios. Existing incremental learning transfer diagnostic methods focus on transfer diagnostics under fixed operating conditions. However, real industrial equipment often adjusts its operating conditions to meet different task requirements, resulting in differences in the distribution of monitoring data features caused by time-varying speeds. This change not only deviates from the assumption of steady-state operation but also exacerbates the complexity and difficulty of transfer diagnostic tasks. Furthermore, when faced with time-series streaming data affected by feature distribution shifts, existing incremental fault diagnosis methods encounter a dilemma of stability (retaining acquired knowledge) versus plasticity (learning new knowledge). This often leads to catastrophic forgetting of the model in new diagnostic tasks, resulting in misclassification of fault modes. Moreover, the knowledge replay technique in incremental fault diagnosis relies on sample replay mechanisms, ignoring the rich high-dimensional feature information in the samples.

[0006] Therefore, under dynamic and time-varying operating conditions, developing a kind of incremental migration diagnostic method that combines stability and flexibility, and combining it with enhanced playback technology, is of great significance for realizing lifelong continuous migration diagnostics of mechanical equipment. Summary of the Invention

[0007] The purpose of this application is to address the limitations of existing incremental learning methods in diagnostic performance, stability and plasticity, and the insufficient generalization ability of dynamic networks under incremental time-varying migration scenarios. It provides a generalized incremental lifetime migration intelligent fault diagnosis method, device, medium, and product for time-varying operating conditions, aiming to improve the incremental fault diagnosis performance of rotating machinery under time-varying operating conditions.

[0008] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions, including: Vibration sensor signals of rotating machinery under different working conditions and health states are acquired to construct an information enhancement source domain and an information enhancement target domain based on continuous wavelet transform. An incremental task based on continuous wavelet transform is constructed to enhance the information source domain and the information enhancement target domain, thereby generating an initial task training set, an incremental task training set, and a test set; the initial task training set, the incremental task training set, and the test set include a source domain training set and a target domain training set; An incremental transfer learning model is trained using an initial task training set. Training parameters are randomly initialized to obtain a dynamic feature extraction network consisting of a feature extractor and a classifier. The gradients of the feature extraction backbone parameters in the dynamic feature extraction network are updated based on a multi-stage training strategy and a double-layer memory buffer. A dynamic feature extraction network is trained and updated with gradients of the backbone parameters of the feature extraction network using an incremental task training set. The initial training parameters of the dynamic network are generated through a dynamic network expansion strategy. The optimal training parameters are obtained by relying on a multi-stage training strategy, and the double-layer memory buffer is updated to obtain the updated incremental transfer learning model. The updated class incremental transfer learning model is evaluated using the test set, and the updated class incremental transfer learning model whose evaluation results meet the set requirements is used as the intelligent fault diagnosis model. Real-time acquisition of vibration sensor signals from rotating machinery to be diagnosed under time-varying operating conditions; Using the aforementioned intelligent fault diagnosis model, based on the vibration sensor signals of the rotating machinery to be diagnosed under time-varying operating conditions, a generalized incremental migration diagnosis result is obtained.

[0009] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions provided above.

[0010] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions provided above.

[0011] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions provided above.

[0012] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a generalized incremental lifetime transfer intelligent fault diagnosis method, device, medium, and product for time-varying operating conditions. By employing a dynamic network expansion strategy to generate initial training parameters for the dynamic network, the stability-plasticity dilemma in incremental time-varying transfer scenarios can be effectively addressed, enabling the incremental transfer learning model to capture new information while retaining prior knowledge. Utilizing a multi-stage training strategy enhances the generalization ability of the dynamic feature extraction network, thereby improving knowledge retention and adaptability. Designing a double-layer memory buffer enhances the incremental diagnostic capability of the dynamic feature extraction network in time-varying transfer scenarios, thus improving the incremental fault diagnosis performance of rotating machinery under time-varying operating conditions. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating a generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the training and testing process of an incremental transfer learning model provided in an embodiment of this application; Figure 3 A schematic diagram of the implementation architecture of a generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions provided in an embodiment of this application; Figure 4 A schematic diagram of the evaluation results provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] In one exemplary embodiment, this application provides a generalized class incremental lifelong transfer diagnosis (GCILTD) method for time-varying operating conditions. This method is executed by a computer device, specifically a terminal or server, either alone or jointly. In this embodiment, the method is described using a server as an example. Figure 1 and Figure 2 As shown, the method includes: Step 100: Acquire vibration sensor signals of rotating machinery under different working conditions and health states to construct an information enhancement source domain and an information enhancement target domain based on continuous wavelet transform.

[0018] Step 101: Construct incremental tasks based on continuous wavelet transform for information enhancement in the source and target domains to generate initial training sets, incremental training sets, and test sets. The initial training sets, incremental training sets, and test sets include source domain training sets and target domain training sets.

[0019] For example, a systematically constructed incremental task can include n The source domain training set (labeled data used to train the model) for each task and containing n The target domain training set for each task (labeled data used for fine-tuning the model, with a very small amount of data) is used as the input data for the incremental transfer learning model adopted in this application.

[0020] Step 102: Train the incremental transfer learning model using the initial task training set, randomly initialize the training parameters, obtain the dynamic feature extraction network consisting of a feature extractor and a classifier, and update the gradient of the feature extraction backbone parameters in the dynamic feature extraction network based on the multi-stage training strategy and the double-layer memory buffer.

[0021] In practical applications, this step primarily relies on training an incremental transfer learning model based on the initial task. For example, for the initial task... First, an incremental transfer learning model is trained using data from the initial task training set, and the training parameters are randomly initialized. The dynamic feature extraction network is obtained, and the gradient of the feature extraction backbone parameters is updated by relying on a multi-stage training strategy and a double-layer memory buffer.

[0022] Step 103: Train the dynamic feature extraction network after updating the gradient of the backbone parameters of the feature extraction using the incremental task training set. Generate the initial training parameters of the dynamic network through the dynamic network expansion strategy. Obtain the optimal training parameters by relying on the multi-stage training strategy and update the double-layer memory buffer to obtain the updated incremental transfer learning model.

[0023] In practical applications, this step is mainly based on training incremental transfer learning models using incremental tasks. For example, for the first... Each incremental task generates the initial training parameters for the dynamic network through a dynamic network expansion strategy. Then, the optimal training parameters are obtained by relying on a multi-stage training strategy. And update the double-layer memory buffer.

[0024] Step 104: Evaluate the updated class incremental transfer learning model using the test set. The updated class incremental transfer learning model whose evaluation results meet the set requirements (e.g., the error between the output results and the actual data is less than the set value) is used as the intelligent fault diagnosis model (hereinafter referred to as the model).

[0025] Step 105: Real-time acquisition of vibration sensor signals from rotating machinery to be diagnosed under time-varying operating conditions.

[0026] Step 106: Using an intelligent fault diagnosis model, based on the vibration sensor signals of the rotating machinery to be diagnosed under time-varying working conditions, the generalized incremental migration diagnosis results are obtained.

[0027] In one exemplary embodiment of this application, to avoid the interference of class imbalance on the feature learning process, while ensuring the consistency of local and global feature learning in the class space, in the class incremental transfer diagnostic scenario, there exists... n An incremental task with non-repeating categories. The first... t The training set for each task is denoted as . ,in and These are the source domain training set and the target domain training set, respectively. Define the first... t The source domain training set for each task is It contains a local training set. and global training set The local training set and the global training set are obtained by randomly dividing the source domain training set in equal proportions according to the principle of maintaining a consistent proportion of categories. For example, each of the two sets accounts for 50% of the source domain training set in terms of sample quantity, and the distribution of category labels is consistent.

[0028] In an exemplary embodiment of this application, step 102 essentially initializes the dynamic feature extraction network, and the implementation process may include: In the multi-stage training strategy phase, based on the initial dynamic feature extraction network, fine-grained feature learning, global feature learning, and cross-domain fast adaptation are combined to enhance the generalization ability of the dynamic network, thereby improving knowledge retention and adaptive capabilities. The feature extractor in the dynamic feature extraction network processes the local training set in the source domain. Perform fine-grained feature learning on the global training set of the source domain. Global feature learning is performed by introducing a dynamically adjusted learning rate during the local training phase (i.e., the fine-grained feature learning process) to enhance the feature representation and generalization ability of the incremental transfer learning model in incremental transfer scenarios. Finally, a small number of samples from the target domain training set are used. Cross-domain fast adaptive fine-tuning of the dynamic feature extraction network parameters yields the optimal training parameters for the initial task. .

[0029] As an optional implementation, to further improve diagnostic performance in class-incremental time-varying transition scenarios, a dual-layer memory buffer was developed by integrating a feature-level buffer and a sample-level buffer. Based on this, the dual-layer memory buffer consists of two complementary components: a feature-level buffer that stores high-dimensional features for replay, and a sample-level buffer that stores a small number of samples from the initial task training set. By integrating the feature-level buffer and the sample-level buffer, the fault diagnosis capability in class-incremental time-varying transition scenarios can be further enhanced. Specifically, the feature-level buffer stores high-dimensional features output by the feature extractor (the feature dimensions are set by the user according to actual needs) and participates in feature replay in subsequent tasks. The key samples stored in the sample-level buffer participate in gradient updates during the training phase through sample replay. In particular, the feature-level buffer focuses on preserving high-dimensional features that capture abstract knowledge across tasks. Conversely, the sample-level buffer maintains a small number of samples to preserve the original data distribution and class separability.

[0030] 1) Feature-level buffer.

[0031] Feature-level buffers retain previous... The high-dimensional features extracted from each incremental task are compared with the features extracted in the current incremental task. Connect them, and then identify the characteristics of the connection. F As input to the classifier, we have: .

[0032] In the formula, It is a feature concatenation operation. and These are the feature extractor and the input sample in the incremental task t, respectively.

[0033] 2) Sample-level buffer.

[0034] Traditional cluster-based sample selection strategies for sample replay often sacrifice intra-class diversity. To better preserve the representativeness of each class, K-means clustering can be used to construct a sample buffer (i.e., a sample-level buffer). For class c in the incremental task t, the normalized features... Divided into m clusters And select the distance from each cluster center The most recent m samples are taken as representative samples, and we have: .

[0035] In the formula, The closest to the incremental task t Clustering The index of the selected sample. It is a Euclidean norm. For the index of class c in incremental task t The corresponding normalized features.

[0036] In an exemplary embodiment of this application, step 103 above essentially involves removing the first... The initial parameters inherited in the network trained for each task As initial parameters, and based on these initial parameters, incremental training is performed, where: (1) The dynamic network expansion strategy is achieved by combining parameter inheritance and expansion in the feature extractor and classifier. It aims to solve the stability-plasticity dilemma in the incremental time-varying transition scenario, so that the final intelligent fault diagnosis model can capture new information while retaining prior knowledge. The specific process is as follows: 1) Inheritance of feature extractor parameters.

[0037] In incremental tasks In the initial parameters of the feature extractor From frozen old subnet parameters and the parameters of the expanded new subnet The components include: .

[0038] Initial parameters of the feature extractor It will not be updated in subsequent training. The specific process is explained below: First, incremental tasks t old subnet parameters Directly inherit the task The training subnets in the network include: .

[0039] in, For the task The parameters of the feature extractor. This inheritance mechanism preserves the discriminative ability of previous tasks while preventing the destruction of previously acquired knowledge during the training of new tasks.

[0040] Secondly, incremental tasks t Parameters of the new subnet extended in the middle Through incremental tasks The parameters of the trained subnet (i.e., the task) The parameters of the feature extractor are passed to initialize it: .

[0041] parameter New subnets can be built upon the knowledge embedded in older subnets, rather than starting from scratch.

[0042] 2) Classifier parameter expansion.

[0043] In incremental tasks t In this model, with the introduction of new failure modes, the classifier is extended to accommodate the expanded class space. Detailed design description is as follows: First, a stability-driven extension is performed on the old fault modes. The stability-driven extension strategy allows the old fault modes to retain their original discriminative ability, while utilizing the enhanced representation introduced by the new subnet, thereby improving the model's expressive power.

[0044] Incremental tasks The classifier weights corresponding to the features learned in the task are retained in the task. t In order to ensure the preservation of prior capabilities, we have: In the formula, and These are incremental tasks. t and Classifier weights for old fault modes in the old subnet. For incremental tasks Fault mode quantity in d is the output dimension of the feature extractor.

[0045] For the new subnet, the classifier weights associated with the old failure modes are randomly initialized, and then in the incremental task... t The training process involves refinement. This strategy allows older fault modes to retain their previously acquired discriminative power while leveraging the rich representations introduced by the new subnet, thereby enhancing the model's expressiveness. This can be described as follows: .

[0046] In the formula, It is an incremental task tThe classifier weights for the old fault modes in the Sino-Singapore subnet. The standard deviation of the normal distribution. With a mean of 0 and a variance of It follows a normal distribution.

[0047] Secondly, a plasticity-driven extension is applied to new failure modes. In step 101, the new failure modes introduced by the incremental tasks are pre-determined and incorporated into the training process as new categories. Each incremental task corresponds to a set of new failure categories. These new failure modes are introduced into the incremental transfer learning model at the data level as new category labels, forming the complete failure categories for the current task together with the old categories. The incremental transfer learning model learns the new failure modes through incremental learning while maintaining the recognition of old failure modes. For new failure modes, classification weights are randomly initialized on both the old and new subnets. This allows the model to learn to combine information from multiple feature spaces, thereby accurately identifying new failure modes while maintaining compatibility with the expanded classifier. There are: .

[0048] In the formula, and These represent incremental tasks. t Classifier weights for new fault modes in old and new subnets. These represent incremental tasks. t The number of fault modes.

[0049] Incremental tasks are achieved by expanding classifier parameters. t The classifier in this system can adapt to the expansion of features in new subnets while maintaining its ability to identify old failure modes. This allows for accurate identification of new failure modes while preserving knowledge of old failure modes. Incremental task t Classifier The initial training parameters are expressed as follows: In the formula, and These are incremental tasks. t The classifier weights and bias vectors in the data. and These are incremental tasks. and t Classifier bias vectors for new and old fault modes.

[0050] Finally, incremental tasks t initial training parameters Represented as: 。

[0051] In the formula, For incremental tasks t The initial training parameters of the classifier, These are the initial training parameters for the feature extractor in incremental task t.

[0052] (2) The multi-stage training strategy is a training strategy that includes fine-grained feature learning, global feature learning and cross-domain fast adaptation. It refines the initial training parameters through the feature learning objectives of specific stages. This effectively enhances the generalization of dynamic networks in incremental time-varying migration scenarios. The specific process is described below: 1) Fine-grained feature learning. The essence of fine-grained feature learning is to capture local differences and subtle details within a class. This is achieved by training data locally. By applying gradient updates at both the micro and macro levels, the dynamic network gradually learns features that are highly sensitive to the incremental task t. The micro-level gradient update can be described as follows: .

[0053] In the formula, It is the first The training parameters of the incremental task t in the next iteration. It is the first The training parameters of the incremental task t in the next iteration are used to update the parameters. . It is local training data The classifier loss in the middle, This represents the learning rate in a dynamic network. Represents the classifier loss In parameters The gradient calculated at that point.

[0054] After R rounds of micro-level gradient updates, macro-level gradient updates are adopted to focus on optimizing the overall feature distribution and enhancing the generalization ability of the dynamic network across tasks. .

[0055] In the formula, For the first Incremental tasks in the next iteration The parameters are used to update the training parameters of the incremental task t. . yes No. The second macro gradient and the first The parameters in the sub-gradient update step This represents the dynamic learning rate. Represents the classifier loss In parameters The gradient calculated at that point.

[0056] 2) Global Feature Learning. Global feature learning is based on global training data. Driven by monitoring and adjusting the dynamic learning rate during fine-grained feature learning. To effectively prevent overfitting, the following measures are taken: .

[0057] In the formula, For global training data The classifier loss in the middle, It is the minimum value. and They are respectively At the kth and the The amount of change in each epoch.

[0058] 3) Cross-domain fast adaptation. Cross-domain fast adaptation utilizes data from a limited target domain. A small number of gradient updates allow dynamic networks to quickly adapt to unknown operating conditions. Training parameters Fine-tuning for e iterations yields: .

[0059] In the formula, For incremental tasks t The Middle e The training parameters of the next iteration are the parameters that have been adjusted quickly and adaptively across domains. For incremental tasks t The Middle e Training parameters for -1 iteration. The learning rate of the dynamic feature extraction network during the cross-domain adaptation phase. For loss function In parameters The gradient calculated at that point, For data in the target domain The loss function calculated above.

[0060] Based on the above description, the overall implementation architecture of the method provided in this application is as follows: Figure 3As shown, the process first collects vibration signals under different operating conditions to construct the source and target domains for information enhancement. Random sampling of the vibration signals is then used to construct a sample matrix. Subsequently, a continuous wavelet transform information enhancement strategy is applied to this sample matrix to enhance fault characteristics under different operating conditions, generating a continuous wavelet transform image. Next, an incremental task is constructed, creating a training set that partitions the source and target domains based on continuous wavelet transform information enhancement. Then, a class-incremental transfer learning model is constructed, including a dynamic network expansion strategy, a multi-stage training strategy, and a two-layer memory buffer module. Subsequently, the class-incremental transfer learning model is trained, including initial task training and incremental learning training. The initial task training includes class-incremental parameter learning, a multi-stage training strategy, and a two-layer memory buffer, while the incremental task training adds a dynamic network expansion strategy on top of this.

[0061] In incremental transfer learning diagnostics, the training process for the incremental task includes three stages: dynamic network expansion, multi-stage training, and a two-layer memory buffer. In the dynamic network expansion stage, the incremental task is expanded through inheritance. The training parameters, feature extractor, and classifier were all expanded to generate the initial training parameters for the dynamic network of the incremental task t. In the multi-stage training strategy phase, the initial training parameters for the incremental task t are... The optimal training parameters are obtained by updating through fine-grained learning, global feature learning, and fast cross-domain adaptation. In the two-layer memory buffer stage, firstly, the high-dimensional features of task t are combined with those from the previous layer... The high-dimensional features extracted from each task are concatenated to update the feature-level memory buffer. Subsequently, a small number of samples containing new fault mode information are selected from the dataset and integrated into the sample-level memory buffer for future learning. Finally, the test set is input into the new model to obtain class incremental transfer diagnostic results. This application fully considers the impact of different operating conditions in industrial scenarios on diagnosis, providing a new method for fault diagnosis and offering important technical support for the safe and reliable operation of equipment.

[0062] Furthermore, the diagnostic accuracy of the GCILTD method proposed in this application was evaluated in four different transfer diagnostic scenarios, and the results are shown in Table 1. Clearly, among these seven methods, the GCILTD method achieves the best average diagnostic accuracy of 93.88%, exceeding the iCaRL-F method by 4.20%, the Replay-F method by 6.36%, the DER-F method by 3.11%, and the Foster-F method by 9.28%. GCILTD is the method provided in this application; iCaRL-F represents the incremental classifier and representation learning method; Replay-F represents the memory replay method; DER-F represents the dynamically scalable representation method; and Foster-F represents the feature enhancement and compression method.

[0063] Table 1. Diagnostic accuracy assessment in four different migration diagnostic scenarios.

[0064] Then, the performance of the four incremental fault diagnosis methods was compared in four types of incremental time-varying migration scenarios. SCLIFD represents the supervised contrastive knowledge distillation incremental fault diagnosis method, MCITL represents the multi-component task incremental learning method, and TCIDN represents the dual-channel bidirectional network fault diagnosis method. As shown in Table 2, the average diagnostic accuracy of the proposed GCILTD method is significantly better than other incremental diagnosis methods, exceeding SCLIFD, MCITL, and TCIDN by 4.76%, 2.42%, and 12.46%, respectively. Table 3 shows the forgetting rate of various incremental diagnosis methods in different types of incremental time-varying migration scenarios. The proposed GCILTD method has the lowest average forgetting rate, which is significantly better than the other three methods, further verifying the task stability and robustness of the GCILTD method.

[0065] Table 2. Diagnostic accuracy of four types of incremental fault diagnosis methods in four types of incremental time-varying migration scenarios.

[0066] Table 3 Forgetting rates in different types of incremental time-varying migration scenarios

[0067] To comprehensively compare the performance of the GCILTD method in this application with the three types of incremental diagnostic methods, five representative indicators—diagnostic accuracy, accuracy stability, anti-forgetting, anti-forgetting stability, and efficiency—were used for evaluation. Specifically, efficiency and anti-forgetting were quantified as the reciprocals of running time and forgetting rate, respectively, while accuracy stability and anti-forgetting stability were quantified as the reciprocals of the variances of five experiments conducted under the four types of incremental time-varying migration scenarios.

[0068] Evaluation results as follows Figure 4 As shown, a larger coverage area indicates better overall diagnostic performance. Although the proposed GCILTD method is less efficient than the TCIBN method due to its reliance on backpropagation, which leads to higher computational costs, the proposed GCILTD method still demonstrates superior overall diagnostic performance when considering five representative indicators.

[0069] In an exemplary embodiment of this application, ablation experiments were conducted to demonstrate the effectiveness of the proposed GCILTD method. Four different ablation studies were considered, including removal of the sample-level memory buffer region (Method 1), removal of the feature-level memory buffer region (Method 2), removal of dynamic network expansion (Method 3), and removal of multi-stage training (Method 4). Table 4 presents the ablation experiment results for different categories of incremental time-varying transfer scenarios. By comparing the GCILTD method with removal of the sample-level memory buffer region (Method 1), removal of the feature-level memory buffer region (Method 2), removal of dynamic network expansion (Method 3), and removal of multi-stage training (Method 4), the proposed GCILTD method achieves a best average diagnostic accuracy of 93.88% in various transfer scenarios, showing improvements of 0.43%, 1.1%, 6.51%, and 26.60%, respectively, as shown in Table 4.

[0070] Table 4 Ablation experimental results under different categories of incremental time-varying migration scenarios

[0071] Based on the above description, the key modules in this application are dynamic network expansion and a multi-stage training strategy, which can significantly improve diagnostic accuracy. This is because the former enhances incremental learning capabilities through adaptive architecture and parameter expansion, while the latter achieves continuous fine-grained and global feature learning and rapid cross-domain adaptation in incremental time-varying transfer scenarios, significantly improving the adaptability, anti-forgetting ability, and model generalization ability of incremental learning in dynamic scenarios, thereby effectively enhancing the performance of incremental transfer diagnosis.

[0072] In summary, this application develops a dynamic network expansion strategy that effectively overcomes the stability-plasticity dilemma, enabling the model to capture new information while retaining prior knowledge. Furthermore, by employing a multi-stage training strategy, the generalization ability of the dynamic network is enhanced, further improving knowledge retention and adaptability. The design of a double-layer memory buffer further enhances the incremental class diagnosis capability in time-varying transfer scenarios. Compared with existing incremental class fault diagnosis methods, this application can more effectively adapt to time-varying operating conditions and achieve robust incremental class transfer diagnosis.

[0073] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores generalized incremental lifetime migration intelligent fault diagnosis data oriented towards time-varying operating conditions. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a generalized incremental lifetime migration intelligent fault diagnosis method oriented towards time-varying operating conditions.

[0074] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0075] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0076] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0077] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0078] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0079] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0080] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] This application uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. A generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions, characterized in that, include: Vibration sensor signals of rotating machinery under different working conditions and health states are acquired to construct an information enhancement source domain and an information enhancement target domain based on continuous wavelet transform. An incremental task based on continuous wavelet transform is constructed to enhance the information source domain and the information enhancement target domain, thereby generating an initial task training set, an incremental task training set, and a test set; the initial task training set, the incremental task training set, and the test set include a source domain training set and a target domain training set; An incremental transfer learning model is trained using an initial task training set. Training parameters are randomly initialized to obtain a dynamic feature extraction network consisting of a feature extractor and a classifier. The gradients of the feature extraction backbone parameters in the dynamic feature extraction network are updated based on a multi-stage training strategy and a double-layer memory buffer. A dynamic feature extraction network is trained and updated with gradients of the backbone parameters of the feature extraction network using an incremental task training set. The initial training parameters of the dynamic network are generated through a dynamic network expansion strategy. The optimal training parameters are obtained by relying on a multi-stage training strategy, and the double-layer memory buffer is updated to obtain the updated incremental transfer learning model. The updated class incremental transfer learning model is evaluated using the test set, and the updated class incremental transfer learning model whose evaluation results meet the set requirements is used as the intelligent fault diagnosis model. Real-time acquisition of vibration sensor signals from rotating machinery to be diagnosed under time-varying operating conditions; Using the aforementioned intelligent fault diagnosis model, based on the vibration sensor signals of the rotating machinery to be diagnosed under time-varying operating conditions, a generalized incremental migration diagnosis result is obtained.

2. The generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions according to claim 1, characterized in that, The source domain training set is randomly divided into local and global training sets according to the principle of maintaining a consistent proportion of categories.

3. The generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions according to claim 2, characterized in that, In the process of updating the gradient of the feature extraction backbone parameters in the dynamic feature extraction network based on the multi-stage training strategy and the double-layer memory buffer, the multi-stage training strategy is implemented by combining fine-grained feature learning, global feature learning and cross-domain fast adaptation based on the dynamic feature extraction network. Specifically, the feature extractor in the dynamic feature extraction network performs fine-grained feature learning on the local training set in the initial task training set and global feature learning on the global training set in the initial task training set. In the fine-grained feature learning process, by introducing a dynamically adjusted learning rate, the parameters of the dynamic feature extraction network are rapidly and cross-domain adaptively adjusted using the target domain training set.

4. The generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions according to claim 3, characterized in that, The process of rapidly adaptively adjusting the parameters of the dynamic feature extraction network across domains is represented as follows: ； In the formula, For incremental tasks t The Middle e The training parameters of the next iteration, i.e. the parameters after fast adaptive adjustment across domains; For incremental tasks t The Middle e Training parameters for -1 iteration. The learning rate of the dynamic feature extraction network during the cross-domain adaptation phase. loss function In parameters The gradient calculated at that point, For data in the target domain The loss function calculated above.

5. The generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions according to claim 1, characterized in that, The dual-layer memory buffer consists of a feature-level buffer region and a sample-level buffer region; the feature-level buffer region is used to store high-dimensional features and participate in feature replay; the high-dimensional features refer to features whose dimension is higher than a set value output by the feature extractor. The sample-level buffer is used to store samples in the initial task training set; the samples stored in the sample-level buffer participate in the update of the gradient of the feature extraction backbone parameters through sample replay during the training of the incremental transfer learning model using the initial task training set.

6. The generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions according to claim 1, characterized in that, A dynamic network expansion strategy is achieved by combining parameter inheritance in the feature extractor and parameter expansion in the classifier.

7. The generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions according to claim 6, characterized in that, Parameter inheritance in the feature extractor includes: the parameters of the old subnet in the current incremental task are directly inherited from the trained subnet in the previous incremental task; the parameters of the new subnet expanded in the current incremental task are initialized by passing the parameters of the trained subnet in the previous incremental task. The parameter expansion in the classifier includes: a stability-driven expansion for old fault modes; a plasticity-driven expansion for new fault modes; the old fault modes are pre-determined fault modes when acquiring vibration sensor signals of rotating machinery under different operating conditions and health states; the new fault modes are pre-determined fault modes during the construction of incremental tasks.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the generalized incremental lifetime migration intelligent fault diagnosis method for time-varying operating conditions as described in any one of claims 1-7.

Images