Rotary part fault diagnosis method based on cooperative training

Abstract

The invention discloses a rotating part fault diagnosis method based on cooperative training, and belongs to the technical field of specific calculation models, and the method comprises the steps: obtaining a first fault type group and a second fault type group according to fault types, and obtaining a first meta-training set and a second meta-training set in combination with a model training data set; based on preset model parameters, according to the first meta-training set, auxiliary binary classification loss is obtained through an auxiliary binary classifier, multi-classifier loss is obtained through a cross entropy function, and overall loss is obtained through combination; according to the overall loss, updating the model parameters, combining with the second meta-training set to obtain the overall loss, and according to the overall loss obtained by the first meta-training set and the second meta-training set, obtaining a target function for training the rotating part fault diagnosis model. Through the meta-training set and the collaborative gradient strategy, the model can adapt to unknown working conditions without target working condition data, unknown interference is suppressed in combination with the auxiliary classifier, the generalization ability is remarkably improved, and the data cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of specific computing models, and particularly relates to a fault diagnosis method for rotating components based on co-training. Background Art

[0002] In the operation and maintenance of industrial equipment, rotating equipment (such as centrifugal pump units) generally serves as the core power equipment in oil and gas transportation, chemical processes, and municipal water supply systems. Fault diagnosis of rotating components (such as bearings, impellers, couplings, etc.) is a key link to ensure the safe operation of rotating equipment.

[0003] With the development of the intelligence of industrial equipment, fault diagnosis methods based on deep learning have gradually matured. Existing methods mainly rely on deep learning models to identify fault types by training classifiers through vibration data. However, existing methods rely on training data with known working conditions and known fault types and cannot effectively handle newly emerging fault types under unknown working conditions (such as sudden load changes and speed fluctuations). Patent CN118643324A improves the cross-condition generalization ability through domain adaptation, but it still requires target working condition data for training and cannot identify unknown fault types. Patent CN118643346B introduces uncertainty measurement to improve the ability to identify unknown faults, but it also requires target working condition data for training and does not solve the problem of unknown fault diagnosis under unknown working conditions.

[0004] In the actual operation of rotating equipment, frequent speed changes (such as dynamic adjustment of pump groups from 2950 to 3500 rpm), sudden load changes (such as sudden increase in the viscosity of chemical media), etc., combined with changes in the dynamic characteristics of the shafting caused by equipment aging (such as coupling of abnormal shaft deflection and early bearing faults), often lead to the sudden emergence of new unknown fault modes (such as adhesive wear of sliding bearings, micro-crack propagation of thrust bearings, and low-temperature shrinkage leakage of mechanical seals). Therefore, fault diagnosis of rotating components under unknown working conditions and unknown faults becomes more important. Although Patent CN117786489A improves generalization through co-training of multi-sensor data, it does not solve the problem of blurred classification boundaries caused by the combined action of working condition offset and fault type offset, which will cause a certain degree of misclassification in the identification of unknown faults under unknown working conditions.

[0005] To solve the above problems, there is an urgent need for a fault diagnosis method that does not require target working condition data for training and simultaneously has cross-condition generalization ability and unknown fault identification ability. Summary of the Invention

[0006] The present invention provides a fault diagnosis method for rotating components based on co-training to solve the problem of joint identification of unknown working conditions and unknown fault types of rotating components, improve the diagnostic accuracy and reliability, and realize the intelligent operation and maintenance of rotating equipment under complex working conditions.

[0007] The technical solution adopted by the present invention is as follows: A fault diagnosis method for rotating components based on co-training, comprising: Obtaining a first fault type group and a second fault type group according to the fault type, and combining the collected model training data set to obtain a first meta-training set and a second meta-training set, wherein the model training data set includes vibration data of the rotating component under multiple working conditions and multiple fault types; Based on the preset model parameters, according to the first meta-training set, obtaining an auxiliary binary classification loss through an auxiliary binary classifier, obtaining a multi-classifier loss through a cross-entropy function, and combining to obtain an overall loss; According to the overall loss, updating the model parameters, combining the second meta-training set to obtain an overall loss, and obtaining an objective function according to the overall losses obtained from the first meta-training set and the second meta-training set for training a fault diagnosis model of rotating components.

[0008] Obtaining a first fault type group and a second fault type group according to the fault type, specifically: Obtaining a first fault type group and a second fault type group through relevance analysis according to the characteristics of the fault type; and / or obtaining a first fault type group and a second fault type group according to the vibration intensity corresponding to the fault type.

[0009] The first meta-training set and the second meta-training set are specifically: Both the first meta-training set and the second meta-training set include vibration data under multiple working conditions; The vibration data of the first meta-training set under at least one working condition belongs to one of the first fault type group and the second fault type group, and the vibration data under at least another working condition belongs to the other of the first fault type group and the second fault type group; The fault type group to which the vibration data of the second meta-training set belongs is different from the fault type group to which the vibration data corresponding to the first meta-training set under the corresponding working condition belongs.

[0010] The overall loss is specifically: Combining the multi-classifier loss and the auxiliary binary classification loss to obtain the overall loss;

[0011] Wherein, represents the overall loss, represents the multi-classifier loss, represents the auxiliary binary classification loss.

[0012] Updating the model parameters according to the overall loss, specifically: Update the model parameters according to the gradient of the overall loss obtained from the first meta-training set under the preset model parameters, in combination with the set learning rate;

[0013] Wherein, and represent the vibration data of the first meta-training set, represents that the vibration data corresponding to the first working condition in the first meta-training set belongs to the first fault type group, represents that the vibration data corresponding to the second working condition in the first meta-training set belongs to the second fault type group; and respectively represent and the gradients of the overall loss; represents the set learning rate; represents the preset model parameters; represents the updated model parameters.

[0014] Obtain an objective function for rotating component fault diagnosis according to the overall loss obtained from the first meta-training set and the second meta-training set, specifically:

[0015] Wherein, and represent the vibration data of the second meta-training set, represents that the vibration data corresponding to the first working condition in the second meta-training set belongs to the second fault type group, represents that the vibration data corresponding to the second working condition in the second meta-training set belongs to the first fault type group; and respectively represent and the overall loss under the preset model parameters; and respectively represent and the overall loss under the updated model parameters.

[0016] The method for rotating component fault diagnosis based on co-training further includes: According to the collected vibration data, through preprocessing, obtain a model training data set; wherein, the preprocessing at least includes data anomaly detection, sliding window segmentation, and time-frequency feature conversion processing, and the data anomaly detection is used to identify and eliminate abnormal data.

[0017] The sliding window segmentation is specifically: Based on the vibration data after anomaly elimination, multiple vibration data segments are obtained through a non-overlapping sliding window segmentation method; wherein the window step size is determined according to the vibration data acquisition length, acquisition frequency, and data vibration period.

[0018] The time-frequency feature conversion process is specifically as follows: Based on the vibration data after sliding window segmentation processing, the negative frequency part in the signal is removed through an instantaneous feature extraction method; Based on the processed vibration data, the instantaneous power spectral density is obtained through a time-frequency analysis method; Among them, the frequency resolution of the time-frequency analysis method is determined according to the highest frequency of the acquired vibration data, and the time resolution is determined according to the minimum bandwidth of the acquired vibration data.

[0019] The rotating component fault diagnosis model is specifically as follows: Based on the vibration data of the rotating component acquired in real time, a confidence score is obtained through the auxiliary binary classifier in the rotating component fault diagnosis model, and the known fault type and unknown fault type are judged through the confidence score; For the known fault type, the probability distribution of the known fault type is obtained through the multi-classifier in the rotating component fault diagnosis model to diagnose the rotating component fault.

[0020] Due to the adoption of the above technical solutions, the beneficial effects achieved by the present invention are as follows: 1. In the present invention, the first fault type group and the second fault type group are obtained according to the fault type, and combined with the acquired model training data set, the first meta-training set and the second meta-training set are obtained, wherein the model training data set includes the vibration data of the rotating component under multiple working conditions and multiple fault types. Using the collaborative gradient update strategy of the working condition and the fault type, the model can still effectively capture the feature commonalities under unknown working conditions after being trained with known working condition data. That is to say, the model can adapt to unknown working conditions without relying on target working condition data, solving the problem of feature space mismatch caused by working condition offset in traditional methods.

[0021] In addition, the auxiliary binary classification loss is obtained through the auxiliary binary classifier, the multi-classifier loss is obtained through the cross-entropy function, and the overall loss is combined. The auxiliary binary classifier is introduced to generate a clear classification boundary by distinguishing between known and unknown fault types (judgment by confidence threshold). Combining the cross-entropy loss and the auxiliary binary classification loss to construct an overall loss function to suppress the interference of unknown faults on known classifications. The model can identify unknown fault types that do not appear in the training data, avoiding the misjudgment problem of traditional closed-set classifiers.

[0022] The objective function is obtained based on the overall loss of the first meta-training set and the second meta-training set for training the rotating component fault diagnosis model. A gradient matching mechanism is adopted to ensure that the parameter update directions of the working condition level and the fault type level are consistent, and to avoid adversarial interference in cross-working condition training. The model's ability to represent fine-grained features is optimized, and the feature confusion between known and unknown faults is reduced.

[0023] Therefore, the model only needs to be trained with multi-fault type data under known working conditions, without covering all possible working condition combinations, which greatly reduces the data acquisition cost. Description of the Drawings

[0024] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 It is a schematic flowchart of the method for diagnosing faults of rotating components based on co-training according to an embodiment of the present invention; Figure 2 It is a logical architecture diagram of the method for diagnosing faults of rotating components based on co-training according to an embodiment of the present invention. Detailed Embodiments

[0025] In order to more clearly illustrate the overall concept of the present invention, the following will be described in detail by way of examples in conjunction with the drawings of the specification.

[0026] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0027] As Figure 1 shown, a method for diagnosing faults of rotating components based on co-training includes: S100: Obtain a first fault type group and a second fault type group according to the fault type, and combine the collected model training data set to obtain a first meta-training set and a second meta-training set, where the model training data set includes vibration data of the rotating component under multiple working conditions and multiple fault types.

[0028] The core objective of this step is to construct a training dataset that can adapt to unknown working conditions and unknown fault types. Through the collaborative design of fault type grouping and working conditions, the cross-condition generalization ability of subsequent model training is improved to ensure that the model can accurately identify known fault types under unknown working conditions; the unknown fault identification ability, by suppressing the interference of unknown faults on known classifications through grouping strategies to generate clear classification boundaries; the data efficient utilization ability, where the model can be trained only with known working condition data, reducing the dependence on target working condition data.

[0029] It can be understood that the model training dataset contains vibration data of the rotating component under multiple working conditions and multiple fault types, providing the model with the joint feature distribution of working conditions and fault types to support cross-domain generalization. The meta-training set contains a training set of known fault type data under multiple working conditions for initial model parameter update. The gradient calculation is used to guide the optimization direction of model parameters to ensure the collaborative feature learning of working conditions and fault types.

[0030] According to the fault type grouping, in the present invention, based on the feature similarity of fault types (such as the vibration spectrum difference between inner ring damage and outer ring damage of a bearing), the fault types are divided into two groups (such as the first group being rolling element faults and the second group being lubrication faults) through statistical methods (such as Pearson correlation coefficient).

[0031] Similarly, the fault types can be divided into a high-risk group and a low-risk group according to the vibration energy intensity corresponding to the fault types (such as high-amplitude faults and low-amplitude faults). In a specific embodiment, in the bearing fault of a centrifugal pump, inner ring damage (high-frequency impact characteristics) and grease contamination (low-frequency modulation characteristics) may be grouped differently. This step reduces the feature confusion of fault types within the group, providing a distinguishable feature space for subsequent gradient matching.

[0032] According to the fault type grouping, the model training dataset is divided into a first meta-training set and a second meta-training set. Both the first meta-training set and the second meta-training set contain vibration data under multiple working conditions, and for the same working condition, the vibration data corresponding to the first meta-training set and the second meta-training set belong to different fault type groups. Ensure that there are differences in both working conditions and fault types between the first meta-training set and the second meta-training set to simulate unknown fault scenarios under unknown working conditions.

[0033] Through the collaborative training of the first meta-training set and the second meta-training set, when the model updates its parameters, it needs to adapt to both operating condition shifts and fault type shifts simultaneously, significantly enhancing the cross-domain generalization ability. It should be noted that both the first meta-training set and the second meta-training set need to contain multiple operating conditions (such as rotational speeds of 735 RPM, 1102.5 RPM, and 1470 RPM) to ensure that the model learns the operating condition invariance features. The fault type groups to which the vibration data of the first meta-training set and the second meta-training set belong are different under the same operating condition, avoiding the overlap of feature distributions. This enables the model to focus on the joint features of the operating condition and the fault type, reducing the misjudgment of known faults under unknown operating conditions. The first meta-training set and the second meta-training set constructed in this step provide a data basis for the subsequent model parameter update, ensuring that the model synchronously optimizes the feature representations of the operating condition and the fault type when updating the parameters.

[0034] Generally speaking, in this step, through fault type grouping and meta-training set design, a training data framework that supports generalization across operating conditions and fault types is constructed, providing a key data basis for the collaborative training of the model, solving the bottleneck problem that traditional methods cannot identify unknown faults under unknown operating conditions, and significantly enhancing the practicality and reliability of rotating component fault diagnosis.

[0035] S200: Based on the preset model parameters, according to the first meta-training set, obtain the auxiliary binary classification loss through the auxiliary binary classifier, obtain the multi-classifier loss through the cross-entropy function, and combine them to obtain the overall loss.

[0036] The core objective of this step is to improve the classification accuracy of known faults and the recognition ability of unknown faults under unknown operating conditions by jointly optimizing the multi-classification task and the auxiliary binary classification task. Specifically, the multi-classification task optimizes the classification accuracy of known fault types through the cross-entropy loss. The auxiliary binary classification task uses the auxiliary binary classifier to distinguish between known faults and unknown faults, generating a clear classification boundary and suppressing the interference of unknown faults on known classifications. The overall loss function combines the gradient information of the two losses to synchronously optimize the model parameters, ensuring the consistency of the feature representations of the operating condition and the fault type.

[0037] For the binary classifier, input the vibration data of the first meta-training set, and output a binary judgment (known fault or unknown fault) through the auxiliary binary classifier. The auxiliary binary classification loss function is as follows:

[0038] where, for the input sample , the probability that the model outputs the correct class , for all non-correct classes , calculate their predicted probabilities , and take the one corresponding to the minimum value among them , combining the two items to generate an auxiliary binary classification loss.

[0039] By maximizing the prediction probability of the correct class to reduce missed judgments. Minimizing the confidence of the model in the most confident incorrect class to avoid misjudgments (such as misjudging grease contamination as bearing wear). Through the dual-objective optimization, the model forms a clearer decision boundary between known / unknown faults or critical classes.

[0040] For the multi-classifier, input the vibration data of the first training set, and output the probability distribution of known fault types (such as inner ring damage, outer ring damage of the bearing, etc.) through the multi-classifier. The loss of the multi-classifier is as follows,

[0041] wherein, the model outputs the prediction probabilities for all classes . According to the true label (one-hot encoding), calculate the difference between the model prediction and the true distribution.

[0042] In this step, by minimizing the cross-entropy, the prediction probability distribution of the model for known fault types is closer to the true label, improving the classification accuracy of known faults. The overall loss is obtained by combining the auxiliary binary classification loss and the multi-classifier loss. Compared with using only the cross-entropy loss, the overall loss function improves the classification accuracy of known faults. Moreover, by jointly optimizing the two types of tasks, the recognition ability of known and unknown faults is improved synchronously. The overall loss in this step provides a direction for the parameter optimization of the subsequent second training set, and finally forms an objective function for the hierarchical coordination of working conditions and fault types.

[0043] Generally speaking, in this step, through the collaborative optimization of the auxiliary binary classification loss and the multi-classifier loss, the problems of fuzzy classification boundary and overfitting of traditional methods under complex working conditions are solved, providing an efficient and robust model training method for the fault diagnosis of rotating components.

[0044] S300: Update the model parameters according to the overall loss, combine the second training set to obtain the overall loss, and obtain the objective function according to the overall losses obtained from the first training set and the second training set for the training of the rotating component fault diagnosis model.

[0045] The core objective of this step is to optimize the model parameters through a collaborative training strategy, enabling the model to have stronger generalization ability under unknown working conditions and unknown fault types. Specifically, by using the loss functions of the first meta-training set and the second meta-training set, the model parameters are synchronously optimized to achieve cross-condition parameter update. The losses of the two meta-training sets are jointly formed into an objective function to ensure that the model can adapt to both working condition shift and fault type shift. Here, the objective function refers to the loss function ultimately used for model training, which synthesizes the task requirements of multiple working conditions and multiple fault types, and is used to guide the model to learn the invariant features of working conditions and the distinguishing features of fault types.

[0046] Based on the first meta-training set, the overall loss is calculated through the obtained overall loss function (including the auxiliary binary classification loss and the cross-entropy loss). The parameter gradients are calculated according to the overall loss to update the model parameters. It should be noted that the regularization term (such as maximizing the dot product of gradients) is used to ensure that the parameter update directions of the working condition and the fault type are consistent.

[0047] In this step, the features learned by the model under known working conditions (such as the bearing vibration spectrum) are generalized to unknown working conditions (such as the scene of sudden speed change), and the known fault classification accuracy is obtained. Moreover, the gradient direction consistency constraint reduces adversarial updates and improves the model convergence speed.

[0048] The updated model parameters are applied to the second meta-training set (including different fault types under the same working condition as the first meta-training set), and its overall loss is calculated. According to the overall losses of the first meta-training set and the second meta-training set, the final objective function is formed. In this step, the different fault data in the second meta-training set force the model to learn general features (such as vibration energy distribution), improving the unknown fault recognition accuracy and the cross-domain generalization ability. The losses of the two meta-training sets are weighted and averaged to form the final objective function for guiding model training. The model can be directly deployed under unknown working conditions without additional data, reducing the operation and maintenance costs. This step solves the generalization ability bottleneck of traditional methods under unknown working conditions and unknown fault types through the collaborative training strategy (joint optimization of the first meta-training set and the second meta-training set) and the objective function.

[0049] As a preferred implementation manner of the present invention, the first fault type group and the second fault type group are obtained according to the fault type, specifically: According to the characteristics of the fault type, the first fault type group and the second fault type group are obtained through correlation analysis; and / or, according to the vibration intensity corresponding to the fault type, the first fault type group and the second fault type group are obtained.

[0050] The core objective of this embodiment is to optimize the division of fault type groups through a scientific grouping strategy, enhancing the generalization ability and classification accuracy of the model for rotating components under multiple working conditions and multiple fault types. Specifically, through correlation analysis or vibration intensity grouping, ensure that the feature differences between fault types in different groups are significant, avoid misjudgment caused by feature similarity in the model, and reduce feature confusion. Through the grouping strategy, the model can be trained only with known working condition data, reducing the dependence on target working condition data and achieving efficient data utilization.

[0051] Example 1: Correlation analysis grouping method. Extract time-domain (such as root mean square value, kurtosis), frequency-domain (such as main frequency energy, frequency band power ratio), or time-frequency domain (such as transient characteristics of Wigner-Ville distribution) features from vibration data. Calculate the Pearson correlation coefficient or cosine similarity of the feature vectors of each pair of fault types. For example, the vibration spectrum correlation coefficient between inner ring damage and outer ring damage of a bearing may be relatively high (such as 0.85), while the correlation coefficient between grease contamination and rolling element failure is relatively low (such as 0.32).

[0052] Set a correlation coefficient threshold (such as 0.6), and classify fault types with correlation coefficients lower than the threshold into different groups. For example, classify high-correlation faults (such as inner ring damage, outer ring damage) into the first group, and low-correlation faults (such as grease contamination, shaft eccentricity) into different groups.

[0053] In this embodiment, the feature differences between fault types in different groups are significant, reducing feature confusion, enhancing the clarity of the classification boundary, and reducing the misjudgment rate.

[0054] Example 2: Vibration intensity grouping method. Calculate the vibration energy (such as RMS value), peak factor, or kurtosis of each type of fault. For example, the kurtosis value of bearing spalling fault is significantly higher than that of grease contamination (such as the kurtosis of spalling fault is 15 and that of grease contamination is 3.2). Set a vibration intensity threshold (such as the RMS threshold is 0.5g), classify high-amplitude faults (such as ball fracture) into the first group, and low-amplitude faults (such as slight wear) into the second group.

[0055] In this embodiment, by enhancing the energy sensitivity, the classification accuracy of the model for faults corresponding to different amplitudes is improved.

[0056] It should be noted that correlation analysis and vibration intensity grouping can be adopted simultaneously to generate multi-dimensional grouping results. For example: The first group is high-correlation and high-amplitude faults (such as inner ring damage, outer ring damage of the bearing); the second group is low-correlation and low-amplitude faults (such as grease contamination, shaft misalignment). It can be understood that in the present invention, the rationality of the grouping results can be verified through clustering analysis (such as K-means) to ensure that the feature distributions of fault types within the group are concentrated.

[0057] In a specific embodiment, there are five types of faults in a centrifugal pump bearing: inner ring damage, outer ring damage, ball fracture, grease contamination, and shaft eccentricity. By analyzing the relevance and calculating the spectral correlation coefficients of each fault, it is found that the correlation coefficient between inner ring damage and outer ring damage is 0.85, and that with grease contamination is 0.32. The first group (high correlation) is divided, including inner ring damage and outer ring damage; the second group includes the other three types.

[0058] Grouping according to vibration intensity, calculating the RMS value, ball fracture (0.8g), grease contamination (0.2g). The first group (high amplitude) is divided, including ball fracture; the second group includes the other four types.

[0059] Carry out combined grouping of relevance and vibration intensity. Finally, the first group includes inner ring damage, outer ring damage, and ball fracture (high correlation and high amplitude); the second group includes grease contamination and shaft eccentricity (low correlation and low amplitude).

[0060] This preferred embodiment uses a combined strategy of relevance analysis and vibration intensity grouping to scientifically divide the fault type groups, significantly improving the classification accuracy and robustness of the model under complex working conditions.

[0061] As a preferred embodiment of the present invention, the first meta-training set and the second meta-training set are specifically: Both the first meta-training set and the second meta-training set contain vibration data under multiple working conditions; The vibration data of the first meta-training set under at least one working condition belongs to one of the first fault type group and the second fault type group, and the vibration data under at least another working condition belongs to the other of the first fault type group and the second fault type group; The fault type group to which the vibration data of the second meta-training set belongs is different from that of the vibration data of the first meta-training set under the corresponding working condition.

[0062] The core objective of this preferred embodiment is to improve the cross-domain generalization ability of the model under unknown working conditions and unknown fault types through the carefully designed first meta-training set and second meta-training set. Specifically, the orthogonal design of working conditions and fault types ensures that the model simultaneously learns the working condition invariance features and fault type discrimination features during the training process. By alternately exposing the working condition data of different fault type groups, the dependence of the model on specific working conditions or fault types is reduced. The reverse fault type group design of the second meta-training set forces the model to distinguish known and unknown faults, improving the clarity of the classification boundary and enhancing the robustness to unknown faults.

[0063] It is understandable that the first meta-training set and the second meta-training set cover multiple working conditions, each containing multiple working conditions (such as rotational speeds of 735 RPM, 1102.5 RPM, and 1470 RPM). Through the alternating allocation of fault type groups, the vibration data of at least one working condition (such as working condition A) in the first meta-training set belongs to the first fault type group (such as bearing damage faults). The vibration data of at least another working condition (such as working condition B) belongs to the second fault type group (such as lubrication fault types). The fault type groups of the corresponding working conditions in the second meta-training set are opposite to those in the first meta-training set. The data of the second meta-training set for working condition A belongs to the second fault type group (such as lubrication faults), and the data of the second meta-training set for working condition B belongs to the first fault type group (such as bearing damage).

[0064] This step ensures that there is data for each working condition in the two meta-training sets, and the distribution of fault type groups is complementary. The model learns general features from the bearing damage data (first meta) of working condition A and the lubrication fault data (first meta) of working condition B, enabling it to accurately classify under unknown working condition C and enhancing its cross-working condition generalization ability. The reverse fault type groups in the second meta-training set force the model to distinguish unknown faults in order to identify unknown faults.

[0065] In a specific embodiment, taking the sample data sets of two working conditions and as an example, and are correspondingly divided into 、 and 、 according to the fault types. Among them, and represent different working conditions, and 1 and 2 represent the first fault type group and the second fault type group respectively.

[0066] That is to say, and have different fault types, and have different fault types, and have the same fault type and belong to the first fault type group, and have the same fault type and belong to the second fault type group. Then is used as the first meta-training set, is used as the second meta-test set.

[0067] Similarly, the present invention can also obtain a test set using the sample data sets under multiple working conditions. Taking the sample data sets under three working conditions as an example, for the sample data set , similarly is correspondingly divided into , . Then use as the first meta-training set, as the second meta-test set; or use as the first meta-training set, as the second meta-test set. The present invention does not limit this, and only needs to satisfy the settings of the first meta-training set and the second meta-training set in this embodiment.

[0068] Through the orthogonal design of working conditions and fault types, this embodiment constructs a first meta-training set and a second meta-test set that support cross-domain generalization, and solves the generalization ability bottleneck of traditional methods under unknown working conditions and unknown fault types.

[0069] As a preferred embodiment of the present invention, the overall loss is specifically: combining the multi-classifier loss and the auxiliary binary classification loss to obtain the overall loss;

[0070] Among them, represents the overall loss, represents the multi-classifier loss, represents the auxiliary binary classification loss.

[0071] The core objective of this embodiment is to improve the classification accuracy and robustness of the rotating component fault diagnosis model by jointly optimizing the multi-classification task and the auxiliary binary classification task. Specifically, by maximizing the prediction confidence of the model for known fault types through the multi-classification cross-entropy loss, the accurate classification of known faults is improved. By forcing the model to distinguish between known and unknown fault types through the auxiliary binary classification loss and reducing misjudgment, the robustness of unknown faults is enhanced.

[0072] The overall loss is . By directly adding, the two types of tasks are jointly optimized to ensure the balance of the global loss. Through the joint optimization of the multi-classification cross-entropy loss and the auxiliary binary classification loss, this embodiment solves the problem of insufficient generalization ability of traditional fault diagnosis models under unknown working conditions and unknown fault types.

[0073] As a preferred embodiment of this embodiment, according to the overall loss, update the model parameters, specifically: according to the gradient of the overall loss obtained by the first meta-training set under the preset model parameters, combined with the set learning rate, update the model parameters;

[0074] Among them, and represent the vibration data of the first meta-training set, It is indicated that the vibration data corresponding to the first working condition in the first meta-training set belongs to the first fault type group. It is indicated that the vibration data corresponding to the second working condition in the first meta-training set belongs to the second fault type group. and respectively represent and the gradients of the overall loss. It represents the set learning rate. It represents the preset model parameters. It represents the updated model parameters.

[0075] The core objective of this embodiment is to improve the robustness and cross-domain generalization ability of model parameters by jointly optimizing the gradient information of multiple working conditions and multiple fault type groups. Specifically, the collaborative optimization of the working condition and the fault type, by integrating the gradient information of different fault type groups under different working conditions, ensures that the model parameter update direction adapts to both the working condition change and the fault type difference simultaneously. By integrating the gradient information of different fault type groups under different working conditions, it ensures that the model parameter update direction adapts to both the working condition change and the fault type difference simultaneously, realizing the collaborative optimization of the working condition and the fault type. Through the way of gradient summation, it avoids the conflict of gradient directions of different working conditions or fault type groups, reduces adversarial updates, and realizes the gradient direction consistency constraint.

[0076] The parameter update formula is , where, for the learning rate , the initial value is set to 0.001 and is dynamically adjusted according to the performance of the second meta-training set (such as adopting the cosine annealing strategy).

[0077] This embodiment solves the problem of insufficient generalization ability of traditional models under complex working conditions and unknown fault types through the joint optimization of gradients of multiple working conditions and multiple fault type groups. In addition, it improves the model convergence speed through the gradient direction consistency constraint.

[0078] Specifically, an objective function for the fault diagnosis of rotating components is obtained according to the overall loss obtained from the first meta-training set and the second meta-training set, specifically:

[0079] where, and represent the vibration data of the second meta-training set, It is indicated that the vibration data corresponding to the first working condition in the second meta-training set belongs to the second fault type group, It is indicated that the vibration data corresponding to the second working condition in the second meta-training set belongs to the first fault type group; and respectively represent and The overall loss under the preset model parameters; and respectively represent and the overall loss under the updated model parameters.

[0080] The core objective of this embodiment is to improve the cross - domain generalization ability of the rotating component fault diagnosis model under unknown working conditions and unknown fault types by jointly optimizing the losses of the first meta - training set and the second meta - training set. Specifically, by jointly optimizing the losses of the two meta - training sets, it is ensured that the model parameters can simultaneously adapt to the characteristic differences of different fault type groups under different working conditions, realizing the collaborative optimization of multiple working conditions and multiple fault types. Through the design of the reverse fault type group of the second meta - training set, the model's dependence on specific working conditions or fault types is suppressed, the risk of overfitting is reduced, and the adversarial loss is inhibited.

[0081] For the first meta - training set, the obtained overall loss is . For the second meta - training set, the obtained overall loss is . The model learns the working condition invariance characteristics during the alternating exposure of the first meta - training set (the first fault type group → the second fault type group) and the second meta - training set (the second fault type group → the first fault type group). The reverse fault type group of the second meta - training set (such as lubrication fault at 735 RPM working condition) forces the model to distinguish unknown faults and suppresses overfitting.

[0082] Construct the objective function . The goal is to find the model parameters that minimize the sum of the losses of the first meta - training set and the second meta - training set. The iterative optimization process of the model parameters is as follows: use the loss of the first meta - training set to update the parameters to , use the updated parameters to calculate the loss of the second meta - training set, and jointly optimize.

[0083] It can be understood that for the above objective function formula, the first - order Taylor formula is used for the following expansion to obtain

[0084] where the second term is the sum of each gradient dot product. Maximizing the gradient dot product can regularize this training process, making the update directions of different tasks match and realizing the synchronous matching of working conditions and fault types. This step solves the problem of insufficient generalization ability of traditional models under unknown working conditions and unknown fault types by jointly optimizing the losses of the first meta - training set and the second meta - training set.

[0085] As a preferred implementation manner of the present invention, the rotating component fault diagnosis method based on collaborative training further includes: Based on the collected vibration data, a model training data set is obtained through preprocessing; wherein, the preprocessing at least includes data anomaly detection, sliding window segmentation, and time-frequency feature transformation processing, and the data anomaly detection is used to identify and eliminate abnormal data.

[0086] The core objective of this embodiment is to improve the quality and applicability of vibration data through data preprocessing, and provide reliable, structured, and feature-rich input data for subsequent model training. Specifically, noise or abnormal data is eliminated through anomaly detection to reduce noise interference in model training and ensure data quality. The continuous vibration signal is transformed into time series segments of a fixed length through sliding window segmentation to adapt to the model input format and achieve the structuring of time series data. The time-frequency joint features of the vibration signal are extracted through time-frequency feature transformation to enhance the ability to distinguish fault types, thereby enhancing features and improving separability.

[0087] For data anomaly detection, the method at least includes any one of statistical methods and machine learning methods. The statistical method is based on the mean and standard deviation threshold detection (such as ). The machine learning method is to use Isolation Forest or LOF (Local Outlier Factor) to detect outliers.

[0088] For abnormal data, the abnormal data points are directly deleted or repaired through interpolation methods (such as linear interpolation). In addition, the anomaly detection threshold is dynamically adjusted according to operating condition parameters (such as rotational speed, load) (for example, the noise threshold is increased by 20% under high rotational speed conditions). This embodiment reduces noise interference, reduces the proportion of abnormal data, improves the stability of model training, and reduces the false positive rate.

[0089] As an embodiment under this embodiment, the sliding window segmentation is specifically as follows: Based on the vibration data after anomaly elimination, multiple vibration data segments are obtained through a non-overlapping sliding window segmentation method; wherein the window step size is determined according to the vibration data acquisition length, acquisition frequency, and data vibration period.

[0090] The core objective of this embodiment is to transform continuous vibration data into independent segments of a fixed length through a non-overlapping sliding window segmentation method to meet the requirements of the model input format and improve the efficiency of feature extraction. Specifically, the continuous vibration signal is segmented into time series segments of a fixed length to adapt to the input format of the model (such as CNN or RNN) through data structuring. The non-overlapping window reduces data redundancy and computational complexity, is applicable to real-time or resource-constrained scenarios, and optimizes computational efficiency. By matching the window step size with the vibration period, it is ensured that each segment contains a complete fault feature period, improving classification accuracy.

[0091] To determine the window parameters, the spectrum of the vibration signal is analyzed through Fourier transform or wavelet transform to determine the main fault characteristic frequencies (such as the impact pulse frequency of bearing faults). The vibration period . Among them, is the peak frequency.

[0092] According to the acquisition frequency and the vibration period T , the window length is set. For example, if , , then (250 sampling points). The non-overlapping window step size is used to ensure that the segments are completely independent.

[0093] In this embodiment, each window contains at least one complete vibration period, improving the capture rate of the vibration pulses of the fault. Compared with the overlapping window, the data volume is reduced, and the inference speed and calculation efficiency are improved. This embodiment solves the problems of computational redundancy and insufficient working condition adaptability brought by the traditional overlapping window through the non-overlapping sliding window segmentation method.

[0094] As another embodiment under this implementation manner, the time-frequency feature conversion processing is specifically as follows: According to the vibration data after the sliding window segmentation processing, the negative frequency part in the signal is removed through the instantaneous feature extraction method; According to the processed vibration data, through the time-frequency analysis method, the instantaneous power spectral density is obtained; Among them, the frequency resolution of the time-frequency analysis method is determined according to the highest frequency of the acquired vibration data, and the time resolution is determined according to the minimum bandwidth of the acquired vibration data.

[0095] The core objective of this embodiment is to generate a high-resolution and low-redundancy instantaneous power spectral density through time-frequency analysis and feature extraction, so as to enhance the time-frequency joint features of the vibration signal and improve the model's ability to distinguish fault types. Specifically, through the set frequency and time resolution, high-resolution time-frequency analysis is performed to capture the transient features of the vibration signal (such as the impact pulse of bearing damage). Remove the negative frequency components, suppress the negative frequency, reduce the signal redundancy, and reduce the computational complexity. Retain the instantaneous power spectral density of the positive frequency part, intuitively reflect the time-frequency distribution of the fault features, and enhance the feature interpretability.

[0096] In this step, the input real signal , with a length of . Through the instantaneous feature extraction method (such as Hilbert transform ), an analytic signal is generated. The spectrum of the analytic signal only contains positive frequency components.

[0097] It should be noted that before the Hilbert transform, low-pass filtering is performed on with a cut-off frequency of to ensure that the signal bandwidth satisfies the Nyquist sampling theorem. In this step, the negative frequency components of the analytic signal are completely removed, avoiding cross-term aliasing in the WVD calculation, improving signal fidelity, and reducing the sampling error of high-frequency components.

[0098] Based on the processed vibration data, through time-frequency analysis methods (such as the WVD method),

[0099] where represents the complex conjugate of the analytic signal. The value range of m is (only positive frequencies are retained). By restricting the range of , it is ensured that the calculation is only for positive frequency components, avoiding spectral folding when it exceeds N / 2. is the standard form of the discrete WVD, which is compatible with the positive frequency characteristics of the analytic signal.

[0100] In this step, through aliasing elimination, the time-frequency energy concentration during faults is improved, and cross-term interference is reduced. In addition, the positive frequency characteristics of the analytic signal reduce the computational amount and optimize the computational efficiency. Generally speaking, this embodiment solves problems such as noise interference, missing time-series patterns, and insufficient feature separability existing in the original vibration data through the joint preprocessing of data anomaly detection, sliding window segmentation, and time-frequency feature conversion.

[0101] As a preferred embodiment of the present invention, the fault diagnosis model for rotating components is specifically as follows: Based on the vibration data of the rotating components collected in real time, the confidence score is obtained through the auxiliary binary classifier in the rotating component fault diagnosis model, and the known fault types and unknown fault types are judged through the confidence score; For known fault types, the probability distribution of the known fault types is obtained through the multi-classifier in the rotating component fault diagnosis model to diagnose the faults of the rotating components.

[0102] The core objective of this embodiment is to achieve accurate diagnosis of rotating component faults through the collaborative work of the auxiliary binary classifier and the multi-classifier, while improving the model's detection ability for unknown fault types. Specifically, for known / unknown fault classification, the auxiliary binary classifier is used to distinguish known faults from unknown faults, reducing the risk of misdiagnosis. For the fine-grained diagnosis of known faults, the multi-classifier is used to predict the probability distribution of known fault types, improving the classification accuracy.

[0103] In this embodiment, a fault diagnosis model for rotating components that has been trained is applied to determine the fault type of rotating components based on real-time acquired vibration data. The vibration data of the rotating components (such as the vibration signals of bearings or gearboxes) is obtained through sensors. It can be understood that the real-time acquired vibration data also needs to be preprocessed to generate feature vectors. The preprocessed feature vectors are input into an auxiliary binary classifier to output a binary classification result (known fault / unknown fault) and a confidence score. If the confidence score , it is determined as a known fault; otherwise, it is an unknown fault. Samples with abnormal fluctuations in the confidence score (such as oscillating around 0.8) are marked as potential noise or marginal cases.

[0104] The samples determined as "known faults" by the auxiliary classifier are input into a multi-classifier for processing to output the probability distribution of known fault types (such as the probability of inner ring damage of the bearing being 0.85 and the probability of outer ring damage being 0.12). The category with the highest probability is selected as the final diagnosis result. If the highest probability < 0.7, secondary verification is triggered (such as combining historical data or an expert system).

[0105] For example Figure 2 shown, in a specific embodiment, an induction motor drives a centrifugal pump, fresh water is taken from a water storage tank and conveyed through a pipeline system, and valves are installed before and after the pump to control the flow. The common categories of motor bearing faults at different operating speeds are analyzed. High-frequency vibration data is used as the original input, and this data is collected by a Wilconxon 786B-10 100mV / g uniaxial accelerometer with a sampling frequency of 20KHz.

[0106] The corresponding number of samples at different operating speeds (multiple working conditions) is shown in Table 1.

[0107] Table 1 Corresponding number of samples under multiple working conditions

[0108] In addition, the bearing fault categories are shown in Table 2.

[0109] Table 2 Bearing fault categories

[0110] To verify the performance of the model in diagnosing known and unknown faults under unknown working conditions, the generalization tasks shown in Table 3 are obtained. In each task, one more bearing fault type is added to the target working condition compared to the known working conditions, and only the data in the known working conditions is involved in the model training. The data in the target working conditions is only used for model performance evaluation.

[0111] Table 3 Generalization tasks for model verification

[0112] In all the above generalization tasks, the target working condition data is not involved in model training and is only used for testing. The evaluation metrics are the known fault classification accuracy (Acc_k) and the unknown fault classification accuracy (Acc_u) in the unknown target domain.

[0113] The performance metrics of the model on 12 groups of generalization tasks are shown in Table 4. The average metrics of the proposed method in 12 task groups reach a known accuracy of 89.413% and an unknown accuracy of 92.481%. For different target working conditions, the average accuracies of known fault classification at 735 RPM, 1102.5 RPM, and 1470 RPM working conditions are 82.208%, 91.118%, and 91.915% respectively, and the average accuracies of unknown fault identification are 87.953%, 97.333%, and 92.158% respectively.

[0114] Table 4 Model Performance Metrics

[0115] Furthermore, deleting the relevant components is denoted as M2, which is used to evaluate the impact of the auxiliary multi-binary classifier and on the model performance. In addition, MLDG (Meta-Learning for Domain Generalization) is denoted as M3, which is used to evaluate the impact of the binary learning strategy on the model performance. The accuracies corresponding to M2 and M3 are shown in the above table. Compared with M2, the average Acc_k and Acc_u of the model in this application on 12 groups of generalization tasks are increased by 12.476% and 16.98% respectively, indicating that the auxiliary multi-binary classifier and provide a more accurate classification boundary for fault identification under unknown working conditions.

[0116] In addition, compared with M3, on the basis of achieving an unknown accuracy of 92.481%, the proposed method increases the known fault identification accuracy by 9.645%, and this application of the present invention realizes a more fine-grained category-level model optimization.

[0117] What is not described in this invention can be realized by adopting or referring to the existing technologies.

[0118] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments.

[0119] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A rotating component fault diagnosis method based on collaborative training, characterized in that: include: Obtain a first fault type group and a second fault type group according to the fault type, and obtain a first element training set and a second element training set in combination with the collected model training data set, wherein the model training data set includes vibration data of the rotating component under multiple working conditions and multiple fault types; Based on the preset model parameters and the first training set, an auxiliary binary classifier is used to obtain an auxiliary binary classification loss, a cross entropy function is used to obtain a multi-classifier loss, and the combination is used to obtain an overall loss; According to the overall loss, the model parameters are updated, and the overall loss is obtained by combining the second element training set. The objective function is obtained according to the overall loss obtained from the first element training set and the second element training set for training the rotating component fault diagnosis model.

2. The rotating component fault diagnosis method based on collaborative training according to claim 1, characterized in that: The first fault type group and the second fault type group are obtained according to the fault type, specifically: According to the characteristics of the fault types, a first fault type group and a second fault type group are obtained through correlation analysis; and / or, According to the vibration intensity corresponding to the fault type, a first fault type group and a second fault type group are obtained.

3. The rotating component fault diagnosis method based on collaborative training according to claim 1, characterized in that: The first training set and the second training set are specifically: The first training set and the second training set both contain vibration data under multiple working conditions; The vibration data of at least one working condition of the first training set belongs to one of the first fault type group and the second fault type group, and the vibration data of at least another working condition belongs to the other of the first fault type group and the second fault type group; The vibration data of the second training set and the vibration data of the first training set under corresponding working conditions belong to different fault type groups.

4. The rotating component fault diagnosis method based on collaborative training according to claim 1, characterized in that: The overall losses are as follows: Combining the multi-classifier loss and the auxiliary binary classification loss to obtain the overall loss; in, represents the overall loss, represents the multi-classifier loss, represents the auxiliary binary classification loss.

5. The rotating component fault diagnosis method based on collaborative training according to claim 4 is characterized in that: According to the overall loss, the model parameters are updated, specifically: Update the model parameters according to the gradient of the overall loss obtained by the first training set under the preset model parameters and in combination with the set learning rate; in, and represents the vibration data of the first element training set, Indicates that the vibration data corresponding to the first working condition in the first element training set belongs to the first fault type group, Indicates that the vibration data corresponding to the second working condition in the first element training set belongs to the second fault type group; and Respectively and The gradient of the overall loss; Represents the learning rate of the setting; represents the preset model parameters; represents the updated model parameters.

6. The rotating component fault diagnosis method based on collaborative training according to claim 5, characterized in that: The objective function for rotating component fault diagnosis is obtained according to the overall loss obtained from the first training set and the second training set, specifically: in, and represents the vibration data of the second element training set, Indicates that the vibration data corresponding to the first working condition in the second element training set belongs to the second fault type group, Indicates that the vibration data corresponding to the second working condition in the second element training set belongs to the first fault type group; and Respectively and The overall loss under the preset model parameters; and Respectively and The overall loss under the updated model parameters.

7. The rotating component fault diagnosis method based on collaborative training according to claim 1, characterized in that: Also includes: According to the collected vibration data, a model training data set is obtained through preprocessing; The preprocessing includes at least data anomaly detection, sliding window segmentation, and time-frequency feature conversion processing, and the data anomaly detection is used to identify and eliminate abnormal data.

8. The rotating component fault diagnosis method based on collaborative training according to claim 7, characterized in that: The specific sliding window segmentation is: According to the vibration data after the abnormality is eliminated, a plurality of vibration data segments are obtained by a non-overlapping sliding window segmentation method; The window step size is determined according to the vibration data acquisition length, acquisition frequency and data vibration period.

9. The rotating component fault diagnosis method based on collaborative training according to claim 7, characterized in that: The time-frequency feature conversion process is as follows: According to the vibration data after the sliding window segmentation process, the negative frequency part of the signal is removed by using the instantaneous feature extraction method; According to the processed vibration data, the instantaneous power spectrum density is obtained through time-frequency analysis method; The frequency resolution of the time-frequency analysis method is determined according to the highest frequency of the collected vibration data, and the time resolution is determined according to the minimum bandwidth of the collected vibration data.

10. The rotating component fault diagnosis method based on collaborative training according to claim 1, characterized in that: Rotating component fault diagnosis model, specifically: According to the rotating component vibration data collected in real time, the confidence score is obtained by the auxiliary binary classifier in the rotating component fault diagnosis model, and the known fault type and the unknown fault type are judged by the confidence score; For known fault types, the probability distribution of known fault types is obtained through multiple classifiers in the rotating component fault diagnosis model to diagnose the rotating component fault.

Images