Self-correction noise tag method for few-sample fault diagnosis of mechanical equipment
By building a bidirectional inference multimodal neural network to filter and correct noise labels, the problem of noise label influence in mechanical equipment small samples is solved, and high-precision fault diagnosis is achieved.
Patent Information
- Application Number
- CN202510665364.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-22
- Publication Date
- 2025-08-26
AI Technical Summary
In the diagnosis of few samples of mechanical equipment, noise labels are common, resulting in a degradation of model performance, and the existing technology is difficult to effectively correct, affecting the diagnostic accuracy.
The self-corrected noise label method is adopted to build a bidirectional inference multimodal neural network, combining the reverse inference module and the forward inference module, filter and correct the noise labels, and use feature extractors and text encoder to extract features, calculate similarity probability, and predict and loss optimization through multi-layer perceptrons.
It significantly improves the fault diagnosis accuracy of mechanical equipment under noise label conditions, and achieves high-precision fault diagnosis of few samples.
Smart Images
Figure CN120541524A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fault diagnosis, and more specifically, relates to a self-correcting noise label method for few-sample fault diagnosis of mechanical equipment. Background Art
[0002] In modern industrial production lines, mechanical equipment often operates under harsh conditions such as high temperature, high humidity, and strong vibration, leading to accelerated wear, fatigue, and corrosion of critical components. These problems can easily lead to unplanned downtime and even fault propagation, resulting in significant economic losses. Therefore, mechanical equipment fault diagnosis is crucial to ensuring production safety and efficiency. Deep learning-based fault diagnosis algorithms have been widely used in recent years due to their high accuracy, but their effectiveness relies heavily on sufficient fault samples. However, in real industrial scenarios, faults occur infrequently and samples are scarce, posing a significant challenge to small-sample fault diagnosis.
[0003] While few-shot learning has made progress in fault diagnosis, it still faces the serious challenge of label noise in industrial applications. Existing research often assumes accurate sample labels, but label errors and noise are common in industrial datasets. Especially when there are very few fault samples (e.g., only one sample per fault type), the negative impact of mislabeling on model performance is particularly significant, even exceeding the challenges of unlabeled data. While unlabeled data can be mitigated through techniques such as pseudo-labeling or distribution alignment, noisy labels directly mislead model learning and severely reduce diagnostic accuracy.
[0004] Therefore, how to achieve self-correction in a noisy label environment and ensure accurate diagnosis of mechanical equipment failures has become a core technical problem restricting the development of intelligent operation and maintenance of mechanical equipment. Summary of the Invention
[0005] In view of the limitations of existing technologies in noisy label environments, the present invention proposes a self-correcting noise label method, aiming to achieve accurate fault diagnosis of mechanical equipment under few-sample conditions.
[0006] To achieve the above objectives, the present invention provides a self-correcting noise labeling method for few-sample fault diagnosis of mechanical equipment, comprising:
[0007] S1. Mechanical equipment dataset in experimental environment The diagnostic model is trained on the machine and transferred to the engineering machinery equipment with very few fault samples and noisy labels. Then fine-tune and perform fault diagnosis; the known fault samples in engineering machinery equipment are defined as Contains clean label samples and noisy label samples
[0008] S2. Construct a bidirectional reasoning multimodal neural network, integrating the reverse reasoning module and the forward reasoning module; the reverse reasoning module is implemented through the reverse text encoder ψ R and reverse prompt template Filter the noisy labels in the input data, and the forward reasoning module relies on the forward text encoder ψ F and positive hint templates Correct the filtered labels, and both modules share the feature extractor φ; in the source domain Pre-train the bidirectional reasoning multimodal neural network to enable it to have bidirectional reasoning capabilities, and calculate the pre-training loss
[0009] S3. Target domain A small sample of known fault types (Including clean samples and noise-like ) is input into the pre-trained bidirectional reasoning multimodal neural network, using the reverse prompt template Generate samples Semantic description of category names, combined with feature extractor φ and reverse text encoder ψ R , extract data features respectively and text features Use similarity calculation to evaluate the probability that the sample does not belong to a certain state type Then the given label of each sample Corresponding Perform threshold comparison. If it exceeds the preset threshold ε, it is determined to be a noise label and a noise mask is generated. Finally, a filtering algorithm is used to filter out the noise label.
[0010] S4. Through forward hint template Target domain Noisy label samples Generate semantic descriptions of category names by combining feature extractor φ and forward text encoder ψ F , extract samples separately Data characteristics and text features And calculate the similarity probability between the two Based on this, the wrong labels are redefined to correct the noise labels;
[0011] S5. Corrected raw data Through the feature extractor φ and the forward text encoder ψ F , respectively extract feature data features and text features And calculate the similarity probability between the two Design a multi-layer perceptron to learn data features Make predictions and get predicted probabilities Through the nonlinear fusion mechanism, the similarity probability and predicted probability Integrated into fusion probability Based on fusion probability Calculating classification loss And label prediction loss label pred ; Introducing the contrastive learning paradigm, based on the fusion probability Select anchor points from each category and positive samples Negative samples It is identified by the reverse reasoning module and the contrast loss is calculated By weighted fusion classification loss and contrast loss Total loss Use this as the optimization target to train a multilayer perceptron;
[0012] S6. In the diagnosis stage, without contrastive learning components, after feature extractor φ, forward text encoder ψ F Based on the synergy of multi-layer perceptron and nonlinear fusion mechanism, the probability of fault type is predicted. in accordance with Diagnose specific fault types of mechanical equipment.
[0013] Furthermore, the expression for similarity calculation evaluation is The noise mask generation is calculated as The calculation of the filtering algorithm is
[0014] Furthermore, the similarity probability The calculation is Redefine the calculation of the original error label as in, are the corrected label set and category name set respectively.
[0015] Furthermore, a multilayer perceptron is designed to calculate data features The similarity probability is The nonlinear fusion mechanism is Where ⊙ represents the Hadamard product, σ and v are hyperparameters; classification loss The calculation is and label predictions are calculated as Under the contrastive learning paradigm, the basis fusion probability Select the class with the highest confidence score from each class samples as anchor points The selection mechanism is And the confidence ranking from arrive The sample is designated as a positive sample Specify the policy as Extracting data features through reverse reasoning module and text features And calculate the similarity probability between the two Then adopt the identification mechanism as Filter out negative samples Contrastive loss The calculation is where <·,·> represents cosine similarity and τ represents a hyperparameter.
[0016] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art.
[0017] Existing few-shot learning methods and other related technologies usually assume that the labels of a small number of known fault samples are accurate. However, in actual industrial environments, label errors and noise are common in data sets. Especially when there are very few fault samples (such as only one sample for each type of fault), the negative impact of mislabeling on model performance is particularly significant, even exceeding the challenges posed by unlabeled data. Unlabeled data can be alleviated through techniques such as pseudo-labeling or distribution alignment, while noisy labels will directly mislead model learning and seriously reduce diagnostic accuracy. To this end, the present invention proposes a self-correcting noise label method for few-shot fault diagnosis of mechanical equipment. By innovatively combining the reverse reasoning module and the forward reasoning module, the noise label is effectively corrected, its interference is significantly reduced, and high-precision mechanical equipment fault diagnosis under few-shot conditions is achieved. In summary, the present invention, with its unique self-correcting noise label technology, provides an efficient and reliable solution for mechanical equipment fault diagnosis under noisy labels and few-shot scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] To make the technical details of this technical solution clearer and easier to understand, the following briefly describes the basic details of the relevant drawings of this technical solution. It should be noted that the drawings shown here are only schematic diagrams of some typical embodiments of this technology. Those skilled in the relevant art can deduce the supporting drawings required for other implementation methods based on the illustrations without any creative work.
[0019] Figure 1 Schematic diagram of a self-correcting noise labeling method provided by an embodiment of the present invention;
[0020] Figure 2 yes Figure 1 Schematic diagram of multilayer perceptron training in ; DETAILED DESCRIPTION
[0021] In order to make the purpose, technical solutions and advantages of the present invention clearer and more thorough, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be emphasized that the specific embodiments described herein are only used to illustrate the present invention and do not limit its scope. In addition, the technical features involved in each embodiment can be flexibly combined to give full play to their synergistic effects, provided that they do not conflict with each other.
[0022] See also Figure 1 and Figure 2 The present invention provides a self-correcting noise label method for few-sample fault diagnosis of mechanical equipment. Figure 1 As shown, the self-correction noise label method provided by the present invention has the following specific steps:
[0023] Step 1: Mechanical equipment dataset in experimental environment The diagnostic model is trained on the machine and transferred to the mechanical equipment with few fault samples and noisy labels in the engineering environment. Achieve high-precision fault diagnosis; a small number of known fault samples in engineering machinery equipment are defined as Involving clean label samples and noisy label samples
[0024] Specifically, the mechanical equipment dataset in the experimental environment is recorded as the source domain Used for initial training of diagnostic models; engineering machinery equipment data is defined as the target domain A small number of known fault samples In, x i Represents data, y i Represents the label, t i Represents the state type name, C is the number of state types, K is the number of samples of each known state type; known fault samples Contains clean label samples and noisy label samples Among them, clean samples Label y i and state type name t i With data x i Complete match, while the noise label sample Noisy labels due to annotation errors and noise state type names
[0025] Step 2: Construct a dual-backward inference multimodal neural network, including a backward inference module and a forward inference module; the backward inference module filters the noise labels in the input data, while the forward inference module further corrects the filtered labels; the backward inference module and the forward inference module each contain a dedicated backward text encoder ψ Rand the forward text encoder ψ F , and share a feature extractor φ, and each contains a dedicated design reverse prompt template and positive hint templates To describe the category name in the known fault sample Utilize the source domain Pre-train the bidirectional reasoning multimodal neural network to form a pre-training loss
[0026] Specifically, the source domain By dataset Composition, where x i represents input data, t i Indicates positive text, Indicates reverse text. Define two prompt templates: forward prompt template Represented as {"The state is not [type]"}, reverse prompt template Represented as {"The state is not [type]"}; source domain After feature extractor φ, forward text encoder ψ F and the reverse text encoder ψ R The feature extraction operation is as follows:
[0027]
[0028]
[0029]
[0030] In the formula and yes The vector in , D is the feature dimension. The dual reverse inference multimodal neural network uses two loss functions to optimize the model; the first loss function is the binary opposition loss Used to align data features Its reverse text feature Define m by matching degree ij for (When i=j, it indicates an opposite match, i.e. the reverse text is semantically opposite to the data) or (When i=j, it means matching but irrelevant, that is, the reverse text is correct for the data but has no semantic connection), guiding the alignment of data and reverse text in the feature space, binary opposition loss function Defined as:
[0031]
[0032] Where, represents the matching probability between the i-th data and the j-th reverse text, which is calculated as
[0033]
[0034] Where <·,·> represents the inner product of two vectors, θ is a learnable parameter; the second loss function is the text semantic opposition loss Ensure positive text features With reverse text features Maintaining semantic opposition in feature space is defined as:
[0035]
[0036] In the formula, ||·||2 represents The distance function, when and When pointing in the opposite direction, Decreases to 0. Total pre-training loss for:
[0037]
[0038] Step 3: Set the target domain A small sample of known fault types (Including clean samples and noise-like ) is input into the pre-trained dual reverse inference multimodal neural network, and the reverse prompt template is used For samples Generate semantic descriptions from the category names and use the feature extractor φ and the reverse text encoder ψ R , extract samples Data characteristics and text features By comparing the similarity between text features and data features, the probability that the sample does not belong to a certain category is calculated The given label of each sample Corresponding The value is compared. If it exceeds the predetermined threshold ε, it is determined to be a noise label and a noise mask is generated; a filtering algorithm is used to filter these noise labels;
[0039] Specifically, the target domain A small sample of known fault types Input bidirectional inference multimodal neural network, respectively through the feature extractor φ and the reverse text encoder ψ R Extracting data features and text features The operation is as follows:
[0040]
[0041]
[0042] The probability that a sample does not belong to a certain state type is evaluated by similarity calculation:
[0043]
[0044] Compare the given labels for each sample and Generate the noise mask by the corresponding value of :
[0045]
[0046] The filtering algorithm is used to filter the noise label data set. The expression is:
[0047] x noist =Filter(x, Mask)
[0048] Step 4: Use the forward prompt template Target domain Noisy label samples Generate semantic descriptions of category names by combining feature extractor φ and forward text encoder ψ F , extract samples separately Data characteristics and text features And calculate the data features and text features The similarity probability between According to the similarity probability Redefine the original wrong label to correct the noisy label;
[0049] Specifically, the noise label samples Input feature extractor φ and forward text encoder ψ R , respectively extract feature data features and text features The operation is as follows:
[0050]
[0051]
[0052] By calculating text features and data characteristics Similarity, evaluate the probability that the sample belongs to a certain state type
[0053]
[0054] According to the similarity probability Redefine the original error label:
[0055]
[0056] Where, Represent the corrected label set and category name set respectively, and Refinement represents the sampling process for correcting labels.
[0057] Step 5: Corrected raw data Through the feature extractor φ and the forward text encoder ψ F Extract data features separately and text features And calculate the similarity probability between the two like Figure 2 As shown, a multi-layer perceptron is designed to predict the features of the data and generate prediction probabilities. And through the nonlinear fusion mechanism, the similarity probability and predicted probability Fusion is fusion probability Based on fusion probability Calculating classification loss And label prediction loss label pred ; Using contrastive learning paradigm, based on fusion probability Select the class with the highest confidence score from each class samples as anchor points Confidence ranking from arrive The sample is designated as a positive sample And identify negative samples through the reverse reasoning module Compute contrast loss based on anchor points, positive samples, and negative samples The classification loss and contrast loss Add up to form the total loss Use this as the optimization target to train a multilayer perceptron;
[0058] Specifically, the corrected original data Through the feature extractor φ and the forward text encoder ψ F Extract data features separately and text features And calculate the similarity probability
[0059]
[0060]
[0061]
[0062] Multilayer Perceptron for data features Make predictions and generate predicted probabilities
[0063]
[0064] Through the nonlinear fusion mechanism, the similarity probability and predicted probability Fusion is fusion probability
[0065]
[0066] Where ⊙ represents the Hadamard product, σ and υ are hyperparameters. Compute the classification loss and label predictions:
[0067]
[0068]
[0069] Under the contrastive learning paradigm, according to the fusion probability Select the class with the highest confidence score from each class samples as anchor points Confidence ranking from arrive The sample is designated as a positive sample
[0070]
[0071]
[0072] Negative samples Then, reverse reasoning is used to identify:
[0073]
[0074]
[0075]
[0076]
[0077] Calculating contrastive loss for:
[0078]
[0079] Where <·,·> represents cosine similarity and τ represents hyperparameter. and contrast loss Add up to the total loss
[0080]
[0081] Where γ is determined by experiment, and the total loss is Train a multilayer perceptron as the optimization objective.
[0082] Step 6: In the diagnosis phase, without contrast learning components, the feature extractor φ and the forward text encoder ψ F Based on the synergy of multi-layer perceptron and nonlinear fusion mechanism, the probability of fault type is predicted. in accordance with Diagnose specific fault types of mechanical equipment.
[0083] To highlight the advantages of our method for fault diagnosis under noisy label conditions and with a limited number of fault samples, we conducted comparative experiments with several advanced methods, including deep convolutional neural networks, twin networks for few-shot learning, matching networks, and relational networks. Table 1 shows the comparison of their few-shot fault diagnosis accuracy under noisy label conditions. The table shows that our method achieves significantly higher fault diagnosis accuracy than the other four methods.
[0084] Table 1 method Fault diagnosis accuracy Deep Convolutional Neural Networks 60.68 Siamese Network 78.96 Matching Network 83.41 Relationship Network 84.64 Method of the present invention 96.14
[0085] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A self-correcting noise label method for few-sample fault diagnosis of mechanical equipment, characterized by: include: S1. Mechanical equipment dataset in experimental environment The diagnostic model is trained on the machine and transferred to the engineering machinery equipment with very few fault samples and noisy labels. Then make fine adjustments and perform fault diagnosis; The known fault samples in engineering machinery equipment are defined as Contains clean label samples and noisy label samples S2. Construct a bidirectional reasoning multimodal neural network, integrating the reverse reasoning module and the forward reasoning module; the reverse reasoning module is implemented through the reverse text encoder ψ R and reverse prompt template Filter the noisy labels in the input data, and the forward reasoning module relies on the forward text encoder ψ F and positive hint templates Correct the filtered labels. Both modules share the feature extractor φ; In the source domain Pre-train the bidirectional reasoning multimodal neural network to enable it to have bidirectional reasoning capabilities, and calculate the pre-training loss S3. Target domain A small sample of known fault types (Including clean samples and noise-like ) is input into the pre-trained bidirectional reasoning multimodal neural network, using the reverse prompt template Generate samples Semantic description of category names, combined with feature extractor φ and reverse text encoder ψ R , extract data features respectively and text features Use similarity calculation to evaluate the probability that the sample does not belong to a certain state type Then the given label of each sample Corresponding Perform threshold comparison. If it exceeds the preset threshold ε, it is determined to be a noise label and a noise mask is generated. Finally, a filtering algorithm is used to filter out the noise label. S4. Through forward hint template Target domain Noisy label samples Generate semantic descriptions of category names by combining feature extractor φ and forward text encoder ψ F , extract samples separately Data characteristics and text features And calculate the similarity probability between the two Based on this, the wrong labels are redefined to correct the noise labels; S5. Corrected raw data Through the feature extractor φ and the forward text encoder ψ F , respectively extract feature data features and text features And calculate the similarity probability between the two Design a multi-layer perceptron to learn data features Make predictions and get predicted probabilities Through the nonlinear fusion mechanism, the similarity probability and predicted probability Integrated into fusion probability Based on fusion probability Calculating classification loss And label prediction loss label pred ; Introducing the contrastive learning paradigm, based on the fusion probability Select anchor points from each category and positive samples Negative samples It is identified by the reverse reasoning module and the contrast loss is calculated By weighted fusion classification loss and contrast loss Total loss Use this as the optimization target to train a multilayer perceptron; S6. In the diagnosis stage, without contrastive learning components, after feature extractor φ, forward text encoder ψ F Based on the synergy of multi-layer perceptron and nonlinear fusion mechanism, the probability of fault type is predicted. in accordance with Diagnose specific fault types of mechanical equipment.
2. The self-correcting noise label method for few-sample fault diagnosis of mechanical equipment according to claim 1 is characterized in that: The expression for similarity calculation evaluation is The noise mask generation is calculated as The calculation of the filtering algorithm is x noisy =Filter(x,Mask).
3. The self-correcting noise label method for few-sample fault diagnosis of mechanical equipment according to claim 1 is characterized in that: Similarity probability The calculation is Redefine the calculation of the original error label as in, are the corrected label set and category name set respectively.
4. The self-correcting noise label method for few-sample fault diagnosis of mechanical equipment according to claim 1 is characterized in that: Design a multilayer perceptron to calculate data features The similarity probability is The nonlinear fusion mechanism is Where ⊙ represents the Hadamard product, σ and υ are hyperparameters; classification loss The calculation is and label predictions are calculated as Under the contrastive learning paradigm, the basis fusion probability Select the class with the highest confidence score from each class samples as anchor points The selection mechanism is And the confidence ranking from arrive The sample is designated as a positive sample Specify the policy as Extracting data features through reverse reasoning module and text features And calculate the similarity probability between the two Then adopt the identification mechanism as Filter out negative samples Contrastive loss The calculation is where <·,·> represents cosine similarity and τ represents a hyperparameter.