Interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN
An interactive zero-sample composite fault diagnosis method based on fuzzy semantics and FNN addresses the shortcomings of existing bearing composite fault diagnosis methods in handling nonlinearity and uncertainty, achieving accurate identification and domain adaptation capabilities, and is applicable to composite fault diagnosis in multiple unknown domains.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-12
AI Technical Summary
Existing zero-shot learning-based methods for diagnosing composite bearing faults struggle to handle the nonlinearity and uncertainty of composite faults, resulting in poor domain adaptability and impacting the reliability and stability of fault diagnosis.
An interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN is adopted. Through a feature extraction module, a semantic construction module, a semantic embedding module, and an interactive extension module, wavelet transform, convolutional neural network, fuzzy neural network and other technologies are used to realize feature extraction and semantic mapping. Combined with precise and fuzzy semantics, composite fault generation semantics are generated, and the fault category is inferred by cosine distance.
It achieves accurate identification even in the absence of complex faults, while also taking into account the adaptive capability of the model domain. It is applicable to the diagnosis of complex faults in multiple unknown domains and provides accurate and reliable auxiliary information and human-computer interaction support for fault diagnosis.
Smart Images
Figure CN122020386A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of complex fault diagnosis technology for rotating machinery, and specifically to an interactive zero-sample complex fault diagnosis method based on fuzzy semantics and FNN. Background Technology
[0002] The primary function of bearings is to support rotating mechanical parts and reduce the coefficient of friction of mechanical loads during transmission. They are critical components of rotating machinery. Therefore, continuous condition monitoring and fault diagnosis of bearings are essential to ensure the safe operation of equipment. Under the combined effects of long-term heavy loads, fatigue wear, and harsh operating conditions, composite bearing failures (i.e., the simultaneous occurrence of two or more mutually influential faults) are a common phenomenon.
[0003] Existing zero-shot learning (ZSL)-based composite bearing fault diagnosis schemes typically rely on precise semantic construction and deterministic feature-semantic mapping. This fails to adequately consider the fuzzy and uncertain characteristics of composite faults, resulting in poor domain adaptability. Consequently, when existing fault diagnosis methods operate under inconsistent training and testing conditions, they struggle to maintain excellent diagnostic performance, leading to misdiagnosis and severely impacting the reliability and stability of fault diagnosis. Therefore, there is an urgent need to develop a new diagnostic method capable of handling the nonlinearity and uncertainty of composite faults while ensuring domain adaptability, to overcome the limitations of existing zero-shot composite fault diagnosis methods.
[0004] 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
[0005] The purpose of this invention is to address the problem that existing zero-shot composite fault diagnosis methods are unable to utilize precise semantics and deterministic mapping to handle the nonlinearity and uncertainty of composite faults, resulting in poor domain adaptation capabilities. This invention provides an interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN. This method can achieve both accurate identification of unseen composite faults and model domain adaptation capabilities, and is applicable to composite fault diagnosis in multiple unknown domains.
[0006] The present invention achieves the above objectives by adopting the following technical solutions: An interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN includes the following steps: S1. Collect vibration signals of bearings in rotating machinery under different operating conditions for different types of faults, and segment the vibration signals to form a dataset. The dataset includes a visible single fault dataset and an unlabeled and invisible composite fault dataset. The single fault dataset is the training set, and the composite fault dataset is the test set. S2. Construct a fault diagnosis framework, which includes a feature extraction module, a semantic construction module, a semantic embedding module, an inference module, and an interactive extension module. The feature extraction module includes wavelet transform, convolutional neural network, and center loss, decoupling loss, and cross-entropy loss to extract features of single fault samples. The semantic construction module includes a precise single fault semantic construction module, a fuzzy single fault semantic construction module, a single fault semantic fusion module, and a composite fault semantic generation module to construct composite fault generation semantics as auxiliary information for composite fault diagnosis. The semantic embedding module uses a fuzzy neural network to elastically map the features of single fault samples to the semantic space to generate predicted semantics. The inference module is used to calculate the cosine distance between the predicted semantics and the fused semantics, and select the fault category with the closest cosine distance as the prediction result. The interactive extension module generates corresponding fault text descriptions by calling DeepSeek to generate composite fault generation semantics and / or predicted semantics. S3. The fault diagnosis framework is trained. In the feature extraction module, the single fault dataset is used as input, wavelet transform is used to separate signals of different frequency bands, convolutional neural networks are used to capture local structural features, center loss and decoupling loss are used to optimize the fault feature distribution, effectively extract single fault sample features, and cross-entropy loss is used to train the feature extraction module. In the semantic construction module, the single fault dataset is used as input, the statistical indicators of each sample are calculated, the statistical indicators are input into the precise single fault semantic construction module to obtain the precise single fault semantics, some statistical indicators are input into the fuzzy single fault semantic construction module to obtain the fuzzy single fault semantics, the single fault semantic fusion module concatenates the precise single fault semantics and the fuzzy single fault semantics to obtain the fused single fault semantics, and the composite fault semantic generation module linearly superimposes the fused single fault semantics to obtain the generated composite fault semantics. In the semantic embedding module, the single fault sample features are used as input, and a fuzzy neural network is used to realize the nonlinear mapping from features to semantics to obtain the final predicted semantics. The semantic embedding module is trained by semantic reconstruction loss and rule activation regularization damage. S4. Input the composite fault dataset into the trained fault diagnosis framework and output the fault diagnosis results.
[0007] Preferably, in step S1, an accelerometer is used to collect vibration signals of bearings in rotating machinery under four different operating conditions for different types of faults; The single-fault dataset is represented as follows:
[0008] In the formula, This represents a visible single-fault dataset. For sample space The first in One visible single-fault sample, Represents tag space Category tags, Semantic space The corresponding semantic attributes in The representative sample shows the total number of single faults. The composite fault dataset is represented as follows:
[0009] In the formula, This represents an unlabeled and invisible composite fault dataset. Representing the sample space The first in A composite fault sample, The total number of composite faults that are not visible in the representative sample.
[0010] Preferably, in step S2, a single fault sample is given. The wavelet image is obtained after wavelet transform. The process is represented as:
[0011] In the formula, variables and They represent the scaling factor and the translation factor, respectively. These are wavelet basis functions.
[0012] Preferably, in step S2, the feature extraction module uses the trainable parameters of the convolutional neural network. The final feature representation is obtained. The training process can be represented as:
[0013] For classification tasks, cross-entropy loss is used to learn discriminative features for a known single fault category, as shown below:
[0014] In the formula, It is a typical Softmax classifier. yes A one-hot vector of dimension; The center loss enhances the compactness of intra-class features by narrowing the distance between features and their corresponding class centers. Its expression is as follows:
[0015] In the formula, Indicates the first The category to which each sample belongs The class center vector, It is the loss weighting coefficient; The decoupling loss minimizes intra-class distance and maximizes inter-class distance through boundary-based constraints, and its expression is:
[0016] In the formula, the first term is the intra-class distance, and the second term is the inter-class distance. Representation and Category Different categories The class center vector, It is the boundary value of the inter-class distance. These are the weighting coefficients for the inter-class loss; The total loss of the feature extraction module is expressed as:
[0017] In the formula, The weighting coefficients represent the decoupling loss.
[0018] Preferably, in step S2, in the precise single-fault semantic construction module, given a single-fault dataset... As input, 29 statistical indicators for each sample are first calculated, including 15 time-domain indicators and 14 frequency-domain indicators. The statistical indicators for each fault category are as follows:
[0019] In the formula, Indicates the first The average value of the corresponding indicator for all samples in the class. This represents a 29-dimensional statistical indicator vector for this category; To accelerate model convergence, the same statistical indicators for all categories are normalized, and the expression is as follows:
[0020] The normalized single-fault statistical index vector is obtained by the following formula:
[0021] Then, through a feature mapping network consisting of two fully connected layers... Mapping the statistical index vector of a single fault to fault characteristics , and The mapping relationship between them is trained by the following objective function:
[0022] In the formula, Equal to the corresponding category statistical indicator vector , therefore, No. Precise semantic representation of a single fault category It can be obtained from the following formula:
[0023] In the formula, and These represent the weights and biases of the FC3 layer, respectively. This represents the ReLU activation function.
[0024] Preferably, in step S2, in the fuzzy single-fault semantic construction module, given a single-fault dataset... As input, four time-domain indicators and three frequency-domain indicators were selected from 29 statistical indicators to form the first... The statistical indicator vectors for each single fault category are shown below:
[0025] Then, based on the extraction and visible fault categories Generate text description The expression is:
[0026] In the formula, Based on fault category Functions for generating text descriptions of indicators; Subsequently, each text description is converted into a 256-dimensional fuzzy semantic vector using a pre-trained Word2Vec model. As shown below:
[0027] If a word in the text description is not in the dictionary, it will be replaced with the word that is semantically closest to it, thus obtaining a fuzzy semantic vector.
[0028] Preferably, in step S2, in the single-fault semantic fusion module, the final single-fault fused semantics is obtained by concatenating the precise single-fault semantics and the fuzzy single-fault semantics. Single-fault fusion semantics Represented as:
[0029] In the formula, This represents the concatenation operation; the single-fault semantic set for all categories is: .
[0030] Preferably, in step S2, in the composite fault semantic generation module, based on the strong correlation between single faults and composite faults, the composite fault semantic generation is generated by the linear superposition of single fault fusion semantics, and its expression is:
[0031] In the formula, .
[0032] Preferably, in step S2, in the semantic embedding module, given the features of a single fault sample... Through the trainable parameters of the fuzzy neural network This yields the final predicted semantics. The training process of this module can be represented by the following mapping function:
[0033] The semantic reconstruction loss The expression for the ability of a fuzzy neural network to learn the mapping from features to semantics is:
[0034] The regularization loss of the rule activation It is used to promote rule sparsity and prevent rule redundancy, and its expression is:
[0035] In the formula, where Indicates the number of fuzzy rules. Indicates the first The first sample The activation level of the rule; The total loss expression for the semantic embedding module is:
[0036] In the formula, yes Hyperparameters.
[0037] Preferably, in step S2, the inference module is based on Generative semantics for each composite fault category The cosine distance between the test composite fault samples is used to infer their category. The cosine distance calculation expression is:
[0038] The class with the closest cosine distance is selected as the prediction result, and its expression is:
[0039] In the formula, This represents the prediction results for composite fault diagnosis.
[0040] Preferably, in step S3, the training process of the fault diagnosis framework is represented as follows:
[0041] The feature extraction module proposes cross-entropy loss, center loss, and decoupling loss, and the total loss can be expressed as:
[0042] The semantic embedding module proposes semantic reconstruction loss and rule activation regularization loss, and the total loss can be expressed as:
[0043] The parameters of the fault diagnosis framework include trainable parameters. Total number of training rounds and training batch Among them, trainable parameters Represented as: .
[0044] Preferably, in step S4, the testing process of the fault diagnosis framework is represented as follows:
[0045] In the interactive extension module, in order to verify the accuracy of the text description corresponding to the predicted semantics and the text description corresponding to the compound fault generation semantics, the text description corresponding to the real semantics is loaded at the same time. The text descriptions of the three semantics should be consistent with each other.
[0046] The present invention, employing the above-described structure, can bring the following beneficial effects: (1) By using wavelet transform to achieve reasonable separation of signals in different frequency bands, using CNN to capture local structural features, and using center loss and decoupling loss to optimize the distribution of fault features, the problem of module reusability is well solved. (2) By constructing precise semantic-fuzzy semantic fusion semantics and generating composite fault generation semantics, accurate and reliable auxiliary information is provided for composite fault diagnosis; (3) In order to cope with the nonlinear characteristics of complex semantics and solve the rigidity of deterministic mapping, this application adopts fuzzy neural network (FNN) as the mapping mechanism to realize nonlinear mapping from features to semantics; (4) The cosine distance between the predicted semantics and the generated semantics of the composite fault is calculated and the nearest category is obtained as the fault diagnosis prediction result; at the same time, an interactive extension module based on the concept of "human-computer interaction" is given to give the fault diagnosis prediction result, and the semantics to text description conversion can be realized. The text description can help engineers make appropriate judgments. This application implements zero-sample composite fault diagnosis based on an interactive zero-sample fault diagnosis framework using fuzzy semantics and fuzzy neural networks. It achieves accurate identification of unseen composite faults while also considering the ability to adapt the model domain, making it suitable for composite fault diagnosis in multiple unknown domains. Attached Figure Description
[0047] Figure 1 This is a schematic diagram of the framework of the composite fault diagnosis method of the present invention; Figure 2 This is the experimental platform for the TYUST dataset of this invention; Figure 3 This is the experimental platform for the PU dataset of this invention; In the diagram, 1 is the magnetic powder brake, 2 is the parallel shaft gearbox, 3 is the bearing housing, 4 is the rotor disk, 5 is the torque sensor, 6 is the three-phase drive motor, 7 is the drive motor, 8 is the torque measuring shaft, 9 is the bearing test module, 10 is the flywheel, and 11 is the load motor. Detailed Implementation
[0048] 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.
[0049] 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.
[0050] 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.
[0051] In this invention, unless otherwise explicitly specified and limited, the terms "provided with," "set up," and "connected" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0052] like Figure 1 As shown, the interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN includes the following steps: S1. Collect vibration signals of bearings in rotating machinery under different operating conditions for different types of faults, and segment the vibration signals to form a dataset. The dataset includes a visible single fault dataset and an unlabeled and invisible composite fault dataset. The single fault dataset is the training set, and the composite fault dataset is the test set. Specifically, accelerometers can be used to collect vibration signals of bearings in rotating machinery under four different operating conditions for different types of faults; In addition, the single-fault dataset is represented as:
[0053] In the formula, This represents a visible single-fault dataset. For sample space The first in One visible single-fault sample, Represents tag space Category tags, Semantic space The corresponding semantic attributes in The representative sample shows the total number of single faults. The composite fault dataset is represented as follows:
[0054] In the formula, This represents an unlabeled and invisible composite fault dataset. Representing the sample space The first in A composite fault sample, The total number of composite faults that are not visible in the representative sample; S2. Construct a fault diagnosis framework, which includes a feature extraction module, a semantic construction module, a semantic embedding module, a reasoning module, and an interactive extension module; The feature extraction module includes wavelet transform, convolutional neural network, and center loss, decoupling loss, and cross-entropy loss, used to extract features from single fault samples. Specifically, wavelet transform can achieve reasonable separation of signals from different frequency bands, given a single fault sample. The wavelet image is obtained after wavelet transform. The process is represented as:
[0055] In the formula, variables and They represent the scaling factor and the translation factor, respectively. These are wavelet basis functions; The feature extraction module captures local structural features using a convolutional neural network (CNN) and its trainable parameters. The final feature representation is obtained. The training process can be represented as:
[0056] For classification tasks, cross-entropy loss is used to learn discriminative features for a known single fault category, as shown below:
[0057] In the formula, It is a typical Softmax classifier. yes A one-hot vector of dimension; To address the module reusability issue, centrality loss and decoupling loss are introduced.
[0058] The center loss enhances the compactness of intra-class features by narrowing the distance between features and their corresponding class centers. Its expression is:
[0059] In the formula, Indicates the first The category to which each sample belongs The class center vector, It is the loss weighting coefficient; The decoupling loss minimizes intra-class distance and maximizes inter-class distance through boundary-based constraints, and its expression is:
[0060] In the formula, the first term is the intra-class distance, and the second term is the inter-class distance. Representation and Category Different categories The class center vector, It is the boundary value of the inter-class distance. These are the weighting coefficients for the inter-class loss; The total loss of the feature extraction module is expressed as:
[0061] In the formula, The weighting coefficients represent the decoupling loss; The semantic construction module includes a precise single-fault semantic construction module, a fuzzy single-fault semantic construction module, a single-fault semantic fusion module, and a composite fault semantic generation module, used to construct composite fault generation semantics as auxiliary information for composite fault diagnosis; specifically, In the precise single-fault semantic building module, given a single-fault dataset... As input, 29 statistical indicators for each sample are first calculated, including 15 time-domain indicators and 14 frequency-domain indicators. The statistical indicators for each fault category are as follows:
[0062] In the formula, Indicates the first The average value of the corresponding indicator for all samples in the class. This represents a 29-dimensional statistical indicator vector for this category; To accelerate model convergence, the same statistical indicators for all categories are normalized, and the expression is as follows:
[0063] The normalized single-fault statistical index vector is obtained by the following formula:
[0064] Then, through a feature mapping network consisting of two fully connected layers... Mapping the statistical index vector of a single fault to fault characteristics , and The mapping relationship between them is trained by the following objective function:
[0065] In the formula, Equal to the corresponding category statistical indicator vector , therefore, No. Precise semantic representation of a single fault category It can be obtained from the following formula:
[0066] In the formula, and These represent the weights and biases of the FC3 layer, respectively. This represents the ReLU activation function.
[0067] In the fuzzy single-fault semantic construction module, given a single-fault dataset... As input, four time-domain indicators (root mean square) were selected from 29 statistical indicators. ), peak ( ), kurtosis ( ) and peak factor ( )) and 3 frequency domain indicators (dominant frequency ( ), spectral centroid ( ) and envelope peak frequency ( )) constitutes the first The statistical indicator vectors for each single fault category are shown below:
[0068] Then, based on the extraction and visible fault categories Generate text description The expression is:
[0069] In the formula, Based on fault category Functions for generating text descriptions of indicators; Subsequently, each text description is converted into a 256-dimensional fuzzy semantic vector using a pre-trained Word2Vec model. As shown below:
[0070] If a word in the text description is not in the dictionary, it will be replaced with the word that is semantically closest to it, thus obtaining a fuzzy semantic vector. For example, in the description of bearing failure, if the word "fatigue" is not in the dictionary, it will be replaced with "wear" from the dictionary because "wear" and "fatigue" are semantically similar, thus constructing a semantic vector.
[0071] In the single-fault semantic fusion module, the final single-fault fused semantics is obtained by concatenating the precise single-fault semantics and the fuzzy single-fault semantics. Single-fault fusion semantics Represented as:
[0072] In the formula, This represents the concatenation operation; the single-fault semantic set for all categories is: .
[0073] In the composite fault semantic generation module, based on the strong correlation between single faults and composite faults, the semantics of composite fault generation are generated by the linear superposition of the fused semantics of single faults, and its expression is:
[0074] In the formula, .
[0075] To address the nonlinear characteristics of hardness composite semantics and overcome the rigidity of deterministic mapping, the semantic embedding module employs a fuzzy neural network to elastically map single-fault sample features to the semantic space to generate predicted semantics; specifically, given single-fault sample features... Through the trainable parameters of the fuzzy neural network This yields the final predicted semantics. The training process of this module can be represented by the following mapping function:
[0076] The loss function for the semantic embedding module consists of two parts: semantic reconstruction loss and rule activation regularization loss.
[0077] Semantic Reconstruction Loss The expression for the ability of a fuzzy neural network to learn the mapping from features to semantics is:
[0078] The regularization loss of the rule activation It is used to promote rule sparsity and prevent rule redundancy, and its expression is:
[0079] In the formula, where Indicates the number of fuzzy rules. Indicates the first The first sample The activation level of the rule; The total loss expression for the semantic embedding module is:
[0080] In the formula, yes Hyperparameters.
[0081] The inference module is used to calculate the cosine distance between the predicted semantics and the fused semantics, and selects the fault category with the closest cosine distance as the prediction result; specifically, the inference module is based on Generative semantics for each composite fault category The cosine distance between the test composite fault samples is used to infer their category. The cosine distance calculation expression is:
[0082] The class with the closest cosine distance is selected as the prediction result, and its expression is:
[0083] In the formula, This represents the prediction results for composite fault diagnosis.
[0084] The interactive extension module, based on the concept of "human-computer interaction," proposes a function that supports semantic-to-text conversion. This module loads the semantics of composite fault samples and then calls DeepSeek to convert them into corresponding text descriptions. These descriptions can serve as engineering hints, helping engineers make appropriate judgments. Specifically, in obtaining the generated semantics... Or predict semantics Next, DeepSeek is called to generate the corresponding fault text description. To verify the accuracy of the two text descriptions mentioned above, the text description corresponding to the actual semantics is also loaded. In theory, the textual descriptions of these three semantics should be consistent with each other.
[0085] S3. Training the fault diagnosis framework: In the feature extraction module, the single fault dataset is used as input. Wavelet transform is used to separate signals of different frequency bands. Convolutional neural networks are used to capture local structural features. Center loss and decoupling loss are used to optimize the fault feature distribution, effectively extracting single fault sample features. Cross-entropy loss is used to train the feature extraction module. In the semantic construction module, the single fault dataset is used as input. Statistical indicators for each sample are calculated. The statistical indicators are input into the precise single fault semantic construction module to obtain precise single fault semantics. Some statistical indicators are input into the fuzzy single fault semantic construction module to obtain fuzzy single fault semantics. The single fault semantic fusion module concatenates the precise single fault semantics and the fuzzy single fault semantics to obtain fused single fault semantics. The composite fault semantic generation module linearly superimposes the fused single fault semantics to obtain composite fault generated semantics. In the semantic embedding module, the single fault sample features are used as input. A fuzzy neural network is used to achieve a nonlinear mapping from features to semantics to obtain the final predicted semantics. The semantic embedding module is trained using semantic reconstruction loss and rule activation regularization corruption. The training process of the fault diagnosis framework is represented as follows:
[0086] The feature extraction module proposes cross-entropy loss, center loss, and decoupling loss, and the total loss can be expressed as:
[0087] The semantic embedding module proposes semantic reconstruction loss and rule activation regularization loss, and the total loss can be expressed as:
[0088] The parameters of the fault diagnosis framework include trainable parameters. Total number of training rounds and training batch Among them, trainable parameters Represented as: .
[0089] S4. Input the composite fault dataset into the trained fault diagnosis framework and output the fault diagnosis results. The testing process for the fault diagnosis framework is represented as follows: .
[0090] To verify the feasibility and superiority of this method, the following verifications were conducted: I. Setting Evaluation Indicators: This study uses the following evaluation metrics to comprehensively measure framework performance: (1) To measure the overall performance of the model; (2) To measure the proportion of correctly classified items; (3) This measures the ability to detect positive samples. (4) ,balance and The comprehensive indicators.
[0091] II. Validation 1: Validation using the Taiyuan University of Science and Technology dataset (TYUST).
[0092] 2.1 This dataset was obtained through experiments on 6205 deep groove ball bearings with different damage types using the parallel shaft gearbox failure mechanism simulation test bench (model PT500PLUS) at Taiyuan University of Science and Technology. For example... Figure 2 As shown, the core components of the experimental platform include a magnetic powder brake (1), a parallel shaft gearbox (2), a bearing housing (3), a rotor disk (4), a torque sensor (5), and a three-phase drive motor (6). The accelerometer (VALENIAN-676090) is positioned along the bearing cage... The device was mounted axially with a fixed load of 10 Nm and a sampling frequency of 10 kHz. The original vibration signals were collected under four operating conditions: 800 rpm, 1200 rpm, 1600 rpm, and 2000 rpm. The dataset includes eight health states: healthy (N), inner race fault (IF), outer race fault (OF), ball fault (BF), combined inner and outer race fault (IF&OF), combined inner and ball fault (IF&BF), combined outer race and ball fault (OF&BF), and combined inner, outer, and ball fault (IF&OF&BF).
[0093] To evaluate the diagnostic performance of the fault diagnosis framework in this application under different operating conditions, several diagnostic tasks with varying difficulty and operating conditions were designed using the TYUST dataset from Taiyuan University of Science and Technology. The settings for the fault diagnosis tasks are shown in Table 1.
[0094]
[0095] 2.2 Model Setup
[0096] First, for single fault signals Perform wavelet transform to obtain a 64×64 wavelet image. The input to the convolutional layer is used as the input. Then, two convolutional layers (C1 and C2) and two pooling layers (P1 and P2) are used to extract fault features, and the output of the FC2 layer is selected as the fault feature. The precise-fuzzy semantic building block contains a feature mapping network consisting of two fully connected layers. It outputs precise semantics at the FC3 layer. The semantic embedding module uses a four-layer fuzzy neural network. Its input is the 2048-dimensional fault features extracted by the feature extraction module. The FNN performs feature semantic mapping and outputs 512-dimensional predicted semantics at the normalization layer. See Table 2 for detailed architecture parameters.
[0097]
[0098] 2.3 Ablation Experiment of Feature Extraction Module
[0099] To verify the effectiveness of the two loss terms in the feature extraction module, we constructed three ablation models for tasks A and B, defined as follows: AB1, removing the center loss; AB2, removing the decoupling loss; AB3, removing both the center loss and the decoupling loss. The comparison results are shown in Table 3.
[0100]
[0101] 2.4 Comparison Experiment of Semantic Construction Modules
[0102] To verify the effectiveness of fault semantic construction, this application compares three different semantic construction methods: the exact single-fault semantic construction method (OPS), the fuzzy single-fault semantic construction method (OFS), and the semantic construction method using a convolutional autoencoder (CAE). To verify the effectiveness of the two statistical semantic representations, this paper develops two methods combining the above approaches: PS-CAE and CAE-FS. To verify the effectiveness of fault semantic fusion, this paper constructs a weighted dot product (WDPIE) of exact and fuzzy semantics. The comparative experimental results are shown in Table 4.
[0103]
[0104] 2.5 Comparative Experiment of Semantic Embedding Modules
[0105] To verify the mapping effect of FNN, we selected two different semantic embedding models, namely fully connected networks (FCN) and variational autoencoders (VAE), for comparison to explore the effectiveness of FNN. The comparative experimental results of different mapping networks are shown in Table 5.
[0106]
[0107] 2.6 Comparative Experiment of Reasoning Module
[0108] To analyze the impact of different distance metrics on inference performance, this paper selects Manhattan distance and Euclidean distance for comparison. Experiments were conducted on tasks A1, B1, and C1, and the comparison results are shown in Table 6.
[0109]
[0110] 2.7 Hyperparameter Settings
[0111] The key hyperparameters of the proposed model are determined based on tasks A and B. The model training parameters are configured as follows: batch size and maximum number of epochs are set to 32 and 50, respectively; the initial learning rate is set to 0.001; the model parameters are updated using the Adam algorithm; the number of membership functions in the semantic embedding module is set to 11, and the number of fuzzy rules is set to 20. The hyperparameters for the loss term are selected as follows: = 0.3, = 0.1. See Table 7 for detailed parameters.
[0112]
[0113] 2.8 Comparison Methods
[0114] To verify the superiority of the proposed method, it was compared with four other methods—CADAVAE, LDSIFD, ZFSL, and SLDF—under Task A and Task B. The specific details of each method are as follows.
[0115] (1) CADAVAE. This method consists of four parts: CNN is used to extract fault features; semantics are obtained by manually annotating attribute vectors; the embedding module maps features and semantics to a shared latent embedding space through two variational autoencoders (VAEs); finally, classification is achieved using Softmax.
[0116] (2) LDSIFD, which consists of three modules: obtaining features through locally connected restricted Boltzmann machines; constructing a label description space (LDS), calculating the spatial representation vector of a single fault and averaging these vectors to obtain composite semantics; and establishing a bidirectional projection between the feature space and the label description space by using a linear supervised autoencoder and classifying the data by calculating the Euclidean distance to obtain the mapping.
[0117] (3) ZFSLM. This method consists of three modules: a fault semantic construction module using a computer-aided autoencoder (CAE); a feature extraction module using a convolutional neural network (CNN); a fault semantic embedding module using an autoencoder; and an inference module using Manhattan distance.
[0118] (4) SLDF. The method consists of three modules: a semantic construction module that integrates manually constructed semantics and CAE semantics; a feature extraction module that uses CNN for feature extraction; and a fault semantic embedding module that uses a two-layer fully connected network. The diagnostic results of the TYUST dataset are shown in Table 8.
[0119]
[0120] III. Validation 2: Using the University of Paderborn (PU) dataset
[0121] 3.1 This dataset was provided by the Department of Mechanical Engineering, University of Paderborn, and collected bearing vibration signals under four different operating conditions: Condition 0: 1500 rpm, load 0.7 Nm, radial force 400 N; Condition 1: 1500 rpm, load 0.7 Nm, radial force 1000 N; Condition 2: 1500 rpm, load 0.1 Nm, radial force 1000 N; Condition 3: 900 rpm, load 0.7 Nm, radial force 1000 N. The data mainly covers seven bearing health states: healthy (N), inner ring single pitting (IS1), outer ring multiple pitting (OM1), outer ring single indentation (OS2), outer ring multiple indentation (OM2), inner and outer ring single pitting (IOS1), and inner and outer ring multiple indentation (IOM2). The sampling frequency of the vibration signals was 64 kHz. The experimental setup is as follows: Figure 3 As shown, the system includes a drive motor (7), a torque measurement shaft (8), a bearing testing module (9), a flywheel (10), and a load motor (11). To evaluate the diagnostic performance of the proposed framework under different operating conditions, we designed several diagnostic tasks with different difficulties and operating conditions on the Paderborn dataset (PU). The settings for the fault diagnosis tasks are shown in Table 9.
[0122]
[0123] 3.2 Comparison of experimental results
[0124] To verify the superiority of the proposed method, a comparative experiment was conducted on the PU dataset using the same comparison method as in Case 1.
[0125] The diagnostic results of the PU dataset are shown in Table 10.
[0126]
[0127] 3.3 Extension Module
[0128] The extended module was validated in Task A, where the semantic text descriptions were generated under constraints in the time and frequency domains. A sample was randomly selected from each of the two composite fault categories, and the generated text descriptions based on the true semantics, generated semantics, and predicted semantics are shown in Table 11.
[0129]
[0130] The above verification demonstrates that the proposed composite fault diagnosis method solves the problem that existing zero-sample composite fault diagnosis methods struggle to utilize precise semantics and deterministic mappings to handle the nonlinearity and uncertainty of composite faults, resulting in poor domain adaptability. This method can achieve both accurate identification of unseen composite faults and model domain adaptability, making it suitable for composite fault diagnosis in multiple unknown domains.
[0131] 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.
[0132] Any aspects of this invention not described in detail are well-known to those skilled in the art.
Claims
1. An interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN, characterized in that, Includes the following steps: S1. Collect vibration signals of bearings in rotating machinery under different operating conditions for different types of faults, and segment the vibration signals to form a dataset. The dataset includes a visible single fault dataset and an unlabeled and invisible composite fault dataset. The single fault dataset is the training set, and the composite fault dataset is the test set. S2. Construct a fault diagnosis framework, which includes a feature extraction module, a semantic construction module, a semantic embedding module, an inference module, and an interactive extension module. The feature extraction module includes wavelet transform, convolutional neural network, and center loss, decoupling loss, and cross-entropy loss to extract features of single fault samples. The semantic construction module includes a precise single fault semantic construction module, a fuzzy single fault semantic construction module, a single fault semantic fusion module, and a composite fault semantic generation module to construct composite fault generated semantics as auxiliary information for composite fault diagnosis. The semantic embedding module uses a fuzzy neural network to elastically map the features of single fault samples to the semantic space to generate predicted semantics. The inference module is used to calculate the cosine distance between the predicted semantics and the fused semantics, and select the fault category with the closest cosine distance as the prediction result. The interactive extension module generates corresponding fault text descriptions by calling DeepSeek to generate semantic and / or predictive semantics of complex faults. S3. Train the fault diagnosis framework; In the feature extraction module, the single fault dataset is used as input, wavelet transform is used to separate signals of different frequency bands, convolutional neural network is used to capture local structural features, center loss and decoupling loss are used to optimize the fault feature distribution, effectively extract single fault sample features, and cross-entropy loss is used to train the feature extraction module. In the semantic construction module, the single-fault dataset is used as input to calculate the statistical indicators for each sample. The statistical indicators are then input into the precise single-fault semantic construction module to obtain the precise single-fault semantics. Some statistical indicators are input into the fuzzy single-fault semantic construction module to obtain the fuzzy single-fault semantics. The single-fault semantic fusion module concatenates the precise and fuzzy single-fault semantics to obtain the fused single-fault semantics. The composite fault semantic generation module linearly superimposes the fused single-fault semantics to obtain the generated composite fault semantics. In the semantic embedding module, the features of the single-fault samples are used as input. A fuzzy neural network is used to achieve a non-linear mapping from features to semantics to obtain the final predicted semantics. The semantic embedding module is trained using semantic reconstruction loss and rule activation regularization corruption. S4. Input the composite fault dataset into the trained fault diagnosis framework and output the fault diagnosis results.
2. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 1, characterized in that, In step S1, an accelerometer is used to collect vibration signals of bearings in rotating machinery under four different operating conditions for different types of faults; The single-fault dataset is represented as follows: In the formula, This represents a visible single-fault dataset. For sample space The first in One visible single-fault sample, Represents tag space Category tags, Semantic space The corresponding semantic attributes in The representative sample shows the total number of single faults. The composite fault dataset is represented as follows: In the formula, This represents an unlabeled and invisible composite fault dataset. Representing the sample space The first in A composite fault sample, The total number of composite faults that are not visible in the representative sample.
3. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 2, characterized in that, In step S2, a single fault sample is given. The wavelet image is obtained after wavelet transform. The process is represented as: In the formula, variables and They represent the scaling factor and the translation factor, respectively. These are wavelet basis functions; The feature extraction module uses the trainable parameters of the convolutional neural network. The final feature representation is obtained. The training process can be represented as: For classification tasks, cross-entropy loss is used to learn discriminative features for a known single fault category, as shown below: In the formula, It is a typical Softmax classifier. yes A one-hot vector of dimension; The center loss enhances the compactness of intra-class features by narrowing the distance between features and their corresponding class centers. Its expression is as follows: In the formula, Indicates the first The category to which each sample belongs The class center vector, It is the loss weighting coefficient; The decoupling loss minimizes intra-class distance and maximizes inter-class distance through boundary-based constraints, and its expression is: In the formula, the first term is the intra-class distance, and the second term is the inter-class distance. Representation and Category Different categories The class center vector, It is the boundary value of the inter-class distance. These are the weighting coefficients for the inter-class loss; The total loss of the feature extraction module is expressed as: In the formula, The weighting coefficients represent the decoupling loss.
4. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 3, characterized in that, In step S2, within the precise single-fault semantic construction module, given a single-fault dataset... As input, 29 statistical indicators for each sample are first calculated, including 15 time-domain indicators and 14 frequency-domain indicators. The statistical indicators for each fault category are as follows: In the formula, Indicates the first The average value of the corresponding indicator for all samples in the class. This represents a 29-dimensional statistical indicator vector for this category; To accelerate model convergence, the same statistical indicators for all categories are normalized, and the expression is as follows: The normalized single-fault statistical index vector is obtained by the following formula: Then, through a feature mapping network consisting of two fully connected layers... Mapping the statistical index vector of a single fault to fault characteristics , and The mapping relationship between them is trained by the following objective function: In the formula, Equal to the corresponding category statistical indicator vector , therefore, No. Precise semantic representation of a single fault category It can be obtained from the following formula: In the formula, and These represent the weights and biases of the FC3 layer, respectively. This represents the ReLU activation function.
5. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 4, characterized in that, In step S2, within the fuzzy single-fault semantic construction module, given a single-fault dataset... As input, four time-domain indicators and three frequency-domain indicators were selected from 29 statistical indicators to form the first... The statistical indicator vectors for each single fault category are shown below: Then, based on the extraction and visible fault categories Generate text description The expression is: In the formula, Based on fault category Functions for generating text descriptions of indicators; Subsequently, each text description is converted into a 256-dimensional fuzzy semantic vector using a pre-trained Word2Vec model. As shown below: If a word in the text description is not in the dictionary, it will be replaced with the word that is semantically closest to it, thus obtaining a fuzzy semantic vector.
6. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 5, characterized in that, In step S2, within the single-fault semantic fusion module, the final single-fault fused semantics is obtained by concatenating the precise single-fault semantics and the fuzzy single-fault semantics. Single-fault fusion semantics Represented as: In the formula, This represents the concatenation operation; the single-fault semantic set for all categories is: .
7. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 6, characterized in that, In step S2, within the composite fault semantic generation module, based on the strong correlation between single faults and composite faults, the composite fault semantics are generated by the linear superposition of single fault fusion semantics, and its expression is: In the formula, .
8. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 7, characterized in that, In step S2, in the semantic embedding module, given the features of a single fault sample... Through the trainable parameters of the fuzzy neural network This yields the final predicted semantics. The training process of this module can be represented by the following mapping function: The semantic reconstruction loss The expression for the ability of a fuzzy neural network to learn the mapping from features to semantics is: The regularization loss of the rule activation It is used to promote rule sparsity and prevent rule redundancy, and its expression is: In the formula, where Indicates the number of fuzzy rules. Indicates the first The first sample The activation level of the rule; The total loss expression for the semantic embedding module is: In the formula, yes Hyperparameters.
9. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 8, characterized in that, In step S2, the inference module is based on Generative semantics for each composite fault category The cosine distance between the test composite fault samples is used to infer their category. The cosine distance calculation expression is: The class with the closest cosine distance is selected as the prediction result, and its expression is: In the formula, This represents the prediction results for composite fault diagnosis.
10. The interactive zero-shot composite fault diagnosis method based on fuzzy semantics and FNN according to claim 9, characterized in that, In step S3, the training process of the fault diagnosis framework is represented as follows: The feature extraction module proposes cross-entropy loss, center loss, and decoupling loss, and the total loss can be expressed as: The semantic embedding module proposes semantic reconstruction loss and rule activation regularization loss, and the total loss can be expressed as: The parameters of the fault diagnosis framework include trainable parameters. Total number of training rounds and training batch Among them, trainable parameters Represented as: In step S4, the testing process for the fault diagnosis framework is represented as follows: In the interactive extension module, in order to verify the accuracy of the text description corresponding to the predicted semantics and the text description corresponding to the compound fault generation semantics, the text description corresponding to the real semantics is loaded at the same time. The text descriptions of the three semantics should be consistent with each other.