Open set fault diagnosis method based on two-stage entropy perception consensus domain confrontation framework

Abstract

The invention relates to an open set fault diagnosis method based on a two-stage entropy perception consensus domain confrontation framework, and the method comprises the following steps: S1, obtaining a training data set of a rotating machine, the training data set comprising a source domain data set and a target domain data set; s2, constructing a model framework, wherein the model framework comprises a feature extractor and loss functions of five modules; the feature extractor comprises a convolutional neural network, a bidirectional long-short-term memory network and a multi-head attention mechanism which are arranged in sequence; the five modules comprise an original classifier module, a two-stage entropy sensing module, an auxiliary classifier module, a self-supervised clustering module and a consensus domain confrontation module; inputting the training data set into a feature extractor, and adjusting parameters of a model framework by using an optimization algorithm to complete training of the whole framework; and S3, obtaining a test data set, inputting the test data set into the trained model framework, and outputting a fault diagnosis result. The problem that an existing domain self-adaption method cannot effectively process unknown faults is solved, and the open set fault diagnosis method is achieved.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, and specifically to an open set fault diagnosis method based on a two-level entropy perception consensus domain adversarial framework. Background Technology

[0002] Rotating machinery fault diagnosis refers to monitoring equipment operating status, promptly identifying potential problems, and optimizing maintenance strategies to ensure equipment safety, extend service life, and reduce operational risks. Existing domain-adaptive fault diagnosis methods are typically based on the closed-set assumption, which assumes that the source and target domains have identical fault category spaces. This fails to adequately consider the possibility of unknown faults occurring in the target domain in real-world industrial scenarios. Once an unknown fault occurs, the performance of traditional domain-adaptive methods often degrades significantly, struggling to accurately distinguish between known and unknown categories, leading to misclassification of unknown faults as known categories and severely impacting diagnostic reliability. Therefore, it is necessary to research a novel method that can effectively handle open-set fault diagnosis to overcome the limitations of existing methods in effectively handling unknown faults.

[0003] It should be noted that the above content falls within the inventor's technical knowledge and does not necessarily constitute prior art. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the prior art and provide an open set fault diagnosis method based on a two-level entropy-aware consensus domain adversarial framework, thereby solving the problem that existing domain adaptive methods cannot effectively handle unknown faults.

[0005] The present invention achieves the above objectives by adopting the following technical solutions: An open-set fault diagnosis method based on a two-level entropy-aware consensus domain adversarial framework includes the following steps: S1. Obtain the training dataset for the rotating machinery. The training dataset includes a source domain dataset and a target domain dataset; S2, Building the Model Framework The model framework includes a feature extractor and loss functions for five modules. The feature extractor consists of a convolutional neural network, a bidirectional long short-term memory network, and a multi-head attention mechanism arranged sequentially. The five modules include a primary classifier module, a two-level entropy perception module, an auxiliary classifier module, a self-supervised clustering module, and a consensus domain adversarial module. The training dataset is input into the feature extractor, and then the parameters of the entire model framework are adjusted using an optimization algorithm to optimize the loss function of each module, thereby completing the training of the entire framework. S3. Collect vibration signals of rotating machinery on site to obtain test datasets, input the test datasets into the trained model framework, and output fault diagnosis results.

[0006] In step S1, the source domain dataset is denoted as: ; The target domain dataset is denoted as:

[0007] In the formula, Represents the source domain, which contains the corresponding Categories There are labeled samples, and these samples follow a probability distribution. ; Describe the target domain, which contains elements that follow a probability distribution. of There are 10 unlabeled samples; since the source and target domains were collected under different working conditions, their probability distributions differ significantly, i.e. .

[0008] In step S2, the feature extractor establishes a mapping relationship between vibration signals and diagnostic results, given an input sample. The expression for the operation process of the feature extractor is:

[0009] In the formula, This represents the trainable parameters.

[0010] In step S2, the original classifier module outputs the output of the feature extractor to... Among the categories, Represents a known class, the first Each sample represents an unknown class; the original classifier module outputs a one-dimensional logits vector for each sample. The Softmax function converts a logarithmic vector into a logarithmic vector. A probability distribution of dimension 2, its expression is:

[0011] For classification tasks, the original classifier module is trained on source domain samples to learn discriminative features of known categories. The original classification loss is expressed as:

[0012] In the formula, Let cross-entropy be the loss function. It is the first A one-hot vector of dimension.

[0013] In step S2, the dual-level entropy perception module includes sample-level entropy perception and domain-level entropy perception. The sample-level entropy perception adaptively assigns weights to target samples. The sample-level entropy perception first quantifies the classification difficulty of each target sample using standard entropy, and its expression is:

[0014] Then, the weighted function based on entropy Assign weights to these samples and introduce these weights into the adversarial loss function; The domain-level entropy perception dynamically adjusts the decision boundary; this level utilizes average entropy. To quantify the overall classification uncertainty in the target domain, the classification threshold will be dynamically adjusted through batch processing, and its expression is as follows:

[0015] In the formula, This represents the momentum in SGD. Indicates the current batch The average entropy; The adversarial loss is introduced based on sample-level entropy weights and dynamic thresholds, and its expression is: .

[0016] In step S2, the auxiliary classifier module is used to enhance the discriminability of known categories. The auxiliary classifier module consists of a fully connected layer and an output layer. The output probability is generated by a leaky-softmax function, making the total probability less than 1. Its expression is:

[0017] In the formula, It is a logits vector. It is the first One element; The optimized loss function of the auxiliary classifier module is expressed as: .

[0018] In step S2, the self-supervised clustering module is used to mine structural information in the target domain, and the output of the feature extractor is processed... After regularization and normalization, the vector is fed into the prototype classifier module; the prototype classifier module obtains the representative vector. Constructing prototype vectors ; Store the features of all target domain samples in a memory bank Each slot Storage number The features corresponding to the nth target sample; in the nth Update the corresponding memory after each small batch iteration. ,in This represents the sample index within a small batch; This represents the target feature in a mini-batch; Represents the set of the source prototype and all target features; for The probability calculation expression for its characteristics being similar to those of the fault prototype or other target characteristics is as follows:

[0019] Furthermore, a clustering loss function is used to optimize the clustering of target samples and help identify edge samples, which is expressed as: .

[0020] In step S2, the consensus domain adversarial module integrates the outputs of multiple modules through a consensus probability metric to achieve consistency of known categories and enhances domain alignment through adversarial learning. The probability of a sample belonging to a known class is calculated based on the estimated probability of the original classifier module belonging to the unknown class. The expression is as follows:

[0021] For the auxiliary classifier, the output is the high probability confidence score of the known class. The probability expression for each sample belonging to the known class is:

[0022] In the formula, The larger the value, the greater the probability that the sample belongs to a known category; conversely, the smaller the value, the greater the probability that the sample comes from an unknown category. In the self-supervised clustering module, the clustering confidence is defined as... ,in The similarity probability is calculated by taking the maximum probability between the target sample's features and the prototype vector to measure the closeness to the known category. Based on the above probabilities and confidence levels, the consensus probability index is calculated. Defined as:

[0023] Therefore, the consensus domain adversarial loss can be expressed as:

[0024] The training process of the model framework can be represented as follows:

[0025]

[0026]

[0027] In the formula, For the total loss, These are the weighting coefficients of the two-level entropy-sensing loss function. For the weight coefficients of the loss function of the auxiliary classifier, These are the weight coefficients of the self-supervised clustering loss function. These are the weight coefficients for the consensus domain adversarial loss function; These are the trainable parameters of the feature extractor. These are the trainable parameters of the original classifier module. These are the trainable parameters for the auxiliary classifier module.

[0028] The present invention, by employing the above method, can bring the following beneficial effects: This method fully considers the presence of unknown faults in the target domain, thereby enabling the alignment of known categories and the identification of unknown categories. It effectively overcomes the limitation of existing domain adaptive diagnostic methods in handling unknown faults and is suitable for open-set fault diagnosis. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the principle of the open set fault diagnosis method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the feature extractor of the present invention; Figure 3 This is the experimental platform for the SEU dataset of this invention; Figure 4 This serves as the experimental platform for the SC dataset of this invention. Figure 5 A photograph of the damaged output tooth surface of the high-speed shaft; Figure 6 This is a visualization of the characteristics of different methods in this invention; In the diagram, 1 is the motor controller, 2 is the motor, 3 is the planetary gearbox, 4 is the parallel gearbox, 5 is the brake, 6 is the brake controller, 7 is the magnetic powder brake, 8 is the parallel shaft gearbox, 9 is the bearing housing, 10 is the rotor disk, 11 is the torque sensor, and 12 is the three-phase drive motor. Detailed Implementation

[0030] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0031] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0032] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0033] like Figure 1-6 As shown, an open-set fault diagnosis method based on a two-level entropy-aware consensus domain adversarial framework includes the following steps: S1. Obtain the training dataset for the rotating machinery. Specifically, accelerometers are used to collect vibration signals of different types of faults in the core components of rotating machinery under different operating conditions, and the vibration signals are segmented to form a training dataset; the training dataset includes a source domain dataset and a target domain dataset. The source domain dataset is denoted as: ; The target domain dataset is denoted as:

[0034] In the formula, Represents the source domain, which contains the corresponding Categories There are labeled samples, and these samples follow a probability distribution. ; Describe the target domain, which contains elements that follow a probability distribution. of There are 10 unlabeled samples; since the source and target domains were collected under different working conditions, their probability distributions differ significantly, i.e. .

[0035] S2, Building the Model Framework

[0036] The model framework includes a feature extractor and loss functions for five modules; S21, the feature extractor ( This includes a Convolutional Neural Network (CNN), a Bidirectional Long Short-Term Memory (Bi-LSTM) network, and a multi-head attention mechanism, arranged sequentially. CNN captures local structural features; Bi-LSTM captures bidirectional dependencies in time series data; and the multi-head attention mechanism highlights key features, improving the feature extractor's ability to model relationships between features at different time steps. Given an input sample... Trainable weights of the feature extractor The final feature representation is obtained. The running expression for the feature extractor is:

[0037] S22. The five modules include the original classifier module, the two-level entropy perception module, the auxiliary classifier module, the self-supervised clustering module, and the consensus domain adversarial module. The training dataset is input into the feature extractor, and then the parameters of the entire model framework are adjusted using optimization algorithms so that the loss function of each module achieves the optimal effect, thereby completing the training of the entire framework. The original classifier module outputs the output of the feature extractor to... Among the categories, Represents a known class, the first Each sample represents an unknown class; the original classifier module outputs a one-dimensional logits vector for each sample. The Softmax function converts a logarithmic vector into a logarithmic vector. A probability distribution of dimension 2, its expression is:

[0038] For classification tasks, the original classifier module is trained on source domain samples to learn discriminative features of known categories. The original classification loss is expressed as:

[0039] In the formula, Let cross-entropy be the loss function. It is the first A one-hot vector of dimension.

[0040] The dual-level entropy perception module includes sample-level entropy perception and domain-level entropy perception. The sample-level entropy perception adaptively assigns weights to target samples. Sample-level entropy perception first quantifies the classification difficulty of each target sample using standard entropy, and its expression is:

[0041] Then, the weighted function based on entropy Assign weights to these samples and introduce these weights into the adversarial loss function; The domain-level entropy perception dynamically adjusts the decision boundary; this level utilizes average entropy. To quantify the overall classification uncertainty in the target domain, the classification threshold will be dynamically adjusted through batch processing, and its expression is as follows:

[0042] In the formula, This represents the momentum in SGD (Stochastic Gradient Descent). Indicates the current batch The average entropy; The adversarial loss is introduced based on sample-level entropy weights and dynamic thresholds, and its expression is: .

[0043] The auxiliary classifier module is used to enhance the distinguishability of known categories. The auxiliary classifier module consists of a fully connected layer and an output layer. The output probability is generated by a leaky-softmax function, ensuring the total probability is less than 1. Its expression is:

[0044] In the formula, It is a logits vector. It is the first One element; The optimized loss function of the auxiliary classifier module is expressed as: .

[0045] The self-supervised clustering module is used to mine structural information in the target domain, and the output of the feature extractor is... After regularization and normalization, the vector is fed into the prototype classifier module; the prototype classifier module obtains the representative vector. Constructing prototype vectors ; Store the features of all target domain samples in a memory bank Each slot Storage number The features corresponding to the nth target sample; in the nth Update the corresponding memory after each small batch iteration. ,in This represents the sample index within a small batch; This represents the target feature in a mini-batch; Represents the set of the source prototype and all target features; for The probability calculation expression for its characteristics being similar to those of the fault prototype or other target characteristics is as follows:

[0046] Furthermore, a clustering loss function is used to optimize the clustering of target samples and help identify edge samples, which is expressed as: .

[0047] The consensus domain adversarial module integrates the outputs of multiple modules through a consensus probability metric to achieve consistency of known categories and enhances domain alignment through adversarial learning. The probability of a sample belonging to a known class is calculated based on the estimated probability of the original classifier module belonging to the unknown class. The expression is as follows:

[0048] For the auxiliary classifier, the output is the high probability confidence score of the known class. The probability expression for each sample belonging to the known class is:

[0049] In the formula, The larger the value, the greater the probability that the sample belongs to a known category; conversely, the smaller the value, the greater the probability that the sample comes from an unknown category. In the self-supervised clustering module, the clustering confidence is defined as... ,in The similarity probability is calculated by taking the maximum probability between the target sample's features and the prototype vector to measure the closeness to the known category. Based on the above probabilities and confidence levels, the consensus probability index is calculated. Defined as:

[0050] Therefore, the consensus domain adversarial loss can be expressed as:

[0051] The training process of the model framework can be represented as follows:

[0052]

[0053]

[0054] In the formula, For the total loss, These are the weighting coefficients of the two-level entropy-sensing loss function. For the weight coefficients of the loss function of the auxiliary classifier, These are the weight coefficients of the self-supervised clustering loss function. These are the weight coefficients for the consensus domain adversarial loss function; These are the trainable parameters of the feature extractor. These are the trainable parameters of the original classifier module. These are the trainable parameters for the auxiliary classifier module.

[0055] The parameters of this framework include trainable parameters. Total number of training rounds Training batch .

[0056] S3. Collect vibration signals from the rotating machinery on-site to obtain a test dataset. Input the test dataset into the trained model framework and output the fault diagnosis results. Specifically, use an accelerometer to collect new vibration signals from the rotating machinery and select segments of the vibration signals as the test dataset.

[0057] To verify the feasibility and technical effectiveness of this invention, the following verifications were conducted: Dataset Acquisition (1) Southeast University Dataset (SEU) The bearing dataset used in this study was provided by the School of Mechanical Engineering, Southeast University. The dataset includes rolling bearing vibration signals under two typical operating conditions: Condition 1: 20Hz (1200 rpm) unloaded (0V / 0 Nm); Condition 2: 30Hz (1800 rpm) loaded (2 V / 7.32 Nm). The data covers five bearing health states: healthy, ball fault, inner ring fault, outer ring fault, and combined inner and outer ring fault. The experiment used three accelerometers to simultaneously acquire vibration signals in the x, y, and z directions at a sampling frequency of 5120 Hz. Each health state contains 1000 samples, including 800 training samples and 200 test samples. Each sample contains 1024×3 data points as a vibration time-series signal. The experimental setup includes: motor controller 1, motor 2, planetary gearbox 3, parallel gearbox 4, brake 5, and brake controller 6, all attached to the test bench. Figure 3 The middle mark.

[0058] (2) Self-collected dataset (SC)

[0059] To verify the generalizability of the method, this study independently collected gear vibration datasets (SC) on a gear testing bench. The core components of the system include: a magnetic powder brake 7, a parallel shaft gearbox 8, a bearing housing 9, a rotor disk 10, a torque sensor 11, and a three-phase drive motor 12, all attached to the test bench. Figure 4 The vibration signals were simultaneously acquired using accelerometers in three directions, encompassing five gear health states: healthy, pitted, worn, broken teeth, and missing teeth. Data acquisition was conducted at three speeds: 800 rpm, 1200 rpm, and 1600 rpm. Each health state under each condition contained 1000 samples, with each sample containing 256 × 3 data points, including 800 training samples and 200 test samples.

[0060] The labels corresponding to the health status of the two datasets are detailed in Table I.

[0061] Table I - Health Status and Labels of the Two Datasets

[0062] To evaluate the stability and effectiveness of the proposed framework in different Open Set Domain Adaptive (OSDA) scenarios, we designed several diagnostic tasks with different degrees of openness. The settings for the OSDA fault diagnosis tasks are shown in Table II.

[0063] Table II - Settings for OSDA Fault Diagnosis Tasks

[0064] This paper selects five comparative methods to verify the superiority of the proposed framework. The details of each method are as follows: (1) Domain Adversarial Neural Network (DANN) A DANN consists of a feature extractor, a classifier, and a domain discriminator. DANN aligns the source and target domain distributions through a minimax game, while the classifier ensures that samples from the source domain are correctly classified.

[0065] (2) OSBP

[0066] OSBP implements Domain Adversarial Classification (OSDA) using an adversarial method. The classifier distinguishes between known classes in the first K dimensions and uses a fixed threshold to determine whether it belongs to an unknown class based on the probability of the last dimension.

[0067] (3) UAN

[0068] The UAN consists of a classifier, an adversarial discriminator, and a non-adversarial discriminator. The adversarial discriminator matches the distribution of features between the source and target domains, while the non-adversarial discriminator quantifies the similarity between samples and the domains. This model combines domain similarity and uncertainty metrics to evaluate the transferability of samples.

[0069] (4) DUAL

[0070] DUAL consists of a domain discriminator, an auxiliary domain discriminator, and an extended classifier. The auxiliary discriminator assigns similarity weights to target samples to distinguish between known and unknown categories and mitigate negative transfer. The feature extractor and the extended classifier utilize adversarial learning to construct a discriminative hyperplane, enhancing the separation between known and unknown categories.

[0071] (5) MAOSDAN

[0072] MAOSDAN consists of a fault classifier and an auxiliary adversarial classifier. The framework utilizes an attention-aware module to distinguish between known and unknown samples; the auxiliary classifier reduces the negative transfer caused by forced alignment of unknown samples; and an entropy suppression module is employed to enhance classification determinism.

[0073] Evaluation indicators

[0074] This study uses the following evaluation metrics to comprehensively measure framework performance: (1) , representing the accuracy of known categories.

[0075] (2) , representing the accuracy of unknown samples.

[0076] (3) , representing the accuracy across all categories (including unknown categories).

[0077] (4) This value comprehensively evaluates the balance between known category classification and unknown category recognition. A higher value indicates a better balance in the model. The calculation formula is as follows:

[0078] Model and hyperparameter settings

[0079] The five convolutional layers have kernel sizes of 11, 5, 3, 3, and 3, corresponding to 64, 192, 384, 256, and 256 output channels, respectively. The max pooling layer has a kernel size of 3 and a stride of 2, while the adaptive average pooling layer has a kernel size of 6. The Bi-LSTM module contains two layers of size 256, and the multi-head attention module contains eight attention heads. The classifier contains three fully connected layers with output dimensions of 128, 64, and K+1, respectively. The auxiliary classifier also contains three fully connected layers, with the Leaky-softmax activation function applied to the last layer. Its output dimensions are 128, 64, and K, respectively.

[0080] Based on Task B1, the key hyperparameters of the proposed framework were determined. The model training parameters were configured as follows: batch size and epochs were set to 64 and 100, respectively; the initial learning rate was set to 0.001; and the model parameters were updated using stochastic gradient descent (SGD) optimization with a momentum of 0.9. The final optimal parameter combination was: , , , Specific parameter information is shown in Table III.

[0081] Table III - Parameter Settings

[0082] ablation experiment

[0083] To verify the effectiveness of the core modules of the proposed framework, ablation experiments were conducted on the SC dataset for tasks B1, B2, and B5. Six ablation models removed different modules, as follows: AB1: Sample level removed from the double-entropy sensing module; AB2: Domain level removed from the double-entropy sensing module; AB3: The entire double-entropy sensing module removed; AB4: Auxiliary classifier module removed; AB5: Self-supervised clustering module removed; AB6: Consensus domain adversarial module removed. All experimental parameters remained consistent. The comparison results of the six ablation models are shown in Table IV.

[0084] Table IV - Ablation Experiment Results

[0085] The diagnostic results of the SEU dataset are shown in Table V.

[0086] Table V - Diagnostic results of different methods on the SEU dataset

[0087] The diagnostic results of the SC dataset are shown in Table VI.

[0088] Table VI - Diagnostic Results of Different Methods on the SC Dataset

[0089] Table VI (continued) - Diagnostic results of different methods on the SC dataset

[0090] Applications under practical conditions: This dataset originates from an actual failure case of the reducer of the left coiler in the coiling area of ​​a cold rolling mill, collected from June to November 2023. Experts determined that the equipment triggered a vibration alarm in June 2023 due to wear on the inner raceway of the bearing. On October 31, 2023, disassembly and inspection revealed severe peeling of the bearing inner raceway and damage to the output tooth surface of the high-speed shaft. Figure 5 As shown. After replacing the high-speed shaft and bearing, the abnormal characteristics disappeared, and normal operation was restored. Based on data analysis, the samples were divided into three health states: healthy state (H), bearing single wear state (SW), and combined failure state of bearing wear and tooth surface damage (BWTS). H and SW are known classes, and BWTS is an unknown class. In this scenario, 1600 samples each of H and SW are used as training samples, and 400 samples each of H, SW, and BWTS are used as test samples. Each sample contains 1024×1 data points.

[0091] Based on the function of each module, we slightly adjusted the hyperparameters of the loss function, and... Upgraded to 2.5 An increase of 0.5 was made to improve overall performance. Experimental results are shown in Table VII.

[0092] Table VII - Comparison of Diagnostic Results Using Different Methods

[0093] Taking task B2 as an example, the characteristics of different methods are visualized, such as... Figure 6 As shown.

[0094] Through the above verification, compared with existing domain adaptive diagnostic methods, this invention fully considers the situation where unknown faults exist in the target domain, thereby enabling the alignment of known categories and the identification of unknown categories. It effectively solves the limitation of existing domain adaptive diagnostic methods in handling unknown faults, achieving open-set fault diagnosis for rotating machinery.

[0095] The above specific embodiments should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, any alternative improvements or modifications made to the embodiments of the present invention shall fall within the scope of protection of the present invention.

[0096] Any aspects of this invention not described in detail are well-known to those skilled in the art.

Claims

1. An open set fault diagnosis method based on a two-stage entropy-aware consensus domain adversarial framework, characterized in that, The method comprises the following steps: S1, obtaining a training data set of a rotating machine The training data set comprises a source domain data set and a target domain data set; S2, constructing a model framework The model framework comprises a feature extractor and loss functions of five modules; the feature extractor comprises a convolutional neural network, a bidirectional long short-term memory network and a multi-head attention mechanism arranged in sequence; the five modules comprise an original classifier module, a double-level entropy perception module, an auxiliary classifier module, a self-supervised clustering module and a consensus domain adversarial module; the training data set is input into the feature extractor, and then the parameters of the entire model framework are adjusted by using an optimization algorithm, so that the effects of the loss functions of the modules are optimal, thereby completing the training of the entire framework; S3, collecting a test data set of a vibration signal of the rotating machine in the field, inputting the test data set into the trained model framework, and outputting a fault diagnosis result.

2. The open set fault diagnosis method based on the two-stage entropy-aware consensus domain adversarial framework according to claim 1, characterized in that, In step S1, the source domain data set is denoted as: ； The target domain data set is denoted as: wherein, denotes the source domain, which contains labeled samples corresponding to classes, which are subject to a probability distribution ; denotes the target domain, which contains unlabeled samples subject to a probability distribution ; since the source domain and the target domain are collected under different working conditions, their probability distributions differ significantly, i.e. .

3. The open set fault diagnosis method based on the two-stage entropy-aware consensus domain adversarial framework according to claim 2, characterized in that, In step S2, the feature extractor establishes a mapping relationship between the vibration signal and the diagnostic result, and gives an input sample The running process expression of the feature extractor is: wherein denote trainable parameters; In step S2, the original classifier module outputs the output of the feature extractor to... Among the categories, Represents a known class, the first Each sample represents an unknown class; the original classifier module outputs a one-dimensional logits vector for each sample. The Softmax function converts a logarithmic vector into a logarithmic vector. A probability distribution of dimension 2, its expression is: For a classification task, the original classifier module is trained on source domain samples to learn discriminative features of known classes, and the original classification loss is represented as: wherein is a cross-entropy loss function, is the dimensional one-hot vector.

4. The open set fault diagnosis method based on the two-stage entropy-aware consensus domain adversarial framework according to claim 3, characterized in that, In step S2, the double-level entropy perception module comprises sample-level entropy perception and domain-level entropy perception, the sample-level entropy perception adaptively allocates weights to target samples, and the sample-level entropy perception first quantifies the classification difficulty of each target sample by using a standard entropy, and the expression is: Then an entropy-based weighting function is used The weights are assigned to the samples and introduced into the adversarial loss function; The domain level entropy-aware dynamic adjustment decision boundary, which level utilizes average entropy Quantifying the overall classification uncertainty in the target domain, the threshold for classification will be dynamically adjusted by batch, expressed as: wherein denotes the momentum in SGD, denotes the average entropy of the current batch of samples. The entropy weight of the sample level and the dynamic threshold are introduced into the adversarial loss, and the expression is: 。 5. The open set fault diagnosis method based on the two-stage entropy-aware consensus domain adversarial framework according to claim 4, characterized in that, In step S2, the auxiliary classifier module is used to enhance the distinguishability of known classes, the auxiliary classifier module is composed of a full connection layer and an output layer, the output probability is generated by a leaky-softmax function, so that the total probability is less than 1, and the expression is: wherein is a logits vector, is the th element; The optimization loss function of the auxiliary classifier module is represented as: 。 6. The open set fault diagnosis method based on the two-stage entropy-aware consensus domain adversarial framework according to claim 5, characterized in that, In step S2, the self-supervised clustering module is used to mine the structural information of the target domain, and the output of the feature extractor is input into the prototype classifier module after normalization After normalization, it is input into the prototype classifier module; Prototype classifier module Constructing prototype vectors ; store the features of all target domain samples in the memory , wherein each slot stores the features corresponding to the th target sample After the first memory is updated after each mini-batch iteration, where denotes the sample index in the mini-batch; denotes the target feature in the mini-batch; denotes the set of source prototypes and all target features; for the probability that its feature is similar to the failure prototype or other target features is calculated as: In addition, a clustering loss function is used to optimize the clustering of target samples and help identify edge samples, and the expression is: 。 7. The open set fault diagnosis method based on the two-stage entropy-aware consensus domain adversarial framework according to claim 6, characterized in that, In step S2, the consensus domain adversarial module integrates the outputs of multiple modules by consensus probability to realize the consistency of known classes, and enhances the field alignment by adversarial learning; The probability that a sample belongs to a known class is calculated according to the estimated probability that the original classifier module belongs to an unknown class, and the expression is: For the auxiliary classifier, the output is a high-probability confidence of a known class. The probability expression that each sample belongs to a known class is: wherein The larger the value of the more likely the sample belongs to the known class; conversely, the smaller the value, the more likely the sample comes from the unknown class. In the self-supervised clustering module, the clustering confidence is defined as where is the similarity probability, taking the maximum probability between the target sample feature and the prototype vector to measure the closeness to the known category; the consensus probability indicator is defined as: Therefore, the consensus domain adversarial loss can be represented as: The training process of the model framework can be represented as: wherein, is the total loss, is a two-stage entropy-aware loss function weight coefficient, is an auxiliary classifier loss function weight coefficient, is a self-supervised clustering loss function weight coefficient, is a consensus domain adversarial loss function weight coefficient; are trainable parameters of the feature extractor, are trainable parameters of the primary classifier module, are trainable parameters of the auxiliary classifier module.

Images