Zero-shot prototypical expansion method and system for alleviating hubness problem and composite fault diagnosis

Abstract

The application relates to a zero-sample prototype expansion method and system for relieving a hubness problem and compound fault diagnosis, and belongs to the field of mechanical fault diagnosis. State signal samples of bearings of different types and under different working conditions are collected through a bearing fault test bed, corresponding fault attribute prototypes are designed, prototype expansion kernel functions are used to map the prototypes to a high-dimensional space, the distance between the designed attribute prototypes is expanded, a neural network feature extractor is designed, single fault signal samples and corresponding fault prototypes are input for training, compound fault signals are input into the trained feature extractor to obtain attribute features, the corresponding compound fault types are inferred according to the distance, and a designed measurement method is used to evaluate the hubness problem. The application can effectively relieve the hubness problem, improve the accuracy of compound fault diagnosis, and realize rolling bearing fault diagnosis for compound faults only by using single fault data.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of mechanical fault diagnosis, and relates to a zero-sample prototype expansion method and system for relieving hubness problem and compound fault diagnosis. BACKGROUND

[0002] Rotating machinery is widely used in wind power generation, aerospace, transportation and other fields, and plays an important role in modern industry. Bearings, as one of the key components of rotating machinery, are crucial to the performance and reliability of equipment. However, bearings are prone to failure due to long-term operation under complex working conditions. With the increasing complexity of rotating machinery, in addition to single faults, machinery may also experience compound faults, which are more challenging to diagnose than single faults. Therefore, it is of great significance to study high-precision diagnosis methods for bearings and other components.

[0003] Existing compound fault diagnosis methods can be mainly divided into three categories: fault mechanism-based methods, signal processing technology-based methods and artificial intelligence-based methods. Fault mechanism-based methods mainly use dynamics and modal analysis to understand the mechanism of compound faults, assisting in developing effective maintenance strategies. Signal processing-based methods mainly use signal processing techniques to extract compound fault information and determine fault types through expert experience. However, fault mechanism and signal processing methods are heavily dependent on expert experience, limiting their application in actual industries. In recent years, artificial intelligence-based methods have received widespread attention in various fields due to their excellent performance. Artificial intelligence-based diagnosis methods can be mainly divided into traditional machine learning methods and deep learning-based methods. Although traditional machine learning and deep learning-based methods achieve high diagnostic accuracy, they rely on a large amount of fault data. However, it is unrealistic to obtain a large amount of fault data, especially compound fault data, in actual industries. In recent years, Lampert et al. developed a zero-sample learning method for image recognition to overcome the difficulty of target data collection. The zero-sample learning-based method does not require target samples to participate in training. Currently, zero-sample learning methods have also been extended to the field of fault diagnosis. Feng et al. first applied zero-sample learning to fault diagnosis, Qin et al. proposed a zero-sample attribute consistent model for fault diagnosis under unknown working conditions, and Hu et al. constructed a semantic consistent embedding model for industrial zero-sample fault diagnosis.

[0004] However, existing zero-sample diagnosis methods have the following problems:

[0005] 1) Current zero-shot diagnosis methods directly map the signal feature space to the attribute prototype space and use nearest neighbor estimation as the measurement method. This method is susceptible to the hubness problem, where some prototypes have samples appearing in the nearest neighbor estimates of other prototypes, negatively affecting the accuracy of fault diagnosis.

[0006] 2) In previous work, the defined attributes only consider fault type and frequency characteristics, lacking consideration of other typical fault characteristics. In addition, existing hubness problem measurement methods are designed for high-dimensional data and are not suitable for zero-shot learning in some cases.

[0007] Hubness is a phenomenon in machine learning where certain sample points are more likely to be misclassified as neighbors of other sample points in some datasets. These sample points act as "central points" or "hubs" in the data space, attracting a large number of other sample points, leading to their easy misclassification in nearest neighbor classification algorithms.

[0008] Reasons for the occurrence of Hubness:

[0009] Uneven data distribution: In some datasets, the number of samples in certain classes is much larger than that in other classes, making these samples more likely to be misclassified as neighbors of other classes.

[0010] High-dimensional feature space: In high-dimensional space, the distance between sample points becomes more difficult to calculate, making it more likely to occur Hubness.

[0011] Inappropriate similarity measurement: Using inappropriate similarity measurement methods, such as Euclidean distance, may lead to the occurrence of Hubness.

[0012] Impact of Hubness:

[0013] Reduced classification accuracy: Hubness can reduce the accuracy of nearest neighbor classification algorithms and other algorithms because Hubness samples are more likely to be misclassified as neighbors.

[0014] Affecting clustering performance: Hubness can affect the performance of clustering algorithms because Hubness samples may be incorrectly clustered into other class clusters.

[0015] Increased computational complexity: To solve the Hubness problem, more complex algorithms or data preprocessing are needed, increasing computational complexity.

[0016] In order to alleviate the hubness problem, the application proposes a prototype expansion mapping method. The method designs a prototype expansion kernel function, which maps linear and nonlinear prototype features to a relatively high-dimensional space to enhance separability. The prototype expansion kernel function can increase the distance between prototypes, thereby alleviating the hubness problem and improving diagnostic accuracy. In view of the problem that the previous attribute definition ignores some important vibration features, the application designs a new attribute prototype definition method, which adds more fault features to the attribute. In addition, the application also proposes a way to measure the hubness problem in zero-shot learning. SUMMARY

[0017] Therefore, the application aims to provide a zero-shot prototype expansion method and system for alleviating the hubness problem and composite fault diagnosis.

[0018] To achieve the above-mentioned purpose, the application provides the following technical solutions:

[0019] As shown in Figure 1 A zero-shot prototype expansion method for alleviating the hubness problem and composite fault diagnosis, comprising the following steps:

[0020] S1: Collecting state signal samples of bearings of different types and different working conditions through a bearing fault test bench;

[0021] S2: For each type of bearing fault signal, combining the characteristics of bearing fault type, signal fault frequency and fault amplitude, designing the corresponding fault attribute prototype;

[0022] S3: For the designed fault attribute prototype, using a prototype expansion kernel function to map the prototype to a high-dimensional space to expand the distance between the designed attribute prototypes;

[0023] S4: Designing a neural network feature extractor, inputting single fault signal samples and corresponding fault prototypes for training;

[0024] S5: Inputting the composite fault signal into the trained feature extractor to obtain the attribute features;

[0025] S6: Using the nearest neighbor estimation method to calculate the distance between the extracted composite fault attribute features and the expanded fault attribute prototype features, and inferring the corresponding composite fault type according to the distance;

[0026] S7: Using the designed measurement method to evaluate the hubness problem.

[0027] Further, in S2, the attribute definition method for the designed rolling bearing includes 9 attributes of fault type, impact frequency and impact amplitude. The attributes designed for the rolling bearing are shown in Table 1. Because the data of single fault can be directly obtained, the corresponding attribute prototype P S can be directly defined. The attribute prototype P C of the compound fault needs to be defined through logical operation of the single fault attribute. Through logical or operation, the attributes related to fault type and impact frequency are obtained:

[0028]

[0029] wherein represents the 1-n elements of the i-th prototype of single fault.

[0030] Through logical or operation, the attributes related to fault impact amplitude are obtained:

[0031]

[0032] wherein, D A represents the dimension of the attribute, represents the n+1-D A elements of the w-th prototype of compound fault, represents the n+1-D A elements of the i-th prototype of single fault.

[0033] Table 1

[0034]

[0035]

[0036] Further, in S3, the prototype expansion kernel function is proposed, which can effectively increase the distance between the prototypes and thus alleviate the hubness problem. The specific definition of the function is as follows:

[0037]

[0038] wherein γ1, γ2 are weight parameters of two parts of the kernel function respectively, c is a constant term, d is the degree of the polynomial, γ1, γ2≥1, d∈N+, c≥0.

[0039] Further, the proof of the function in S3 is as follows. First, the symmetry of the function is proved:

[0040]

[0041] Secondly, the positive definiteness of the function is proved, that is, it is proved that for any non-zero vector a TKa≥ 0, where K is the kernel matrix of the proposed function, which is as follows:

[0042]

[0043] According to the properties of linear kernel function and polynomial kernel function, we have:

[0044]

[0045] Obviously, when γ1, γ2≥ 1:

[0046] a T Ka≥ 0 (8)

[0047] In summary, the feasibility of the designed function as a kernel function is verified.

[0048] Further, in S3, the proposed kernel function can expand the distance between the prototypes, which is proved as follows: Let P1= [α1, α2, …, α m ] and P2= [β1, β2, …, β m ] represent the prototypes of any two faults, respectively. The prototypes after mapping by the kernel function can be represented as:

[0049]

[0050]

[0051] The Euclidean distance between the two prototypes before mapping is:

[0052]

[0053] The Euclidean distance between the two prototypes after mapping is:

[0054]

[0055] The difference between the two distances can be calculated as:

[0056]

[0057] When P1 and P2 are not zero vectors, we have:

[0058]

[0059] When α i , β i ∈ {0, 1}:

[0060]

[0061] Therefore, according to the above analysis, the proposed kernel function can expand the distance between the prototypes.

[0062] Further, S4 designs a convolutional neural network as a feature extractor, which mainly consists of four convolutional modules, one flatten layer and three fully connected layers. The first convolutional module includes a wide kernel convolutional layer, a batch normalization layer and a max-pooling layer. The second to fourth convolutional modules have the same configuration, including a normal convolutional layer, an expanded convolutional layer, a batch normalization (BN) layer and a max-pooling layer. Nonlinear rectified linear unit (ReLU) is used to activate all network layers. The specific parameters of the feature extractor are shown in Table 1. Figure 2 For example, in the fourth convolutional layer, (750, 32) represents the input shape of the feature, and 3-1-64-2 represents that the convolution kernel size is 3, the stride is set to 1, the channel number is set to 64, and the expansion rate is set to 2.

[0063] Further, in the training process of the model in S4, the mean square error loss is used to align the deep features and the mapped attribute prototypes:

[0064]

[0065] wherein, represents the jth element of the attribute feature of the ith sample, represents the jth element of the attribute prototype mapped by the kernel function of the ith sample, N S is the number of samples of a single fault, and m is the dimension of the designed prototype.

[0066] Further, in S6, the nearest neighbor estimation is used to infer the fault type of diagnosis:

[0067]

[0068] wherein, y c represents the label of a composite fault, represents the ith element of the bth row of the attribute prototype matrix M c mapped by the kernel function.

[0069] Further, the evaluation index of the hubness problem in the zero sample in S7 is shown as follows:

[0070]

[0071] wherein N P is the number of test prototypes, N k (i) represents that the test sample belonging to the ith prototype appears in the k-nearest neighbors of other prototypes, and the Hub value is larger, indicating that the hubness problem is more serious.

[0072] The beneficial effects of the present application are:

[0073] (1) The present application can realize the fault diagnosis of rolling bearings with compound faults only through single fault data.

[0074] (2) The model constructed by the present application has higher diagnostic accuracy than existing classic and cutting-edge zero-shot models.

[0075] (3) The prototype expansion kernel function proposed by the present application can effectively enlarge the distance between prototypes, alleviate the hubness problem, and thus improve the diagnostic accuracy.

[0076] (4) The present application proposes a hubness problem measurement method that is more suitable for zero-shot learning than the original measurement method.

[0077] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, in conjunction with the accompanying drawings, and the advantages of the present application will be realized and attained by the practice of the present application. The objects and other advantages of the present application will be realized and attained by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to make the purposes, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be made below in conjunction with the drawings, in which:

[0079] Figure 1 Diagnosis process for zero-shot prototype expansion method

[0080] Figure 2 Detailed parameters of feature extractor

[0081] Figure 3 Diagnostic accuracy of different methods under different loads in NEEPU experiment

[0082] Figure 4 Diagnostic accuracy of different methods under different loads in BCF experiment

[0083] Figure 5 Diagnostic accuracy of the proposed model in different tasks; (a) 20Hz→25Hz; (b) 25Hz→20Hz; (c) 25Hz→30Hz; (d) 30Hz→25Hz

[0084] Figure 6 Hub values of different methods in NEEPU and BCF bearing experiments DETAILED DESCRIPTION

[0085] The present application can be implemented or applied by other different specific embodiments, and in the specification, various details can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the features in the following embodiments and examples can be combined with each other without conflict.

[0086] The accompanying drawings are only used for exemplary illustration, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation of the present application; in order to better illustrate the embodiments of the present application, some components in the drawings are omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings can be omitted.

[0087] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, therefore the terms describing the positional relationship in the drawings are only used for exemplary illustration, and cannot be understood as a limitation of the present application, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0088] The present application evaluates the proposed zero-shot prototypical expansion model through three experiments of composite fault diagnosis of rolling bearings under different working conditions. The first experiment is based on the bearing data set NEEPU collected by Northeast Electric Power University. The second experiment uses self-made bearing composite fault experiment platform data (BCF) for verification. The third experiment is verified based on the data set (HUST) of Huazhong University of Science and Technology.

[0089] The NEEPU dataset includes 7 different types of health conditions: normal condition NC, outer ring fault OF, inner ring fault IF, rolling element fault BF, outer ring and inner ring fault OI, rolling element and outer ring fault BO, inner ring and ball fault IB. Various loads are applied to the bearing through a magnetic brake, including 0 N-m (load 0), 0.1 N-m (load 1), 0.2 N-m (load 2), and 0.3 N-m (load 3). The rotational speed range is [1443, 1478] rpm, and the sampling frequency is 12 kHz. The number of samples for each state in the training set is 700. The number of samples for each state in the test set is 300. That is, the number of fault samples is 2800, and the number of composite fault samples in the test set is 900. In the proposed model, only single faults are involved in training, while composite faults are used for testing. The length of each sample is taken as 3072 to obtain more state information, and a sliding window sampling technique is used to obtain additional samples. By comparing the seven classical and frontier zero-shot learning methods of DAP, IAP, SJE, ESZSL, DEM, FDAT and SCE, the advantages of the proposed method are shown. In addition, the prototype expansion method proposed is verified by ablation experiments, and the model without prototype expansion is denoted as ZSPE-WE.

[0090] Figure 3 The diagnostic accuracy of different zero-shot methods under four different loads is shown. Without the participation of composite faults in training, the average diagnostic accuracy of the ZSPE model for composite faults under four loads is 69.31%, 82.58%, 86.56%, and 83.93%, respectively. The proposed model has higher accuracy than the other seven classical and advanced zero-shot learning models. Compared with the ZSPE-WE method, this method has better improvement in accuracy, which verifies the effectiveness of the prototype expansion method proposed.

[0091] To further verify the generalization of the model, a zero-shot diagnosis experiment was conducted on a self-made bearing composite fault test platform. The platform mainly consists of a load motor, a rotational speed-torque measuring instrument, a planetary gear reducer, a bearing test module, and a driving motor. The experimental data set includes 7 types of conditions of rolling bearings under 4 working conditions, which is the same as the NEEPU experiment. The type of test bearing is SKF61908, and 4 loads are applied through the load motor, which are 0 N-m (load 0), 2 N-m (load 1), 4 N-m (load 2), and 6 N-m (load 3), the rotational speed is set to 500 rpm, and the sampling frequency is taken as 12 kHz.

[0092] The average diagnostic accuracy of different zero-shot learning methods in the BCF experiment is as follows Figure 4The average diagnosis accuracy of load 0 to load 3 composite faults is 65.87%, 75.64%, 96.71% and 87.33% respectively when only single fault data is involved in training. Figure 4 As can be seen from

[0093] In order to further verify the generalization ability of the model, the present application carries out cross-domain zero sample diagnosis experiments of multiple tasks based on the HUST bearing data set. In these experiments, only the data under the known working condition is used for training, and the data under the unknown working condition is directly tested. The bearing fault data set is collected by a mechanical fault simulator, which mainly consists of a speed controller, a motor, a shaft, an acceleration sensor, a test bearing and a data acquisition board. In this experiment, five different states of rolling bearings are used, including normal state, outer ring fault, inner ring fault, ball fault, and outer ring and inner ring composite fault. This experiment adopts three different speeds of 20Hz, 25Hz and 30Hz, and the sampling frequency is set to 25.6kHz.

[0094] In the four different cross-domain zero sample diagnosis experiments, the average diagnosis accuracy of the model is 92.53%±1.14, 92.87%±2.15, 97.69%±1.06 and 97.68%±0.88 respectively. It can be known that in the four cases, the diagnosis accuracy is more than 90%. Among them, the diagnosis accuracy of two cases is more than 97%, which further proves the generalization performance of the proposed model. In addition, the confusion matrix obtained in the four fault diagnosis experiments is shown as Figure 5 (a) is 20Hz→25Hz; (b) is 25Hz→20Hz; (c) is 25Hz→30Hz; (d) is 30Hz→25Hz; which further verifies the performance of the proposed model.

[0095] In addition, in the NEEPU and BCF two experiments, the proposed measure (equation (18)) is used to measure the hubness problem to illustrate the effectiveness of the proposed method in alleviating the hubness problem. As shown in Figure 6 In all eight working conditions of the two experiments, the Hub value of ZSPE-WE is much higher than that of ZSPE. The results show that the method reduces the number of times the corresponding sample features appear in the nearest neighbor estimation of other prototypes. It can be concluded that the proposed method can identify the features corresponding to the correct prototype, thereby improving the accuracy of diagnosis.

[0096] Finally, it is to be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the purpose and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A zero-shot prototype extension method for alleviating hubness problems and complex fault diagnosis, characterized by: The following steps are involved: S1: Collect state signal samples of bearings of different types and working conditions through the bearing fault test bench; S2: For various bearing fault signals, design corresponding fault attribute prototypes based on the characteristics of the bearing fault type, signal fault frequency, and fault amplitude. The attribute prototypes include the attributes of fault type, impact frequency, and impact amplitude. Compound Fault Prototype Prototype with a single fault The logical operation is: The properties related to the fault type and impact frequency are obtained through logical OR operation: in, Indicates the w The 1st to nth elements of a composite fault prototype, Represents the 1st to nth elements of the i-th single fault prototype; The properties related to the fault impulse amplitude are obtained through logical AND operation: in, represents the dimension of the attribute, Indicates the w The n+1th compound fault prototype elements, Indicates the i The n+1th of a single fault prototype elements; S3: For the designed fault attribute prototype, the prototype expansion kernel function is used to map the prototype to a high-dimensional space to expand the distance between the designed attribute prototypes; the prototype expansion kernel function is defined as: in, are the weight parameters of the two parts of the kernel function, is a constant term, is the degree of the polynomial, , N+ represents a positive integer; S4: Design a neural network feature extractor and input a single fault signal sample and the corresponding fault prototype for training; S5: Input the composite fault signal into the trained feature extractor to obtain attribute features; S6: Use the nearest neighbor estimation method to calculate the distance between the extracted composite fault attribute features and the expanded fault attribute prototype features, and infer the corresponding composite fault type based on the distance; S7: Use the designed measurement method to evaluate the hubness problem.

2. The zero-shot prototype extension method for alleviating hubness problems and complex fault diagnosis according to claim 1 is characterized by: In S4, the neural network feature extractor is a convolutional neural network; the convolutional neural network consists of four convolution modules, a flattening layer, and three fully connected layers; the first convolution module includes a wide kernel convolution layer, a batch normalization layer, and a maximum pooling layer; the second to fourth convolution modules have the same configuration, including a normal convolution layer, a dilated convolution layer, a batch normalization BN layer, and a maximum pooling layer; The nonlinear rectified linear unit ReLU is used to activate all network layers; during the training process of the convolutional neural network, the mean squared error loss is used to align the deep features and the mapping attribute prototypes: in, Indicates the i The attribute features of the sample j elements, Indicates the i The first attribute prototype of the sample after the kernel function mapping j elements, is the number of samples of a single fault, m Dimensions for designing prototypes.

3. The zero-shot prototype extension method for alleviating hubness problems and complex fault diagnosis according to claim 1 is characterized by: In S6, the nearest neighbor estimation is used to infer the type of fault to be diagnosed: in, A label indicating a compound fault, Represents the attribute prototype matrix mapped by the kernel function No. b The first i elements, m Dimensions for designing prototypes.

4. The zero-shot prototype extension method for alleviating hubness problems and complex fault diagnosis according to claim 1 is characterized by: In S7, the evaluation index of the hubness problem is defined as: in, is the number of tested prototypes, Indicates that it belongs to i The number of test samples of a prototype that appear in the k nearest neighbors of other prototypes. The larger the Hub value, the more serious the Hubness problem.

5. A system for alleviating hubness problems and diagnosing complex faults using the zero-sample prototype expansion method according to any one of claims 1 to 4, characterized in that: include: A signal acquisition module is used to collect status signal samples of the bearing; The attribute prototype design module is used to design the corresponding fault attribute prototype based on the characteristics of the bearing fault type, signal fault frequency and fault amplitude; The prototype expansion module is used to map the designed fault attribute prototype to a high-dimensional space using the prototype expansion kernel function, thereby expanding the distance between the designed attribute prototypes; Feature extraction module, used to design a neural network feature extractor, which inputs single fault signal samples and corresponding fault prototypes for training; The diagnosis module is used to input the composite fault signal into the trained feature extractor to obtain attribute features, and use the nearest neighbor estimation method to calculate the distance between the extracted composite fault attribute features and the expanded fault attribute prototype features, and infer the corresponding composite fault type based on the distance; The evaluation module is used to evaluate the hubness problem using the designed measurement method.

6. The system for alleviating hubness problems and diagnosing complex faults according to claim 5, characterized in that: The attribute prototype design module further includes a fault type identification unit, a fault frequency extraction unit and a fault amplitude extraction unit; All three modules receive bearing status signal samples from the signal acquisition module; The fault type identification unit is responsible for identifying the type of bearing fault, the fault frequency extraction unit is responsible for extracting the fault frequency characteristics of the bearing fault signal, and the fault amplitude extraction unit is responsible for extracting the fault amplitude characteristics of the bearing fault signal; These three modules pass their respective output results to the attribute prototyping module; The attribute prototype design module designs the corresponding fault attribute prototype based on the output results of these three modules.

7. The system for alleviating hubness problems and diagnosing complex faults according to claim 5, characterized in that: The prototype extension module further includes a kernel function calculation unit and a mapping unit; Both modules receive fault attribute prototypes from the attribute prototyping module; The kernel function calculation unit is responsible for calculating the kernel function value between attribute prototypes according to the preset kernel function formula; The mapping unit receives the kernel function values ​​from the kernel function calculation unit and uses these values ​​to map the attribute prototype to a high-dimensional space; The mapping unit outputs the expanded attribute prototype to the feature extraction module.

8. The system for alleviating hubness problems and diagnosing complex faults according to claim 5, characterized in that: The evaluation module further includes a hubness problem calculation unit and an output unit; Both modules receive the fault type inference results from the diagnosis module; The Hubness problem calculation unit is responsible for calculating the value of each test sample in the hubness problem metric based on the fault type inference result; The output unit receives the calculation results from the Hubness problem calculation unit and outputs the evaluation results of the hubness problem.

Images