Hybrid hint learning network and system for rolling bearing cross-domain fault diagnosis

Abstract

The application provides a hybrid prompt learning network and system for rolling bearing cross-domain fault diagnosis, comprising a self-step prototype prompt generator, a context-aware prompt retriever and a prompt-guided refinement and adaptation module. In this way, a ProtoRAP framework is proposed to realize rolling bearing cross-domain fault diagnosis through prototype-guided knowledge encoding, context-aware dynamic retrieval and gated-driven feature adaptation. The hybrid prompt learning network and system for rolling bearing cross-domain fault diagnosis provided in the embodiment are significantly superior to existing advanced methods in terms of diagnosis accuracy and cross-domain stability in the variable working condition diagnosis task of multiple rolling bearing fault data sets, verifying the effectiveness and robustness of the hybrid prompt learning network and system for rolling bearing cross-domain fault diagnosis in complex industrial scenarios.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and in particular to a hybrid cue learning network and system for cross-domain fault diagnosis of rolling bearings. Background Technology

[0002] Rolling bearings are core components of rotating machinery, widely used in critical industrial sectors such as aerospace, wind power generation, and rail transportation. Their health directly impacts equipment safety and production efficiency. As one of the main sources of rotating machinery failures, undiagnosed bearing defects can trigger cascading damage or even catastrophic accidents. With the deepening advancement of intelligent manufacturing, deep learning-based fault diagnosis technology has attracted significant attention due to its powerful feature learning capabilities, achieving remarkable results under controlled experimental conditions. However, the success of deep learning models heavily relies on the fundamental assumption that training and testing data follow an independent and identically distributed (IID) pattern. In actual industrial scenarios, the statistical characteristics of rolling bearing vibration signals are coupled with multiple factors, including operating parameters (speed, load, ambient temperature) and the arrangement of measuring points, leading to significant data distribution shifts (domain shift) between different operating conditions. When a model trained in the source domain is directly deployed to the target domain, domain shift often results in a sharp decline in diagnostic performance. This cross-domain transfer challenge has become a core bottleneck restricting the industrial application of intelligent diagnostic technology. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a hybrid cue learning network and system for cross-domain fault diagnosis of rolling bearings, so as to systematically solve the cross-domain fault diagnosis problem from three levels: semantically anchored fault knowledge representation, operating condition adaptive dynamic knowledge extraction, and multi-granular feature refinement and fusion.

[0004] In a first aspect, embodiments of the present invention provide a hybrid cue learning network for cross-domain fault diagnosis of rolling bearings, comprising: a self-paced prototype cue generator, a context-aware cue retrieval unit, and a cue-guided refinement and adaptation module; the self-paced prototype cue generator is used to progressively filter high-confidence samples by explicitly anchoring abstract cue vectors to semantic prototypes of typical bearing fault modes, combined with an uncertainty-driven course learning strategy; the context-aware cue retrieval unit is used to capture cross-domain stable vibration characteristics under different speeds and loads using coarse-grained context representations in the intermediate layers of the network; the cue-guided refinement and adaptation module is used to achieve adaptive weighted aggregation of retrieval cue through a multi-head cross-attention mechanism, and to selectively enhance fault-sensitive dimensions by using a gating network to inversely modulate the target domain operating condition characteristics.

[0005] In optional embodiments of this application, the self-paced prototype prompt generator is used to establish an explicit semantic mapping between prompts and failure modes in the feature space through a prototype anchoring strategy, and to construct a prompt knowledge base; the self-paced prototype prompt generator is also used to select high-confidence samples by combining an uncertainty-driven adaptive course learning strategy.

[0006] In an optional embodiment of this application, the above-mentioned context-aware cue retrieval device is used to extract coarse-grained contextual representations of target domain samples and dynamically retrieve the K most relevant cue from the cue knowledge base by means of cosine similarity, where K is an integer greater than or equal to 1.

[0007] In an optional embodiment of this application, the above-mentioned prompt guidance refinement and adaptation module is used to generate adaptive prompts by retrieving prompts through multi-head cross-attention fusion, refine target domain features using a gating mechanism, selectively enhance fault-related dimensions and suppress domain-specific noise, and finally obtain domain-invariant representations.

[0008] In optional embodiments of this application, the self-paced prototype cue generator includes: a first collaborative submodule, a second collaborative submodule, and a third collaborative submodule; the first collaborative submodule is used to dynamically filter high-confidence samples through Shannon entropy measurement based on an uncertainty-aware course learning mechanism; the second collaborative submodule is used to establish an explicit semantic mapping between cue vectors and physical fault modes based on a prototype anchoring strategy; and the third collaborative submodule is used to maximize the feature differences between cue vectors based on orthogonality regularization.

[0009] In an optional embodiment of this application, the aforementioned context-aware cue retrieval device is used to construct a coarse-grained context representation using intermediate layer features of the target domain samples; wherein, the construction of the coarse-grained context representation preserves the basic signal patterns that are stable across domains; the context-aware cue retrieval device is also used to complete the cue retrieval process by calculating the similarity between the coarse-grained context representation and each cue in the cue knowledge base.

[0010] In an optional embodiment of this application, the above-mentioned prompt guidance refinement and adaptation module constructs a two-stage cascaded reasoning architecture; the prompt guidance refinement and adaptation module is used to transform passively retrieved prompts into active knowledge guidance signals through attention prompt fusion and prompt gating feature refinement.

[0011] In optional embodiments of this application, the hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings described above is used to construct a prompting knowledge base through multi-objective joint optimization during the source domain training phase; the hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings is also used to achieve dynamic retrieval and adaptive fusion of knowledge through a context-aware prompting retrieval device and a prompting guidance refinement and adaptation module during the target domain reasoning phase.

[0012] Secondly, embodiments of the present invention also provide a hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings, the hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings comprising: the aforementioned hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings.

[0013] The embodiments of the present invention bring the following beneficial effects: This invention provides a hybrid prompting learning network and system for cross-domain fault diagnosis of rolling bearings. It proposes a ProtoRAP (Prototypical Retrieval-Adaptation Prompt Learning Framework) framework, which achieves cross-domain fault diagnosis of rolling bearings through prototype-guided knowledge encoding, context-aware dynamic retrieval, and gated feature adaptation. In variable-condition fault diagnosis tasks on multiple rolling bearing fault datasets, the hybrid prompting learning network and system provided in this embodiment significantly outperform existing state-of-the-art methods in terms of diagnostic accuracy and cross-domain stability, validating its effectiveness and robustness in complex industrial scenarios.

[0014] Other features and advantages of this disclosure will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0015] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A schematic diagram comparing a traditional cross-domain transfer learning method with the method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the structure of a hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings provided in an embodiment of the present invention; Figure 3 A schematic diagram of a ProtoRAP framework provided in an embodiment of the present invention; Figure 4A flowchart of a hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings is provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of a hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings, provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0019] Currently, domain adaptation techniques have become a research hotspot in fault diagnosis to address the cross-domain migration challenges caused by the aforementioned domain offset. Existing methods are mainly divided into two categories: those based on discrepancy metrics and those based on adversarial learning. Discrepancy-metric-based methods achieve feature alignment by minimizing the distribution distance between the source and target domains. However, these methods have limited ability to model complex nonlinear distribution offsets, and the global distribution alignment strategy makes it difficult to explicitly encode discriminative fault knowledge for each source domain, lacking adaptive knowledge retrieval capabilities tailored to the characteristics of target domain samples. Adversarial learning-based methods borrow from generative adversarial networks, learning domain-invariant features through adversarial training between the domain discriminator and feature extractor. However, adversarial methods are susceptible to training instability due to distributional differences between source domains in multi-source domain scenarios. Therefore, existing methods all employ implicit feature alignment strategies, making it difficult to explicitly encode the fault semantic knowledge of the source domains and achieve on-demand transfer.

[0020] In recent years, Large Language Models (LLMs) have achieved revolutionary breakthroughs in the field of Natural Language Processing, with their powerful knowledge representation and reasoning capabilities attracting widespread attention from academia and industry. Prompt Learning, as a new paradigm in the LLM era, efficiently adapts pre-trained models to downstream tasks by injecting a small number of learnable cue vectors into the input or feature space, achieving a good balance between parameter efficiency and transfer performance. This technology has been widely applied in fields such as computer vision, providing a new technical path for the explicit encoding and efficient transfer of domain knowledge. In the fields of intelligent manufacturing and fault diagnosis, LLM-related technologies are in a phase of rapid development.

[0021] However, most existing research uses LLM as an auxiliary decision-making tool or directly transfers visual cueing techniques, and has not yet formed a systematic cueing learning framework for intelligent manufacturing scenarios. On the one hand, industrial vibration signals have unique time-frequency domain fault characteristics, and existing cueing learning methods are unable to effectively encode this type of domain-specific knowledge; on the other hand, the multi-source domain distribution shift and negative transfer problems that are common in actual production environments have not yet been addressed specifically. Applying deep cueing learning to multi-source domain fault transfer diagnosis requires systematically addressing the following three challenges: (1) Static decoupling of features and prompts. Existing methods use a serial architecture, in which features are first extracted by a fixed encoder and then prompts are used for retrieval. This one-way interaction ignores the guiding role of prompt knowledge in feature learning, resulting in the inability to dynamically optimize feature representations based on domain knowledge.

[0022] (2) The semantic ambiguity of the prompts and the instability of the learning process. Traditional prompt vectors lack physical semantic constraints, making it difficult to establish an interpretable mapping with the fault mechanism. At the same time, the learning process is easily affected by difficult and noisy samples, resulting in unstable quality and high redundancy of the prompt library.

[0023] (3) Granularity mismatch and negative transfer risk. Relying solely on high-level abstract features for knowledge retrieval while ignoring intermediate and low-level signal patterns leads to granularity mismatch. Naive suggestion fusion strategies cannot adaptively balance the contributions of multi-source knowledge and are prone to introducing noisy knowledge that triggers negative transfer.

[0024] Based on this, the present invention provides a hybrid prompting learning network and system for cross-domain fault diagnosis of rolling bearings, which can be found in [reference needed]. Figure 1 This diagram illustrates a comparison between a traditional cross-domain transfer learning method and the method provided in this embodiment. Figure 1 As shown in (a), traditional methods employ domain-independent models for full model transfer and adaptation; Figure 1 As shown in (b), the method provided in this embodiment achieves lightweight knowledge encoding, selective transfer, and adaptive fusion through domain-specific prompts.

[0025] like Figure 1 As shown, traditional methods follow a "model-driven" paradigm: all source domain data are uniformly input into a domain-independent model for training, all model parameters are directly transferred, and full model adaptation is performed in the target domain. This paradigm has inherent limitations: (1) the domain-independent model is difficult to capture the discriminative fault features of each source domain; (2) the full model transfer lacks selectivity and may transfer source domain noise along with it; (3) the computational cost of full parameter fine-tuning is high, and it is prone to overfitting when the sample size is small. In contrast, the method provided in this embodiment adopts a lightweight "cue-driven" paradigm: a domain-specific cue vector is independently learned for each source domain, and the adaptive aggregation of multi-source cues is achieved through attention weighting. Only the fusion of cues is needed to guide the refinement of target domain features without modifying the backbone network parameters.

[0026] To facilitate understanding of this embodiment, a hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings, as disclosed in this embodiment of the invention, will first be described in detail.

[0027] Example 1: This invention provides a hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings. (See also...) Figure 2 The diagram illustrates the structure of a hybrid cue learning network for cross-domain fault diagnosis of rolling bearings. This network includes: a self-paced prototype cue generator, a context-aware cue retrieval unit, and a cue-guided refinement and adaptation module. The self-paced prototype cue generator explicitly anchors abstract cue vectors to semantic prototypes of typical bearing fault modes, progressively filtering high-confidence samples using an uncertainty-driven learning strategy. The context-aware cue retrieval unit captures cross-domain stable vibration characteristics under different speeds and loads using coarse-grained contextual representations in the network's intermediate layers. The cue-guided refinement and adaptation module adaptively weights and aggregates retrieved cue suggestions through a multi-head cross-attention mechanism, and uses a gating network to modulate the target domain operating condition features, selectively enhancing the fault-sensitive dimension.

[0028] This embodiment provides a ProtoRAP (Prototype Retrieval-Adaptive Hint Learning) framework, which systematically addresses cross-domain fault diagnosis from three levels: semantically anchored fault knowledge representation, condition-adaptive dynamic knowledge extraction, and multi-granularity feature refinement and fusion. The main contributions of this embodiment are summarized below: (1) A new cross-domain diagnostic paradigm of feature-cue bidirectional collaboration is proposed, which breaks through the architectural limitations of feature extraction and domain adaptation being independent in traditional methods. It realizes bidirectional interaction between feature space and domain knowledge through learnable cue vectors, shifting from a "model-driven" to a "cue-driven" learning paradigm, and providing new exploration for cue learning technology in the era of large language models in the field of intelligent manufacturing. (2) Design a prompt generation mechanism (SPPG) that combines prototype anchoring and self-paced learning. The semantic interpretability is guaranteed by fault prototype constraints, and the interference of noise samples is effectively suppressed by the self-paced strategy, thereby enhancing the quality and robustness of the prompt library. (3) Construct an adaptive knowledge transfer strategy that integrates hierarchical retrieval and attention fusion. A dynamic prompt selection is achieved through the context-aware retrieval system (CAPR), and adaptive weighted aggregation is performed with the help of the refined attention module (PGRA), thereby effectively avoiding the risk of negative transfer in multi-source domain scenarios. (4) Experiments on single-source and multi-source domains on three bearing fault datasets, CWRU, JNU and MFS, show that the proposed method is significantly better than existing advanced methods in terms of diagnostic accuracy, providing an effective technical solution for equipment health monitoring and predictive maintenance in intelligent manufacturing environments.

[0029] This invention provides a hybrid cueing learning network and system for cross-domain fault diagnosis of rolling bearings. A ProtoRAP framework is proposed, which achieves cross-domain fault diagnosis of rolling bearings through prototype-guided knowledge encoding, context-aware dynamic retrieval, and gated feature adaptation. In variable-condition fault diagnosis tasks on multiple rolling bearing fault datasets, the hybrid cueing learning network and system provided in this embodiment significantly outperform existing state-of-the-art methods in terms of diagnostic accuracy and cross-domain stability, validating its effectiveness and robustness in complex industrial scenarios.

[0030] Example 2: This invention provides another hybrid cue learning network for cross-domain fault diagnosis of rolling bearings, implemented based on the above embodiments. The specific implementation of the hybrid cue learning network for cross-domain fault diagnosis of rolling bearings is described in detail. This embodiment provides a ProtoRAP framework to address the cross-domain fault diagnosis problem of rolling bearings under varying operating conditions. This method aims to address the core challenge of transferring diagnostic knowledge from a labeled source domain to an unlabeled target domain, while maintaining robustness to distribution shifts caused by changes in operating conditions and environmental noise.

[0031] I. Problem Definition and Framework Overview: In the multi-source domain adaptive fault diagnosis task, let... express A collection of labeled source domain datasets, Let s denote the source domain. For the k-th source domain, its sample set is defined as... ,in This represents the i-th vibration signal sample in the k-th source domain. Represents the set of real numbers. Indicates the length of the vibration signal sample. This represents the fault category label corresponding to the sample, where C is the total number of fault categories. Let be the total number of samples in the k-th source domain. The target domain dataset is defined as follows: , where t represents target (target domain). For the j-th sample in the target domain, This represents the total number of samples in the target domain. The target domain data was collected under different operating conditions and does not include label information.

[0032] The goal of the multi-source domain fault diagnosis problem is to learn a mapping function. This enables it to accurately predict the fault category of samples in the target domain. Indicates the input space, Represents the tag space. Let X be the classification mapping function from the input space X to the label space Y. The core challenge of this problem lies in the difference in marginal distributions between the source and target domains. and potential conditional distribution shift ,in, and Let the marginal probability distributions of the k-th source and target domains be represented respectively. and Let represent the conditional probability distributions of the k-th source domain and the target domain, respectively. This study establishes the following basic assumptions: (1) Label space consistency: all source domains and target domains share the same fault category definition, i.e. , among which, The tag space representing the source domain. The label space representing the target domain. (2) Label availability: source domain data are labeled samples, and target domain data are unlabeled samples.

[0033] In some embodiments, the self-paced prototype cue generator is used to establish an explicit semantic mapping between cue and failure mode in the feature space through a prototype anchoring strategy, and to build a cue knowledge base; the self-paced prototype cue generator is also used to select high-confidence samples by incorporating an uncertainty-driven adaptive curriculum learning strategy.

[0034] In some embodiments, the context-aware cue retrieval unit is used to extract coarse-grained contextual representations of target domain samples and dynamically retrieve the K most relevant cue points from the cue knowledge base by cosine similarity, where K is an integer greater than or equal to 1.

[0035] In some embodiments, the prompt-guided refinement and adaptation module is used to generate adaptive prompts by retrieving prompts through multi-head cross-attention fusion, refine target domain features using a gating mechanism, selectively enhance fault-related dimensions and suppress domain-specific noise, and finally obtain a domain-invariant representation.

[0036] Inspired by the cue learning paradigm in Large Language Models (LLM), this embodiment proposes the ProtoRAP framework, transferring the idea of ​​efficiently adapting learnable cue vectors to pre-trained models to the field of cross-domain fault diagnosis. (See [link to relevant documentation]). Figure 3 The diagram shows a ProtoRAP framework, which consists of three collaborative modules to achieve domain-invariant fault representation learning. (1) The Adaptive Prototype Hint Generator (SPPG) establishes an explicit semantic mapping between hints and failure modes in the feature space through a prototype anchoring strategy and constructs a hint knowledge base (PKR). At the same time, it combines an uncertainty-driven adaptive curriculum learning strategy to gradually select high-confidence samples to ensure the purity of domain knowledge extraction; (2) The context-aware cue retrieval (CAPR) extracts coarse-grained contextual representations of target domain samples and dynamically retrieves the Top-K most relevant cue from the PKR through cosine similarity, realizing the knowledge transfer strategy of "on-demand retrieval". (3) The prompting guidance refinement and adaptation module (PGRA) generates adaptive prompts by retrieving prompts through multi-head cross-attention fusion. Then, it uses a gating mechanism to refine the target domain features, selectively enhance fault-related dimensions and suppress domain-specific noise, and finally obtains a domain-invariant representation.

[0037] II. Self-paced Prototype Hint Generator: In the paradigm of intelligent manufacturing empowered by Large Language Models (LLM), cue learning has become a key technology for efficiently transferring pre-trained models to domain-specific industrial tasks. However, unlike Natural Language Processing (NLP), which mainly deals with discrete text tags, bearing fault diagnosis in industrial scenarios faces two fundamental challenges: (1) how to transform continuous, noisy vibration signals into physically interpretable fault knowledge representations; and (2) how to ensure the purity and robustness of knowledge extraction under the interference of sample noise and annotation uncertainty. To address these challenges, this embodiment proposes a Self-Paced Prototypical Prompt Generator (SPPG) module.

[0038] In some embodiments, the self-paced prototype cue generator includes: a first collaborative submodule, a second collaborative submodule, and a third collaborative submodule; the first collaborative submodule is used to dynamically filter high-confidence samples through Shannon entropy measurement based on an uncertainty-aware curriculum learning mechanism; the second collaborative submodule is used to establish an explicit semantic mapping between cue vectors and physical failure modes based on a prototype anchoring strategy; and the third collaborative submodule is used to maximize the feature differences between cue vectors based on orthogonality regularization.

[0039] The core idea of ​​SPPG is to construct a Prompt-Knowledge Repository (PKR) that simultaneously possesses physical semantic consistency and feature orthogonality. Unlike traditional methods that merely treat prompts as implicit parameters for fine-tuning, SPPG explicitly encodes source domain failure patterns into semantic descriptors with enhanced generalization capabilities by designing uncertainty metrics and prototype constraints. To achieve this, SPPG consists of three collaborative sub-modules: First, an uncertainty-aware curriculum learning mechanism dynamically filters high-confidence samples using Shannon entropy to ensure the purity of knowledge extraction; second, a prototype anchoring strategy establishes an explicit semantic mapping between prompt vectors and physical failure patterns, endowing prompts with interpretability; and finally, orthogonality regularization maximizes the feature differences between prompt vectors, preventing pattern collapse and improving the coverage of the knowledge repository. Through the synergistic effect of these three modules, a high-quality prompt knowledge repository with clear semantics and decoupled features is constructed.

[0040] Data acquisition in real-world manufacturing environments is often accompanied by background noise and fluctuations in operating conditions, resulting in long-tailed or fuzzy data distributions that severely weaken the reliability and traceability of fault diagnosis models. Directly using all data for feature aggregation during the fitting of noise patterns often leads to a misalignment between the prompts and the noise patterns. To address this issue, drawing inspiration from data quality screening in LLM training, this implementation introduces a self-paced learning strategy. This strategy aims to simulate the gradual learning mechanism in human cognition, prioritizing the selection of high-confidence samples to ensure the purity of knowledge extraction.

[0041] This embodiment uses Shannon entropy as a model to analyze the first... Sample The cognitive uncertainty measure of the prediction results. Let the probability distribution vector of the model output be... ,in For parameters θ The feature extractor, and the Softmax function is an activation function used for classification problems. The Each element indicates that the sample belongs to the fault category. The predicted probability, c =1,2,…, C Where C is the total number of fault categories, then the uncertainty is... Defined as: (1) Based on this, a dynamic sample weighting function (binary filter mask) is constructed. This is used to determine the validity of a sample during the t-th training round. In this embodiment, the selection logic is explicitly defined as: (2) in, To control the pace of course learning, a monotonically increasing learning threshold parameter is used. This mechanism follows the learning principle of "from easy to difficult": in the initial stage of training... The initial settings are low, selecting only high-confidence "clean samples" for prototype computation and knowledge extraction to avoid noise interference; as training progresses, Gradually increase the difficulty level, gradually incorporating samples of moderate difficulty to broaden the scope of knowledge coverage. When a sample is considered a reliable sample, it participates in subsequent optimization; when When this happens, the sample is identified as noise or an outlier and is temporarily masked.

[0042] To overcome the lack of physical interpretability in continuous cue vectors and enhance the model's reasoning ability under complex conditions, this embodiment develops a Prototypical Anchoring Prompt Mechanism in the SPPG module. This mechanism aims to establish a semantic mapping relationship between abstract cue vectors and specific bearing failure modes in the feature manifold space, thereby achieving explicit injection of physical semantics.

[0043] First of all, the The characteristic prototype of each source domain Online estimation is performed using a weighted sample set filtered through a self-stepping learning strategy. To ensure the temporal stability of the prototype and mitigate the impact of batch fluctuations, the calculation formula is designed as follows: (3) in, Indicates from the A sample set from each source domain Indicates sample The feature embedding vector, This is a smoothing term used to prevent the denominator from being zero.

[0044] Subsequently, semantic alignment loss is constructed. Used to constrain each cue vector Converging towards the most semantically relevant fault prototype: (4) The gradient of the semantic alignment loss with respect to the cue vector is: (5) Among them, the stopping gradient operator is introduced. This is used to block the backpropagation path, ensuring that the optimization process only adjusts the cue vector to adapt to the actual physical signal distribution, rather than influencing the prototype in the reverse direction. The process transforms cue into semantic anchors to characterize domain-specific mechanisms (such as outer ring peeling or cage damage). Similar to the process of mapping discrete words to continuous semantic embeddings in large language models, this mechanism endows the system with deeper semantic representation capabilities.

[0045] To construct a high-capacity knowledge base capable of supporting subsequent cross-domain adaptation, it is necessary to ensure that the domain-specific cue vectors cover the broadest possible feature subspace in the source domain and prevent pattern collapse. Therefore, this embodiment introduces orthogonality constraints in SPPG to maximize the distinguishability between cue basis vectors from different domains.

[0046] First, the cue vector is row-normalized, defined as: (6) in The normalized cue vectors are then stacked row-wise to form a row-normalized cue matrix. (in d For each prompt vector (feature dimension) This is the L2 norm (also known as the Euclidean norm, representing the length of a vector). Furthermore, an orthogonal regularization loss is introduced. This is used to penalize the cosine similarity between cue vectors from different domains: (7) in, Let Ns be the identity matrix of size Ns×Ns. This represents the square of the Frobenius norm.

[0047] The orthogonal loss is related to the cue vector. The gradient is: (8) in, This represents the j-th cue vector. express The transpose of 1 transforms a column vector into a row vector. This represents the squared L2 norm, which is the square of the vector length.

[0048] This gradient causes the cue vectors from different domains to separate from each other on the unit hypersphere, thus forming orthogonal basis vectors. Minimize This approach promotes mutual orthogonality between cue vectors in Hilbert space. This feature decoupling strategy ensures that each cue in the PKR captures unique domain information, thereby improving the model's generalization ability and knowledge transfer efficiency when facing target conditions.

[0049] III. Context-Aware Prompt Retrieval System: After completing the domain semantic construction and orthogonal decoupling of the source domain prompt vectors, relying solely on static prompts for knowledge transfer in cross-domain scenarios still has significant limitations. Intelligent manufacturing scenarios are often accompanied by fluctuations in operating conditions, noise disturbances, and structural changes, causing the feature distribution of target domain samples to deviate from the applicable scope of source domain prompts. Inspired by the LLM domain retrieval-Augmented Generation (RAG) paradigm, this embodiment proposes a Context-Aware Prompt Retriever (CAPR). Similar to RAG enhancing the model's reasoning ability by retrieving external knowledge bases, CAPR dynamically selects the Top-K most relevant source domain prompts from the prompt knowledge base, achieving "on-demand invocation" and "context alignment" at the prompt level.

[0050] In some embodiments, the context-aware cue retrieval system is used to construct a coarse-grained context representation using intermediate-layer features of target domain samples; wherein, the construction of the coarse-grained context representation preserves cross-domain stable basic signal patterns; the context-aware cue retrieval system is also used to complete the cue retrieval process by calculating the similarity between the coarse-grained context representation and each cue in the cue knowledge base.

[0051] The core idea of ​​CAPR is to construct a "coarse-grained contextual representation" using intermediate-layer features of the target domain samples. This representation preserves cross-domain stable basic signal patterns (such as impulse periodicity and frequency domain structure), enabling a more robust characterization of the underlying physical properties of the target domain data. Subsequently, by calculating the similarity between this representation and all hints in the Prompt-Knowledge Repository (PKR), a Top-K retrieval process for the hints is completed, thereby ensuring that the final set of hints participating in the inference is semantically relevant to the current target domain samples.

[0052] For a single target domain input sample If the high semantic features of the network output layer are directly used for cue relevance calculation, it is easily affected by cross-domain decision boundary drift, thus reducing the accuracy of retrieval. In contrast, intermediate layer features usually have the characteristics of "low semantics and high structure", which are more likely to capture the basic vibration mode of bearing fault signals, thus providing a more stable cross-domain consistency basis for cue retrieval. To enhance cross-domain robustness, this embodiment uses the output of the intermediate layer of the feature extraction network as a coarse-grained context representation of the sample: (9) in, For parameters φ The intermediate layer mapping function of the feature extractor (d (The feature dimension is consistent with the cue vector dimension.) These are intermediate layer features. This mapping makes... It can preserve consistent structural information across different domains.

[0053] To further eliminate local noise interference and improve representation compactness, global pooling is used to obtain the final context vector: (10) Contextual representation of a given target domain sample CAPR is calculated Hint retrieval is performed based on the similarity between the hint vectors and all hint vectors in the PromptKnowledge Repository (PKR). The PKR is optimized during training in the source domain. Domain-specific cues are constructed. Because normalized cosine similarity is insensitive to amplitude changes and has strong robustness to noise disturbances, it is used as a measure of cue relevance. (11) in, For the first The normalized cosine similarity between the source domain cue vector and the current target domain sample.

[0054] Sort all similarities in descending order of score, and select the top-K highest-scoring prompts to form a candidate prompt set for the target domain. : (13) in, For target domain samples The candidate suggestion set; Top-K is the Top-K selection operation, which selects all... Sort by score from highest to lowest, then select the top... The highest-scoring item; It is calculated by formula (11) A set of similarity scores. These hints will serve as the knowledge subspace most relevant to the target domain samples, and will be injected into the model inference process through subsequent modules to achieve hint-level dynamic adaptation.

[0055] IV. Refined and Adapted Prompt-Based Guidance Module: While CAPR effectively filters out Top-K source domain hints that are potentially relevant to the target domain samples through coarse-grained matching, directly applying these discrete, unprocessed prior knowledge to target domain diagnosis still has limitations. In the dynamic production environment of smart manufacturing, target domain data often faces complex nonlinear distribution shifts and background noise interference, and simple "feature-hint" splicing strategies cannot fully explore the domain semantics contained in the hints.

[0056] In some embodiments, the prompt-guided refinement and adaptation module constructs a two-stage cascaded reasoning architecture; the prompt-guided refinement and adaptation module is used to transform passively retrieved prompts into active knowledge guidance signals through attention prompt fusion and prompt gating feature refinement.

[0057] Drawing inspiration from the LLM concept of dynamically fusing contextual information through attention mechanisms, this embodiment proposes a Prompt-Guided Refinement and Adaptation Module (PGRA). PGRA constructs a two-stage cascaded reasoning architecture that transforms passively retrieved prompts into proactive knowledge guidance signals through attention-based prompt fusion and prompt-gated feature refinement, thereby achieving co-evolution between the feature space and the prompt space.

[0058] To address the potential semantic misalignment between the retrieved Top-K suggestions and the current target domain samples, a mechanism is needed to dynamically measure the contribution of each candidate suggestion and generate an adaptive suggestion "tailor-made" for the current situation. Inspired by the Transformer, a core component of Large Language Models (LLM), this embodiment designs an Attentional Prompt Fusion layer.

[0059] set up This represents the final layer features of the target domain samples obtained through the feature extractor; candidate cue matrix. This indicates the Top-level results retrieved by CAPR from the Tipping Knowledge Base (PKR). Domain-specific tips. In this article, we will... Mapped to query vector (Query, ), and candidate hint matrix These are then mapped to keys (Key, ) and Value Thus, a multi-head cross-attention mechanism is constructed: (14) in, This is a learnable linear projection matrix. By calculating the dot product scaling attention between the Query and Key, the model can "infer" which source domain fault modes best fit the current target domain samples: (15) in, This is the attention weight vector (reflecting the model's dependence on different cues). This is the attention head dimension. It intuitively reflects the model's dependence on different prior knowledge, enhancing the traceability of the diagnostic process. The final adaptive prompt... It is obtained by weighted aggregation, and residual connections and layer normalization (LayerNorm) are introduced to maintain the stability of gradient flow: (16) By integrating semantic information from multiple sources, a smooth mapping from a "discrete knowledge base" to a "continuous semantic vector" is achieved through an attention mechanism, providing high-quality guiding signals for subsequent feature refinement.

[0060] After obtaining adaptive cues, the key challenge lies in how to leverage this domain knowledge to optimize noisy features in the target domain. Traditional feature adaptation methods often overlook the inverse effect of knowledge on data. To address this, this embodiment proposes a prompt-gated feature refinement mechanism. This mechanism simulates the cognitive process by which human experts use prior experience (Prompt) to focus on key evidence (Feature), dynamically modulating feature channels by generating gating signals.

[0061] This embodiment will generate adaptive prompts. The input is fed into a lightweight gating network, which consists of linear layers and a sigmoid activation function, designed to generate a channel attention mask. : (17) in, This represents the Sigmoid function, which constrains the output to... Within the interval, This is the gating vector. and This is the gating parameter. Subsequently, the gating signal is used to adjust the original features. Perform element-wise feature modulation, and the output is : (18) in, This represents the Hadamard Product. This process achieves knowledge-driven feature reconstruction: gating mechanism. It can adaptively enhance feature dimensions that are highly correlated with the currently inferred failure mode, while suppressing irrelevant dimensions caused by environmental noise or cross-domain distribution differences.

[0062] V. Optimization Target and Parameter Update Strategy: In some embodiments, the hybrid cue learning network for cross-domain fault diagnosis of rolling bearings is used to construct a cue knowledge base through multi-objective joint optimization during the source domain training phase; the hybrid cue learning network for cross-domain fault diagnosis of rolling bearings is also used to achieve dynamic retrieval and adaptive fusion of knowledge through a context-aware cue retrieval device and a cue-guided refinement and adaptation module during the target domain reasoning phase.

[0063] This embodiment illustrates the optimization objective and parameter update strategy of the ProtoRAP framework. Within the LLM-era paradigm of cue learning, the proposed method constructs a high-quality cue knowledge base through multi-objective joint optimization during the source domain training phase, achieving domain-centric fault knowledge encoding. During the target domain inference phase, dynamic knowledge retrieval and adaptive fusion are achieved through the CAPR and PGRA modules, requiring no additional supervision signals. Based on the above analysis, the overall optimization objective function of the ProtoRAP framework during the source domain training phase is defined as: (19) in, Cross-entropy classification loss is used to learn discriminative fault features; The semantic alignment loss ensures semantic consistency between the cue vector and the fault prototype; The orthogonality regularization loss ensures the diversity and decoupling of the cue vectors. Non-negative hyperparameters ( (representing the set of positive real numbers), used to balance the trade-off between semantic consistency and feature diversity.

[0064] Let the set of learnable parameters of the ProtoRAP framework be... ,in These are the parameters for the feature extractor. For classifier parameters, To provide hints for the knowledge base (by N) s It is composed of stacked prompt vectors. d (for feature dimensions) These are the parameters of the PGRA module. The gradients of each parameter are calculated through backpropagation: (20) (twenty one) (twenty two) (twenty three) Parameter updates are performed using the Adam optimizer: (twenty four) in, This is the learning rate.

[0065] Self-paced learning threshold (Consistent with the previous text) Scheduling by linear strategy: (25) in, For the total number of training rounds, and The threshold boundary is defined by this progressive threshold scheduling strategy, which simulates the human cognitive learning process from easy to difficult. This ensures that pure domain knowledge is extracted first from high-confidence samples in dynamic production environments, effectively improving the robustness and generalization ability of the model in resource-constrained scenarios.

[0066] Example 3: This invention provides an experimental verification and analysis of a hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings, which is implemented based on the above embodiments.

[0067] To verify the effectiveness and generalization ability of the ProtoRAP-based rolling bearing fault diagnosis model, this embodiment uses three bearing fault diagnosis cases for systematic empirical evaluation. Specifically, using variable speed operating condition data contained in the CWRU (Case Western Reserve University Bearing Fault Dataset), the evolution law of bearing fault characteristics under dynamic speed conditions is analyzed in depth; the JNU (Jiangnan University) dataset is used to further verify the adaptability of the proposed model under different data distributions and variable speed operating conditions; at the same time, real bearing vibration signals are collected based on the Mechanical Fault Simulation (MFS) experimental platform of the Rotating Machinery Intelligent Operation and Maintenance Laboratory to fully reflect the impact of complex operating conditions such as speed fluctuations on the bearing operating status in actual industrial scenarios.

[0068] Extensive experiments on three public benchmark datasets—CWRU, JNU, and MFS—validate the effectiveness and superiority of the proposed method. In single-source domain transfer tasks, ProtoRAP achieves average accuracies of 98.19%, 93.11%, and 95.45%, respectively, representing improvements of 2.50%, 5.07%, and 2.65% over the best comparison method. In multi-source domain transfer tasks, the method demonstrates superior multi-source knowledge fusion capabilities, achieving average accuracies of 99.40%, 96.30%, and 98.18%, respectively, representing improvements of approximately 2.97%, 5.83%, and 2.36% over the best comparison method. Ablation experiments, using t-SNE combined with KDE visualization analysis, intuitively verify the irreplaceable role of each core module in feature alignment and discriminative representation learning. Parameter sensitivity analysis shows that the model has good robustness within the optimal hyperparameter range. Confusion matrix and feature visualization results further confirm ProtoRAP's advantages in feature clustering compactness and inter-class separability in cross-domain scenarios. This study provides a new paradigm for the application of cue learning technology in the field of fault diagnosis in intelligent manufacturing. The proposed fault prototype cue generation and context-aware retrieval mechanism offers an effective solution to the problem of domain knowledge adaptation under limited labeled data conditions in manufacturing, verifying the enormous application potential of cue learning technology in industrial fault diagnosis scenarios. Future research will explore the fusion mechanism of multimodal sensor information such as vibration, acoustics, and temperature with cue learning to construct a more comprehensive semantic representation of faults, thereby further improving the model's fault diagnosis performance.

[0069] In summary, this embodiment innovatively introduces the prompt learning paradigm from the era of large language models into the field of intelligent manufacturing fault diagnosis. Addressing the domain offset problem faced in diagnosing rolling bearing faults under varying operating conditions in complex industrial environments, it proposes a cross-domain fault diagnosis framework (ProtoRAP) based on prototype retrieval and adaptation. This framework explicitly encodes source domain fault knowledge into semantic prototype prompt vectors through a self-paced prototype prompt generation module (SPPG), achieves dynamic matching between target domain samples and source domain fault knowledge using a context-aware prompt retrieval unit (CAPR), and utilizes a prompt-guided refined attention module (PGRA) to achieve deep fusion of prompts and features. This effectively solves the performance bottleneck of implicit feature alignment and insufficient generalization ability under varying operating conditions in traditional domain adaptation methods.

[0070] Example 4: This invention provides a method for diagnosing rolling bearing faults using a hybrid cueing learning network, based on the aforementioned embodiments. It is applied to the hybrid cueing learning network for cross-domain fault diagnosis of rolling bearings provided in the previous embodiments. See also... Figure 4The flowchart shown illustrates a hybrid cueing learning network for cross-domain fault diagnosis of rolling bearings. This hybrid cueing learning network includes the following steps: Step S402: The self-paced prototype cue generator explicitly anchors abstract cue vectors to semantic prototypes of typical bearing failure modes, and progressively filters high-confidence samples by combining uncertainty-driven course learning strategies. Step S404: The context-aware cue retrieval system uses the coarse-grained context representation of the intermediate layer of the network to capture transdomain stable vibration characteristics under different rotational speeds and loads. In step S406, the prompting and guidance refinement and adaptation module realizes adaptive weighted aggregation of retrieval prompts through a multi-head cross-attention mechanism, and uses a gating network to inversely modulate the target domain operating condition features to selectively enhance the fault sensitivity dimension.

[0071] This invention provides a hybrid cueing learning network-based method for cross-domain fault diagnosis of rolling bearings. A ProtoRAP framework is proposed, which achieves cross-domain fault diagnosis of rolling bearings through prototype-guided knowledge encoding, context-aware dynamic retrieval, and gated feature adaptation. The hybrid cueing learning network and system provided in this embodiment significantly outperform existing state-of-the-art methods in terms of diagnostic accuracy and cross-domain stability across multiple rolling bearing fault datasets under varying operating conditions, validating its effectiveness and robustness in complex industrial scenarios.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the rolling bearing fault diagnosis method using a hybrid cueing learning network for cross-domain fault diagnosis of rolling bearings described above can be referred to the corresponding process in the aforementioned embodiments of the hybrid cueing learning network for cross-domain fault diagnosis of rolling bearings, and will not be repeated here.

[0073] Example 5: This invention also provides a hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings, used to run the above-described hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings in the rolling bearing fault diagnosis method; see also Figure 5 The diagram shows a hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings. The hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings includes a memory 100 and a processor 101. The memory 100 is used to store one or more computer instructions, and the one or more computer instructions are executed by the processor 101 to implement the hybrid prompting learning network method for cross-domain fault diagnosis of rolling bearings.

[0074] Furthermore, Figure 5The hybrid prompting learning system shown for cross-domain fault diagnosis of rolling bearings also includes a bus 102 and a communication interface 103, with the processor 101, communication interface 103 and memory 100 connected via the bus 102.

[0075] The memory 100 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 102 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0076] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 100, and processor 101 reads information from memory 100 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0077] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0078] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0080] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hybrid cueing learning network for cross-domain fault diagnosis of rolling bearings, characterized in that, include: Self-paced prototype prompt generator, context-aware prompt retrieval unit, and prompt guidance refinement and adaptation module; The self-paced prototype cue generator is used to progressively filter high-confidence samples by explicitly anchoring abstract cue vectors to semantic prototypes of typical bearing failure modes, combined with an uncertainty-driven course learning strategy. The context-aware cue retrieval system is used to capture transdomain stable vibration characteristics under different rotational speeds and loads by utilizing the coarse-grained context representation of the intermediate layer of the network. The refinement and adaptation module for the prompt guidance is used to achieve adaptive weighted aggregation of retrieval prompts through a multi-head cross-attention mechanism, and to selectively enhance the fault sensitivity dimension by using a gating network to inversely modulate the target domain operating condition features.

2. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to claim 1, characterized in that, The self-paced prototype cue generator is used to establish an explicit semantic mapping between cue and failure mode in the feature space through a prototype anchoring strategy, and to build a cue knowledge base. The self-paced prototype cue generator is also used to select high-confidence samples by incorporating an uncertainty-driven adaptive curriculum learning strategy.

3. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to claim 2, characterized in that, The context-aware cue retrieval device is used to extract coarse-grained contextual representations of target domain samples and dynamically retrieve the K most relevant cue from the cue knowledge base using cosine similarity, where K is an integer greater than or equal to 1.

4. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to claim 1, characterized in that, The prompt guidance refinement and adaptation module is used to generate adaptive prompts by retrieving prompts through multi-head cross-attention fusion, refine target domain features using a gating mechanism, selectively enhance fault-related dimensions and suppress domain-specific noise, and finally obtain domain-invariant representations.

5. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to claim 1, characterized in that, The self-paced prototype prompt generator includes: a first collaborative submodule, a second collaborative submodule, and a third collaborative submodule; The first collaborative submodule is used to dynamically filter high-confidence samples through Shannon entropy measurement based on the uncertainty-aware course learning mechanism; The second collaborative submodule is used to establish an explicit semantic mapping between the cue vector and the physical fault mode based on the prototype anchoring strategy; The third collaborative submodule is used to maximize the feature differences between cue vectors based on orthogonality regularization.

6. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to claim 2, characterized in that, The context-aware cue retrieval system is used to construct a coarse-grained context representation using intermediate-layer features of target domain samples; wherein, the construction of the coarse-grained context representation preserves cross-domain stable basic signal patterns; The context-aware prompt retrieval device is also used to complete the prompt retrieval process by calculating the similarity between the coarse-grained context representation and each prompt in the prompt knowledge base.

7. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to claim 1, characterized in that, The prompt-guided refinement and adaptation module is constructed with a two-stage cascaded reasoning architecture. The refinement and adaptation module for prompt guidance is used to transform passively retrieved prompts into active knowledge guidance signals through attention prompt fusion and prompt gating feature refinement.

8. The hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings according to any one of claims 1-7, characterized in that, The hybrid cue learning network for cross-domain fault diagnosis of rolling bearings is used to construct a cue knowledge base through multi-objective joint optimization during the source domain training phase. The hybrid cue learning network for cross-domain fault diagnosis of rolling bearings is also used to achieve dynamic retrieval and adaptive fusion of knowledge in the target domain reasoning stage through the context-aware cue retrieval device and the cue-guided refinement and adaptation module.

9. A hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings, characterized in that, The hybrid prompting learning system for cross-domain fault diagnosis of rolling bearings includes: the hybrid prompting learning network for cross-domain fault diagnosis of rolling bearings as described in any one of claims 1-8.

Images