Method for bearing fault diagnosis based on uncertainty quantification

Abstract

The application discloses a bearing fault diagnosis method based on uncertainty quantification, collects a training sample set, each training sample including two different modal signals of a bearing, constructs a bearing fault diagnosis model, adopts two feature extraction modules to extract modal features from the two modal signals respectively, and a prediction module adopts uncertainty quantification based on D-S evidence theory to generate the probability that the bearing belongs to each fault state; the training sample set is used to train the bearing fault diagnosis model, and the trained bearing fault diagnosis model is used for fault detection of the bearing. The application constructs a macro uncertainty quantification strategy based on D-S evidence theory for multi-modal signals, adaptively solves the uncertainty probability of each overall modal, and then realizes effective fusion of multi-source monitoring signals, and improves the accuracy and reliability of bearing fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of bearing diagnostic technology, and more specifically, relates to a bearing fault diagnosis method based on uncertainty quantification. Background Technology

[0002] As a core supporting component of mechanical equipment, the operating status of bearings directly affects the overall stability and safety of the equipment. Accurate fault diagnosis is a crucial means to ensure reliable equipment operation. With the development of monitoring technology, multimodal monitoring signals (such as vibration signals, current signals, and temperature signals) are widely used in bearing fault diagnosis. By fusing information from different modes, the limitations of single-mode information can be effectively compensated for, improving the accuracy and comprehensiveness of fault diagnosis. However, existing multimodal fusion diagnostic technologies still face many unresolved technical challenges in practical applications, severely restricting the improvement of fault diagnosis performance. These challenges are as follows:

[0003] On the one hand, existing multimodal fusion methods lack fine-grained screening of fault information within a single mode, making fault features easily diluted by interference components. The original monitoring signals typically contain a large amount of background noise, operating condition fluctuations, and other irrelevant interference components. Current mainstream multimodal fusion schemes mostly adopt a direct fusion approach, indiscriminately fusing all signal components within each mode. This processing method causes effective fault features to be masked or even diluted by irrelevant interference components, failing to fully highlight the key information of the fault features. This can reduce diagnostic accuracy at best, and at worst cause the entire multimodal fusion framework to fail, making effective identification of bearing faults impossible.

[0004] On the other hand, existing technologies cannot quantify the uncertainty of a single mode's contribution to the fault category during multimodal fusion, and fixed weight allocation patterns easily lead to insufficient reliability of diagnostic results. During model training, influenced by the adaptive changes in the decision boundary, the contribution of different modes to fault features is not constant but exhibits dynamic fluctuations. However, existing multimodal fusion schemes typically employ fixed weight allocation strategies, failing to quantify the dynamic uncertainty of each mode's contribution. This allows low-quality modes (such as those severely disturbed or with unclear fault features) to negatively interfere with the overall fusion process, causing the model to blindly output incorrect diagnostic decisions, significantly reducing the reliability and stability of diagnostic results, and making it difficult to meet the accuracy requirements of fault diagnosis in real-world industrial scenarios. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a bearing fault diagnosis method based on uncertainty quantification. For multimodal signals, a macroscopic uncertainty quantification strategy based on DS evidence theory is constructed to adaptively solve the uncertainty probability of each global mode, thereby realizing the effective fusion of multi-source monitoring signals and improving the accuracy and reliability of bearing fault diagnosis.

[0006] To achieve the above-mentioned objectives, the bearing fault diagnosis method based on uncertainty quantification of the present invention includes the following steps:

[0007] S1: Collect bearing data under normal conditions and Multimodal signal samples under various fault conditions, each multimodal signal sample including two different modes of signal within the same time period. , , Indicates time, sample label Set to fault status. The training sample set is constructed.

[0008] S2: Construct a bearing fault diagnosis model, including a first feature extraction module, a second feature extraction module, and a prediction module, wherein:

[0009] The first feature extraction module is used to extract the first mode signal from the input multimodal signal. Feature extraction is performed to obtain the first modality features. And send it to the prediction module;

[0010] The second feature extraction module is used to extract the second mode signal from the input multimodal signal. Feature extraction is performed to obtain the second modality features. And send it to the prediction module;

[0011] The prediction module is used to predict based on two modal features. The bearing is predicted to belong to The probability of each fault state; the prediction module includes a first evidence vector generation module, a second evidence vector generation module, a Dirichlet sampling module, an adaptive weight allocation module, a correction module, a basic probability allocation fusion module, and a probability generation module, wherein:

[0012] The first evidence vector generation module is used to generate evidence vectors based on the first modality features. Generate a non-negative first evidence vector And send it to the Dirichlet sampling module:

[0013] ,

[0014] in, Indicates the first mode in the fault category Evidence value, , This represents the weights and biases corresponding to the first mode. Represents a smooth nonlinear activation function;

[0015] The second evidence vector generation module is used to generate evidence based on the second modality features. Generate a non-negative second evidence vector And send it to the Dirichlet sampling module:

[0016] ,

[0017] in, Indicates the second mode in the fault category Evidence value, , This represents the weights and biases corresponding to the second mode;

[0018] Dirichlet sampling module is based on two evidence vectors The basic probability assignments for the DS evidence theory are generated by sampling and mapping using the Dirichlet distribution. The specific method is as follows:

[0019] Based on each evidence vector Calculate the total intensity of the distribution :

[0020] ,

[0021] in, Indicates the first In the mode, the th Concentration parameters of the Dirichlet distribution of the class;

[0022] Then, the basic probability distribution of a single-element focal element is calculated using the following formula. Basic probability allocation of the focal element of the entire set Basic probability allocation of compound focal elements :

[0023] ,

[0024] in, Indicates the first Fault-like states, Represents a set of fault states. Indicates the first The modality pair of the first Determinism of class-specific faults, Indicates the first Overall cognitive uncertainty across modalities;

[0025] Then, the overall cognitive uncertainty The data is sent to the adaptive modal weighting module to assign the basic probability of the single-element focal element. Basic probability allocation of the focal element of the whole set Send to the correction module;

[0026] The adaptive modal weighting module is used to adjust the overall cognitive uncertainty. The weights of the two modes are calculated. And send it to the correction module, weight The calculation formula is as follows:

[0027] ,

[0028] in, Indicates the adjustment parameter. Indicates the compensation coefficient;

[0029] The correction module is used to adjust the weights. Basic probability assignment of a single-element focal element Basic probability allocation of the focal element of the whole set After making corrections, we obtain the corrected basic probability distribution. , And send it to the basic probability allocation fusion module; the corrected formula is as follows:

[0030] ;

[0031] The basic probability assignment fusion module is used to assign the basic probability of the modified single-element focal element using the DS orthogonal sum rule. Basic probability allocation of the focal element of the whole set By performing fusion, we obtain the basic probability distribution of the fusion of single-element focal elements and universal focal elements. , And send it to the probability vector generation module; the fusion formula is as follows:

[0032] ,

[0033] ,

[0034] in, The conflict coefficient is defined as follows:

[0035] ,

[0036] The probability generation module assigns a basic probability based on the fusion of single-element focal elements and the total set of focal elements. , Obtain the probability that the final bearing belongs to each fault state. :

[0037] ;

[0038] S3: The bearing fault diagnosis model is trained using the training sample set to obtain a trained bearing fault diagnosis model;

[0039] S4: When bearing fault detection is required, two different modal signals within the same time period are collected in real time and input into the trained bearing fault diagnosis model to obtain the fault diagnosis result.

[0040] This invention relates to a bearing fault diagnosis method based on uncertainty quantification. A training sample set is collected, with each training sample including two different modal signals of the bearing. A bearing fault diagnosis model is constructed, and two feature extraction modules are used to extract modal features from the two modal signals respectively. The prediction module uses uncertainty quantification based on DS evidence theory to generate the probability of the bearing belonging to each fault state. The bearing fault diagnosis model is trained using the training sample set, and the trained bearing fault diagnosis model is used to detect bearing faults.

[0041] The present invention has the following beneficial effects:

[0042] 1) For feature extraction of modal signals, this invention proposes a microphysical embedding selection mechanism to evaluate the weights corresponding to different components within a single mode, thereby achieving fine-grained screening of single-mode fault features and enhancing the ability to characterize intrinsic fault features.

[0043] 2) This invention constructs a macroscopic uncertainty quantification strategy based on DS evidence theory, adaptively solves the uncertainty probability of each global mode, and thus realizes the effective fusion of multi-source monitoring signals, improving the accuracy and reliability of fault diagnosis.

[0044] 3) In terms of training the bearing fault diagnosis model, this invention also designs a joint optimization strategy that combines global evidence calibration and local weight regularization to improve modal performance and ensure fault diagnosis efficiency. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a specific implementation of the bearing fault diagnosis method based on uncertainty quantification according to the present invention.

[0046] Figure 2 This is a structural diagram of the bearing fault diagnosis model in this invention;

[0047] Figure 3 This is an example diagram of the CEEMDAN decomposition result of the original signal in this embodiment;

[0048] Figure 4 yes Figure 3 The probability density distribution of the amplitudes of each order of IMF components is shown below;

[0049] Figure 5 This is a schematic diagram of the weighted fusion of IMF component features;

[0050] Figure 6 This is a schematic diagram of the framework of the macroscopic uncertainty quantification strategy in this invention. Detailed Implementation

[0051] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0052] Example

[0053] Figure 1 This is a flowchart illustrating a specific implementation of the bearing fault diagnosis method based on uncertainty quantification according to the present invention. Figure 1 As shown, the specific steps of the bearing fault diagnosis method based on physical embedding and uncertainty quantification of the present invention include:

[0054] S101: Obtain the training sample set:

[0055] Collect bearing data under normal conditions and Multimodal signal samples under various fault conditions, each multimodal signal sample including two different modes of signal within the same time period. , , Indicates time, sample label Set to fault status. This allows us to construct a training sample set.

[0056] Generally, signals used for bearing fault diagnosis include vibration signals, current signals, acoustic emission signals, temperature signals, etc. In this embodiment, a classic combination of vibration signals and current signals is used.

[0057] S102: Constructing a bearing fault diagnosis model:

[0058] To achieve bearing fault diagnosis, this invention constructs a bearing fault diagnosis model. Figure 2 This is a structural diagram of the bearing fault diagnosis model in this invention. (See diagram below.) Figure 2 As shown, the bearing fault diagnosis model in this invention includes a first feature extraction module, a second feature extraction module, and a prediction module, wherein:

[0059] The first feature extraction module is used to extract the first mode signal from the input multimodal signal. Feature extraction is performed to obtain the first modality features. And send it to the prediction module.

[0060] The second feature extraction module is used to extract the second mode signal from the input multimodal signal. Feature extraction is performed to obtain the second modality features. And send it to the prediction module.

[0061] Because fault features in early monitoring signals are often weak and easily masked by noise and operating conditions, and fault-irrelevant components are often coupled with effective components, they are difficult to accurately model and directly remove. Therefore, this embodiment no longer relies on explicit noise reconstruction and removal, but proposes a micro-physical embedding selection mechanism. The core idea is to first decompose each modal signal into several intrinsic mode functions (IMFs), and then construct a reference benchmark using normal signals. Physical difference quantization reflects the degree of deviation of the input signal's IMFs from the normal signal; a greater deviation indicates a more obvious fault feature. Based on this, the weights of the IMFs are adaptively evaluated, guiding the network to extract and weight modal features, thereby achieving effective representation of fault features.

[0062] like Figure 2 As shown, in this embodiment, the first feature extraction module and the second feature extraction module adopt the same structure, which respectively include an empirical mode decomposition module, a physical difference quantization module, a mode weight generation module, a mode feature extraction module, and a weighted fusion module, wherein:

[0063] The ensemble empirical mode decomposition module is used to analyze modal signals. The IMF components are obtained by performing ensemble empirical module decomposition. And send it to the physical difference quantization module and the modal feature extraction module. , This represents the number of components in the set's empirical mode.

[0064] In this embodiment, the signal is decomposed using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). Figure 3 This is an example diagram of the CEEMDAN decomposition result of the original signal in this embodiment. Figure 3 Taking the decomposition of China into six IMF components as an example. Figure 3 As shown, through this decomposition method, the original complex signal is mapped onto multiple frequency bands, enabling the model to measure the physical differences between different frequency bands at a microscopic level.

[0065] The physical difference quantification module is used to quantify each IMF component. Physical difference quantization is performed on the normal reference IMF component of the corresponding mode in the normal state multimodal signal sample to obtain the physical difference quantization vector of each IMF component in that mode. And send it to the modal weight generation module.

[0066] To comprehensively quantify physical differences, this embodiment characterizes them from two perspectives: distribution shift and shock mutation. The former is represented by KL divergence. Quantification describes whether the overall distribution has changed; the latter is determined by kurtosis. Quantification is used to characterize common early-stage failure features such as shocks.

[0067] First, KL divergence is used to characterize the degree of deviation in statistical distribution between the IMF component of the current modal signal and the normal reference IMF component. Let the i-th component in the input modal signal be... The first mode The probability distribution of each IMF component is as follows The probability distribution corresponding to the normal reference IMF component is as follows: The KL divergence between the two is defined as:

[0068] ,

[0069] in, and These are the probability density functions of the corresponding distributions.

[0070] Figure 4 yes Figure 3 The probability density distribution of the amplitudes of each order of IMF components is shown in the diagram. Figure 4 As shown, the amplitudes of each IMF component follow a Gaussian distribution. Therefore, the probability distributions of different components in the input multimodal signal and the normal reference are as follows:

[0071] , ,

[0072] in, , These represent the first and second modal components of the input multimodal signal, respectively. The first mode Mean and standard deviation of each IMF component , They represent the statistical results obtained from the normal signal samples in the training sample set. The first mode The mean and standard deviation of each IMF component.

[0073] At this point, the KL divergence can be obtained. The expression is:

[0074] .

[0075] The above formula can be used to characterize the degree of abnormality of the input multimodal signal relative to the normal state.

[0076] Next, kurtosis is used to characterize the impulse response of local components from the perspective of waveform morphology. Let the kurtosis be the first component in the input multimodal signal. The first mode The kurtosis of each IMF component is denoted as The first normal signal sample obtained from the training sample set The first mode The kurtosis of each IMF component is denoted as Then define the first The first mode Kurtosis between each IMF component and its corresponding normal reference IMF component for:

[0077] .

[0078] In this embodiment, kurtosis The following formula is used for calculation:

[0079] ,

[0080] in, Indicates the length of the signal. This indicates the prevention of the minimum value where the denominator is 0.

[0081] When the bearing is in normal condition, even if the change in operating conditions causes the overall energy of the signal to increase, its waveform usually remains relatively stable and the kurtosis will not increase significantly. However, when local damage causes mechanical shock, the pulse component of the signal is significantly enhanced, thereby increasing the kurtosis difference.

[0082] By leveraging the complementarity of KL divergence and kurtosis difference, a physical difference quantization vector is constructed. This enables a joint characterization of the physical differences of local components.

[0083] The modal weight generation module is used to quantize vectors based on physical differences. Generate corresponding weights And send it to the weighted fusion module.

[0084] Due to the KL divergence in this embodiment and kurtosis The representation of faults does not satisfy a fixed linear relationship. Therefore, in this embodiment, the modal weight generation module adopts a state-aware network, that is, the modal weight generation module includes a multilayer perceptron and a Softmax layer, wherein:

[0085] Multilayer perceptrons are used to quantize physical difference vectors. Generate score values ​​by performing nonlinear mapping And then sent to the Softmax layer. The processing of the multilayer perceptron can be represented by the following formula:

[0086] ,

[0087] in, , , and ...

[0088] The Softmax layer is used to apply the Softmax function to the score values. Normalization is performed to obtain the weights of the corresponding IMF components. Weight The calculation formula can be expressed as follows:

[0089] , .

[0090] The modal feature extraction module is used to extract features from each IMF component. Extract the corresponding features And send it to the weighted fusion module.

[0091] In this embodiment, the specific method for feature extraction by the modal feature extraction module is as follows:

[0092] For each IMF component Normalization is performed to obtain the normalized components. In this embodiment, mean-variance normalization is used. Then, a 1D-CNN feature extractor is used to process the normalized components. Features were extracted , can be represented as follows:

[0093] ,

[0094] in, Indicates the first A 1D-CNN feature extractor for each modality.

[0095] The weighted fusion module is used to calculate the weights. Features of each component in each mode Weighted fusion is performed to obtain the final features of each modality. :

[0096] .

[0097] Figure 5 This is a schematic diagram of the weighted fusion of IMF component features. For example... Figure 5 As shown, in this embodiment, the input signal is decomposed into three IMF components. Each component is processed by feature extraction to generate a corresponding feature vector. The feature vectors are then fed into a weighted fusion box, and each feature vector is weighted by adaptive weights. Finally, the summation and aggregation are performed to output a fused feature vector, thus realizing adaptive feature fusion of multiple IMF components.

[0098] The prediction module is used to predict features based on two modalities. The predicted bearing belongs to The probability of a fault state.

[0099] Traditional multimodal fusion assumes equal reliability among modes. However, in scenarios with strong noise interference or complex operating conditions, a single mode can easily generate high uncertainty. In such cases, equal-weighted fusion blindly introduces uncertainty into the decision-making process, severely weakening the model's reliability. Therefore, this invention constructs a macroscopic uncertainty quantification strategy based on DS evidence theory to achieve prediction. Figure 6 This is a schematic diagram illustrating the framework of the macroscopic uncertainty quantification strategy in this invention. For example... Figure 6 As shown, the multimodal features obtained at the microscopic level are further mapped to Dirichlet distribution parameters, achieving a unified representation of fault category determinism and overall cognitive uncertainty. Based on this, the uncertainty probability of each mode is adaptively solved according to the uncertainty of the mode in relation to the fault category, thereby achieving effective fusion of multi-source monitoring signals and improving the reliability of diagnostic results. Figure 2 and Figure 6 As shown, the prediction module in this invention includes a first evidence vector generation module, a second evidence vector generation module, a Dirichlet sampling module, an adaptive weight allocation module, a correction module, a basic probability allocation fusion module, and a probability generation module, wherein:

[0100] The first evidence vector generation module is used to generate evidence vectors based on the first modality features. Generate a non-negative first evidence vector And send it to the Dirichlet sampling module:

[0101] ,

[0102] in, Indicates the first mode in the fault category Evidence value, , This represents the weights and biases corresponding to the first mode. This represents a smooth nonlinear activation function that guarantees non-negativity of modal evidence.

[0103] The second evidence vector generation module is used to generate evidence based on the second modality features. Generate a non-negative second evidence vector And send it to the Dirichlet sampling module:

[0104] ,

[0105] in, Indicates the second mode in the fault category Evidence value, , This represents the weights and biases corresponding to the second mode.

[0106] Dirichlet sampling module is used for sampling based on two evidence vectors The Basic Probability Assignment (BPA) of the DS evidence theory is generated by sampling and mapping using the Dirichlet distribution. The specific method is as follows:

[0107] For any mode , its first Class evidence value The larger the value, the more likely the mode response belongs to a fault category. The stronger the support, the better. Therefore, this invention will use evidence vectors. The concentration parameters are transformed into a Dirichlet distribution to quantify the uncertainty of the mode for the fault category, thus obtaining its total distribution intensity. :

[0108] ,

[0109] in, Indicates the first In the mode, the th Concentration parameters of the Dirichlet distribution of the class.

[0110] Given that the fault diagnosis task of this invention has a clearly defined category set, only two scenarios are considered: support for specific fault categories and overall ignorance. Therefore, the basic probability assignment (BPA) of a single-element focal element is employed. BPA (unknown terms) of the entire set of focal elements. BPA of compound coke elements They are defined as follows:

[0111] ,

[0112] in, Indicates the first Fault-like states, Represents a set of fault states. Indicates the first The modality pair of the first Determinism of class-specific faults, Indicates the first Overall cognitive uncertainty across modalities.

[0113] Then, the overall cognitive uncertainty The data is sent to the adaptive modal weighting module to assign the basic probability of the single-element focal element. Basic probability allocation of the focal element of the whole set Send to the correction module.

[0114] The adaptive modal weighting module is used to adjust the overall cognitive uncertainty. The weights of the two modes are calculated. And send it to the correction module. This invention utilizes the uncertainty of each modal output to dynamically allocate weights, thereby reducing the proportion of unreliable modes in the final fusion. Specifically, the weights... The calculation formula is as follows:

[0115] ,

[0116] in, This represents the adjustment parameter, used to control the magnitude of weight changes due to uncertainty. This represents the compensation coefficient, used to adjust the degree of influence of the other party's uncertainty on the weights of this mode. Therefore, it can be seen that... and It is simultaneously constrained by both the uncertainty of this mode and the uncertainty of the other mode.

[0117] The correction module is used to adjust the weights. Basic probability assignment of a single-element focal element Basic probability allocation of the focal element of the whole set After making corrections, we obtain the corrected basic probability distribution. and basic probability allocation Send to the basic probability allocation fusion module. The corrected formula is as follows:

[0118] .

[0119] After correction, when the uncertainty of a single mode is high, evidence pointing to a specific fault will be adaptively suppressed. The suppressed evidence is redistributed to the unknown terms, thereby minimizing blind decision-making.

[0120] The basic probability assignment fusion module is used to assign the basic probability of the modified single-element focal element using the DS orthogonal sum rule. Basic probability allocation of the focal element of the whole set By performing fusion, we obtain the basic probability distribution of the fusion of single-element focal elements and universal focal elements. , And send it to the probability vector generation module. The fusion formula is as follows:

[0121] ,

[0122] ,

[0123] Among them, the denominator Used for normalization, The conflict coefficient is defined as follows:

[0124] ,

[0125] The probability generation module assigns a basic probability based on the fusion of single-element focal elements and the total set of focal elements. , Obtain the probability that the final bearing belongs to each fault state. :

[0126] .

[0127] S103: Training the bearing fault diagnosis model:

[0128] The bearing fault diagnosis model is trained using a training sample set to obtain a well-trained bearing fault diagnosis model.

[0129] To improve the training performance of the bearing fault diagnosis model, this embodiment designs a loss function based on global evidence calibration. The aim is to simultaneously ensure classification accuracy and effectively constrain cognitive uncertainty. Loss function Classification of risk items With evidence calibration items It consists of two parts, namely:

[0130] ,

[0131] in, This represents the weighting coefficient; in this embodiment, the annealing coefficient is used. , Indicates the number of iterations. Indicates the annealing cycle.

[0132] First, by minimizing prediction bias and prediction variance, the classification accuracy and prediction stability of the model are improved; therefore, the classification risk term... The calculation formula is:

[0133] ,

[0134] in, This indicates that the input multimodal signal obtained based on the real label belongs to the first... The probability of a fault state. This indicates that the input multimodal signal predicted by the model belongs to the . The probability of a fault state. This represents the total concentration parameter of the Dirichlet distribution. ,in Indicates the first The fusion concentration parameter of the Dirichlet distribution is calculated using the following formula:

[0135] .

[0136] Secondly, by minimizing the KL divergence between the modified Dirichlet distribution and the uniform distribution, the model is calibrated for incorrect category evidence and cognitive uncertainty, ensuring that the model maintains high uncertainty even when evidence is insufficient. Therefore, the evidence calibration term... The calculation formula is:

[0137] ,

[0138] in, This represents the corrected Dirichlet distribution concentration parameter, if the actual fault state of the input multimodal signal is... ,but ,otherwise . Represents the Gamma function. This represents the Digamma function.

[0139] When using a feature extraction module with physical difference quantization, to ensure that the IMF weights learned by the modal weight generation module are consistent with the physical differences of the local components and to avoid the weights being evenly distributed across all components, a local weight regularization term can be introduced. Physical alignment item With weight regularization term It consists of two parts, namely:

[0140] ,

[0141] in, , Representing physical alignment items With weight regularization term The weight.

[0142] First, cosine similarity is used to achieve structural alignment of IMF weights and physical differences, hence the physical alignment term. The calculation formula is:

[0143] ,

[0144] in, Indicates the first IMF weight vectors for each modality Indicates the first A vector of physical differences between each modality. .

[0145] Secondly, by minimizing the entropy of the IMF weight distribution, the average distribution of IMF weights is suppressed, hence the weight regularization term... The calculation formula is:

[0146] .

[0147] Finally, by combining the loss function based on global evidence calibration and the local weight regularization term, the overall loss function is obtained. The calculation formula is:

[0148] .

[0149] S104: Bearing Fault Diagnosis:

[0150] When bearing fault detection is required, two different modal signals within the same time period are collected in real time and input into the trained bearing fault diagnosis model to obtain the fault diagnosis result.

[0151] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the invention. To verify the effectiveness of the present invention under complex working conditions, this embodiment uses the rotating machinery multi-sensor dataset from the Korea Advanced Institute of Science and Technology (KAIST) for experimental verification. In this embodiment, vibration and current signals are selected as inputs, including four states: normal, inner race fault, outer race fault, and shaft misalignment. Data acquisition is performed under three load conditions: 0, 2, and 4 Nm, with a sampling frequency of 25.6 kHz for all conditions. In the data preprocessing stage, a sliding window technique is used, with a fixed interval length of 3072 sampling points, extracting 500 intervals for each state and load combination. The constructed multimodal dataset is divided into 7 cross-load diagnostic tasks; Table 1 shows the task division data table in this embodiment.

[0152] A1 0 A4 0、2 A2 2 A5 0、4 A3 4 A6 2、4 A7 0、2、4 \ \

[0153] Table 1

[0154] As shown in Table 1, in this embodiment, the model is randomly divided into a training set and a test set in a 4:1 ratio to evaluate its diagnostic performance.

[0155] In this embodiment, five existing methods were selected as comparison methods to verify the present invention, including two single-modal deep learning models (WDCNN and LSS) and three multi-modal information fusion models (DCAE, M2FN, and MIFN). Table 2 shows the accuracy comparison of the present invention and the comparison methods in seven diagnostic tasks in this embodiment.

[0156] A1 55.25±8.07 66.22±12.65 80.65±1.37 96.62±2.38 90.11±2.47 99.75±0.11 A2 68.60±13.70 73.25±13.77 90.30±2.03 61.38±2.88 86.08±4.35 98.75±1.20 A3 83.46±13.97 88.36±8.40 89.65±3.75 77.83±5.39 86.25±3.65 98.80±0.55 A4 64.11±9.43 79.51±13.92 78.85±5.39 96.38±2.86 87.23±3.26 98.77±0.57 A5 66.11±14.26 82.25±10.27 75.50±14.10 86.50±10.83 86.00±2.40 97.03±2.66 A6 75.39±11.46 82.15±12.53 86.25±5.55 80.46±13.65 93.73±6.13 99.75±0.10 A7 66.90±12.65 82.60±8.64 83.20±4.32 92.72±6.70 86.38±3.26 99.25±0.20 mean 68.55 79.19 83.49 84.56 87.97 99.15

[0157] Table 2

[0158] As shown in Table 2, single-modal methods (LSS, WDCNN) rely on only a single sensor, making them highly susceptible to interference from changes in operating conditions, resulting in low overall diagnostic accuracy. Other multimodal fusion methods (DCAE, M2FN, and MIFN), while mitigating the limitations of single-modal methods by utilizing complementary multimodal information, still exhibit unstable performance under different mixed operating conditions, with significant differences in diagnostic performance across tasks, and average accuracy ranging only from 83.49% to 87.97%. In contrast, this invention maintains extremely high diagnostic accuracy across all seven tasks, with an average accuracy of 99.15%, significantly outperforming other comparative methods. This fully demonstrates the strong diagnostic performance of this invention under complex operating conditions.

[0159] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A bearing fault diagnosis method based on uncertainty quantification, characterized in that, Includes the following steps: S1: Collect bearing data under normal conditions and Multimodal signal samples under various fault conditions, each multimodal signal sample including two different modes of signal within the same time period. , , Indicates time, sample label Set to fault status. The training sample set is constructed. S2: Construct a bearing fault diagnosis model, including a first feature extraction module, a second feature extraction module, and a prediction module. in: The first feature extraction module is used to extract the first mode signal from the input multimodal signal. Feature extraction is performed to obtain the first modality features. And send it to the prediction module; The second feature extraction module is used to extract the second mode signal from the input multimodal signal. Feature extraction is performed to obtain the second modality features. And send it to the prediction module; The prediction module is used to predict based on two modal features. The bearing is predicted to belong to The probability of each fault state; the prediction module includes a first evidence vector generation module, a second evidence vector generation module, a Dirichlet sampling module, an adaptive weight allocation module, a correction module, a basic probability allocation fusion module, and a probability generation module, wherein: The first evidence vector generation module is used to generate evidence vectors based on the first modality features. Generate a non-negative first evidence vector And send it to the Dirichlet sampling module: ， in, Indicates the first mode in the fault category Evidence value, , This represents the weights and biases corresponding to the first mode. Represents a smooth nonlinear activation function; The second evidence vector generation module is used to generate evidence based on the second modality features. Generate a non-negative second evidence vector And send it to the Dirichlet sampling module: ， in, Indicates the second mode in the fault category Evidence value, , This represents the weights and biases corresponding to the second mode; Dirichlet sampling module is based on two evidence vectors The basic probability assignments for the DS evidence theory are generated by sampling and mapping using the Dirichlet distribution. The specific method is as follows: Based on each evidence vector Calculate the total intensity of the distribution : ， in, Indicates the first In the mode, the th Concentration parameters of the Dirichlet distribution of the class; Then, the basic probability distribution of a single-element focal element is calculated using the following formula. Basic probability allocation of the focal element of the entire set Basic probability allocation of compound focal elements : ， in, Indicates the first Fault-like states, Represents a set of fault states. Indicates the first The modality pair of the first Determinism of class-specific faults, Indicates the first Overall cognitive uncertainty across modalities; Then, the overall cognitive uncertainty The data is sent to the adaptive modal weighting module to assign the basic probability of the single-element focal element. Basic probability allocation of the focal element of the whole set Send to the correction module; The adaptive modal weighting module is used to adjust the overall cognitive uncertainty. The weights of the two modes are calculated. And send it to the correction module, weight The calculation formula is as follows: ， in, Indicates the adjustment parameter. Indicates the compensation coefficient; The correction module is used to adjust the weights. Basic probability assignment of a single-element focal element Basic probability allocation of the focal element of the whole set After making corrections, we obtain the corrected basic probability distribution. , And send it to the basic probability allocation fusion module; the corrected formula is as follows: ； The basic probability assignment fusion module is used to assign the basic probability of the modified single-element focal element using the DS orthogonal sum rule. Basic probability allocation of the focal element of the whole set By performing fusion, we obtain the basic probability distribution of the fusion of single-element focal elements and universal focal elements. , And send it to the probability vector generation module; the fusion formula is as follows: ， ， in, The conflict coefficient is defined as follows: ， The probability generation module allocates basic probabilities based on the fusion of single-element focal elements and universal focal elements. , Obtain the probability that the final bearing belongs to each fault state. : ； S3: The bearing fault diagnosis model is trained using the training sample set to obtain a trained bearing fault diagnosis model; S4: When bearing fault detection is required, two different modal signals within the same time period are collected in real time and input into the trained bearing fault diagnosis model to obtain the fault diagnosis result.

2. The bearing fault diagnosis method according to claim 1, characterized in that, The loss function for training the bearing fault diagnosis model in step S3 The calculation formula is as follows: ， in, Indicates the weighting coefficient. Indicates the category of risk items, This indicates evidence calibration items and classification risk items. The calculation formula is: ， in, This indicates that the input multimodal signal obtained based on the real label belongs to the first... The probability of a fault state. This indicates that the input multimodal signal predicted by the model belongs to the . The probability of a fault state. This represents the total concentration parameter of the Dirichlet distribution. ,in Indicates the first The fusion concentration parameter of the Dirichlet distribution is calculated using the following formula: ； Evidence calibration item The calculation formula is: ， in, This represents the corrected Dirichlet distribution concentration parameter, if the actual fault state of the input multimodal signal is... ,but ,otherwise , Represents the Gamma function. This represents the Digamma function.

3. The bearing fault diagnosis method according to claim 2, characterized in that, The weighting coefficient , Indicates the number of iterations. Indicates the annealing cycle.

4. The bearing fault diagnosis method according to claim 1, characterized in that, The feature extraction module includes an ensemble empirical mode decomposition module, a physical difference quantization module, a mode weight generation module, a mode feature extraction module, and a weighted fusion module, wherein: The ensemble empirical mode decomposition module is used to analyze modal signals. IMF components are obtained by performing ensemble empirical module decomposition. And send it to the physical difference quantization module and the modal feature extraction module. , Indicates the number of components in the set's empirical modes; The physical difference quantification module is used to quantify each IMF component. Physical difference quantization is performed on the normal reference IMF component of the corresponding mode in the normal state multimodal signal sample to obtain the physical difference quantization vector of each IMF component in that mode. And send it to the modal weight generation module; The modal weight generation module is used to quantize vectors based on physical differences. Generate corresponding weights And send it to the weighted fusion module; The modal feature extraction module is used to extract features from each IMF component. Extract the corresponding features And send it to the weighted fusion module; The weighted fusion module is used to calculate the weights. Features of each component in each mode Weighted fusion is performed to obtain the final features of each modality. : 。 5. The bearing fault diagnosis method according to claim 4, characterized in that, The physical difference quantization vector ,in Indicates the first The first mode The KL divergence between each IMF component and its corresponding normal reference IMF component is calculated using the following formula: ， in, , These represent the first and second modal signals in the input multimodal signal, respectively. The first mode Mean and standard deviation of each IMF component , They represent the statistical results obtained from the normal signal samples in the training sample set. The first mode Mean and standard deviation of each IMF component; Indicates the first The first mode The kurtosis difference between an IMF component and its corresponding normal reference IMF component is calculated using the following formula: ， in, Indicates the first of the input multimodal signals The first mode Kurtosis of each IMF component The first normal signal sample obtained from the training sample set represents the... The first mode The kurtosis of each IMF component.

6. The bearing fault diagnosis method according to claim 5, characterized in that, The kurtosis The following formula is used for calculation: ， in, Indicates the length of the signal. This indicates the prevention of the minimum value where the denominator is 0.

7. The bearing fault diagnosis method according to claim 4, characterized in that, The modality weight generation module includes a multilayer perceptron and a Softmax layer, wherein: Multilayer perceptrons are used to quantize physical difference vectors. Generate score values ​​by performing nonlinear mapping And send it to the Softmax layer; The Softmax layer is used to apply the Softmax function to the score values. Normalization is performed to obtain the weights of the corresponding IMF components. .

8. The bearing fault diagnosis method according to claim 4, characterized in that, The specific method for feature extraction by the modality feature extraction module is as follows: For each IMF component Normalization is performed to obtain the normalized components. Then, a 1D-CNN feature extractor is used to extract the normalized components. Features were extracted .

9. The bearing fault diagnosis method according to claim 4, characterized in that, The loss function for training the bearing fault diagnosis model in step S3 The calculation formula is as follows: ， in, The loss function based on global evidence calibration is calculated as follows: ， in, Indicates the weighting coefficient. Indicates the category of risk items, This indicates evidence calibration items and classification risk items. The calculation formula is: ， in, This indicates that the input multimodal signal obtained based on the real label belongs to the first... The probability of a fault state. This indicates that the input multimodal signal predicted by the model belongs to the . The probability of a fault state. This represents the total concentration parameter of the Dirichlet distribution. ,in Indicates the first The fusion concentration parameter of the Dirichlet distribution is calculated using the following formula: ； Evidence calibration item The calculation formula is: ， in, This represents the corrected Dirichlet distribution concentration parameter, if the actual fault state of the input multimodal signal is... ,but ,otherwise , Represents the Gamma function. Represents the Digamma function; The local weight regularization term is represented by the following formula: ， in, , Representing physical alignment items With weight regularization term Weights; Physical alignment terms The calculation formula is: ， in, Indicates the first IMF weight vectors for each modality Indicates the first A vector of physical differences between each modality. ; Weight regularization term The calculation formula is: 。 10. The bearing fault diagnosis method according to claim 9, characterized in that, The weighting coefficient , Indicates the number of iterations. Indicates the annealing cycle.

Images