Rotary mechanical equipment fault diagnosis method and related device

Abstract

The invention discloses a rotary mechanical equipment fault diagnosis method and a related device. The method comprises the following steps: acquiring a vibration signal of rotary mechanical equipment to be diagnosed; and inputting the vibration signal into a pre-trained fault diagnosis model, and outputting a diagnosis result. Wherein the fault diagnosis model is obtained by training a knowledge distillation guided cost-sensitive ensemble learning network by adopting training data, and the training data comprises vibration signals corresponding to different faults and fault labels; the knowledge distillation guided cost-sensitive ensemble learning network comprises a multi-scale feature coding module, a multi-depth knowledge integration module and a classifier module which are connected in sequence, and the multi-scale feature coding module comprises a multi-scale feature extraction module and a high-order interaction fusion module. The multi-depth knowledge integration module comprises a multi-depth feature extraction module and a multi-depth knowledge fusion module, training of the knowledge distillation guided cost-sensitive ensemble learning network is divided into a first training stage and a second training stage, a loss function adopted in the first training stage is a multi-depth knowledge adaptive migration loss function, and a loss function adopted in the second training stage is a multi-depth knowledge adaptive migration loss function. The model parameters are frozen after training in the first training stage is finished. The objective of the invention is to solve the problem of data imbalance in rotary mechanical equipment fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and in particular relates to a rotating mechanical equipment fault diagnosis method and related devices. Background Art

[0002] With the advent of the era of Industry 4.0, modern industrial systems are gradually moving towards intelligence and automation. Rotating machinery, as the core equipment in these systems, is spread across many critical industrial fields such as aerospace, shipbuilding, and energy supply. Rotating machinery usually operates continuously in harsh and complex working environments, which makes them extremely prone to failure, which may lead to significant economic losses and safety accidents. In recent years, the vigorous development of deep learning technology has provided new opportunities for intelligent fault diagnosis. Deep learning-driven intelligent fault diagnosis methods have achieved remarkable results in many fields with their excellent information mining and feature representation capabilities. Specifically in the industrial field, a variety of deep network models, such as autoencoders, convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and transformers, have been widely used in fault diagnosis of industrial equipment and key components. These models usually take raw signals as input and directly convert data into health status information of the equipment through end-to-end mapping, thereby significantly improving the accuracy of fault diagnosis. However, in practical applications, intelligent diagnosis models face severe challenges of data category imbalance. In industrial sites, since most equipment operates in normal conditions, while failures are sporadic and expensive to detect, the data categories between healthy and faulty states are often extremely unbalanced, and there is a serious shortage of labeled fault samples. In addition, the probability of failure of different components varies, resulting in data imbalance between different fault categories (inter-class imbalance), and the probability of failures of different severity is also variable, which in turn causes data imbalance within a class (intra-class imbalance). These problems seriously restrict the effective application of intelligent diagnosis models in different industrial scenarios.

[0003] In response to the problem of data imbalance, academia and industry have carried out a lot of research work and proposed a variety of solutions. These solutions can be roughly divided into two categories: data-level methods and algorithm-level methods. Data-level methods mainly generate fault samples through resampling technology, generative adversarial networks (GANs) and variational autoencoders (VAEs) to alleviate data set bias. However, the performance of such methods is heavily dependent on sample distribution. When fault samples are scarce, it is difficult to ensure the reliability, authenticity and diversity of generated data. Algorithm-level methods mainly include ensemble learning methods and cost-sensitive learning methods. Ensemble learning methods achieve excellent diagnostic results by combining the diagnostic capabilities of multiple branch networks. However, the design of branch networks often lacks the embedding of imbalanced diagnostic knowledge, which makes it difficult to ensure the differences between branches, thus affecting the integration effect. Cost-sensitive learning methods improve the overall diagnostic ability of the algorithm by increasing the model's attention to minority classes. However, the distribution of weights depends on expert experience, making it difficult to stably guarantee the generalization performance of this method. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a method for diagnosing a fault of a rotating mechanical device and a related device, which aims to solve the problem of data imbalance in the fault diagnosis of the rotating mechanical device.

[0005] In order to solve the above technical problems, the present invention is implemented by the following technical solutions:

[0006] According to a first aspect of the present invention, there is provided a method for diagnosing a fault of a rotating mechanical device, comprising:

[0007] Obtaining vibration signals of rotating mechanical equipment to be diagnosed;

[0008] Inputting the vibration signal into a pre-trained fault diagnosis model and outputting a diagnosis result;

[0009] Among them, the fault diagnosis model is obtained by training a cost-sensitive integrated learning network guided by knowledge distillation using training data, the training data includes vibration signals and fault labels corresponding to different faults, the cost-sensitive integrated learning network guided by knowledge distillation includes a multi-scale feature encoding module, a multi-depth knowledge integration module and a classifier module connected in sequence, the multi-scale feature encoding module includes a multi-scale feature extraction module and a high-order interactive fusion module, the multi-depth knowledge integration module includes a multi-depth feature extraction module and a multi-depth knowledge fusion module, the training of the cost-sensitive integrated learning network guided by knowledge distillation is divided into a first training stage and a second training stage, the loss function used in the first training stage is a multi-depth knowledge adaptive migration loss function, the model parameters are frozen after the first training stage is completed, and the second training stage is entered, and the loss function used in the second training stage is a cost-sensitive loss function.

[0010] In a possible implementation of the first aspect, the multi-scale feature extraction module uses a plurality of parallel connected dilated convolution layers to filter the vibration signal through convolution kernels with different time-frequency resolutions to extract multi-scale features of the vibration signal;

[0011] The high-order interactive fusion module includes a first normalization layer, a recursive gated convolution module, a second normalization layer and a perceptron connected in sequence, and residual connections are respectively provided at the recursive gated convolution module and the perceptron. The high-order interactive fusion module is used to encode and fuse the multi-scale features of the vibration signal to obtain fused features;

[0012] The multi-depth feature extraction module includes a shallow feature extraction module and a deep feature extraction module, wherein the shallow feature extraction module is composed of a plurality of first convolutional layers connected in sequence, and the deep feature extraction module is composed of a plurality of second convolutional layers connected in parallel; the shallow feature extraction module is used to perform shallow feature extraction on the fused features to obtain a plurality of shallow features; the deep feature extraction module is used to perform deep feature extraction on a plurality of shallow features to obtain a plurality of deep knowledge;

[0013] The multi-depth knowledge fusion module includes a feature matching module and a knowledge fusion module, wherein the feature matching module is composed of a third convolution layer corresponding to a plurality of first convolution layers one by one, and the feature matching module is used to perform feature shape matching on a plurality of shallow features to obtain a plurality of shallow knowledge; the knowledge fusion module is used to fuse a plurality of shallow knowledge with a plurality of deep knowledge to obtain fused knowledge;

[0014] The classifier module includes a plurality of parallel classifiers, each of which is used to classify the faults of the fused knowledge and output the predicted probability of each fault.

[0015] In a possible implementation manner of the first aspect, the shallow feature extraction module is used to perform shallow feature extraction on the fused features to obtain multiple shallow features, specifically:

[0016]

[0017] In the formula, is the shallow feature extracted by the mth first convolutional layer; M is the number of first convolutional layers; represents the network parameters of the first convolutional layer of the mth layer; ° represents function synthesis, i.e. h°g(f I )=h(g(f I )), is the fused feature, W and C represent the length and number of channels of the feature respectively.

[0018] In a possible implementation manner of the first aspect, the deep feature extraction module is used to perform deep feature extraction on multiple shallow features to obtain multiple deep knowledge, specifically:

[0019]

[0020] In the formula, The deep knowledge extracted by the r-th second convolutional layer; is the network parameter of the second convolutional layer of the rth layer; It is the shallow feature extracted by the last second convolutional layer; R represents the number of second convolutional layers, M ≥ R.

[0021] In a possible implementation manner of the first aspect, the feature matching module is used to perform feature shape matching on multiple shallow features to obtain multiple shallow knowledge, specifically:

[0022]

[0023] In the formula, It is the shallow knowledge after the mth original shallow feature matching transformation; is the network parameter of the mth third convolutional layer.

[0024] In a possible implementation manner of the first aspect, the knowledge fusion module is used to fuse multiple shallow knowledge and multiple deep knowledge to obtain fused knowledge, specifically:

[0025]

[0026] In the formula, It is the integrated knowledge; It is the shallow knowledge after the rth original shallow feature matching transformation; is element-wise addition.

[0027] In a possible implementation manner of the first aspect, the multi-depth knowledge adaptive transfer loss function It consists of the self-distillation loss of the target class, the dynamic knowledge transfer loss of the non-target class, and the cross entropy loss, specifically:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] In the formula, represents the self-distillation loss of the target class; KL(·) represents the computational divergence; p j and p k Represent the predicted probabilities of the jth and kth classifiers respectively; Represents the dynamic knowledge transfer loss of non-target classes, and converts the output logits of the classifier Break down into goals and non-target in, is the index set of non-target classes, is the average logits of R classifiers; The most difficult category; For super teachers; and The non-target probabilities of the super teacher and the classifier r = [1, 2, ... R] as students, respectively; represents the cross entropy loss, is the output of the rth classifier regarding the kth type of fault, K is the total number of fault classes; α and β are hyperparameters used to balance the contribution of the three losses during training.

[0037] In a possible implementation manner of the first aspect, the cost-sensitive loss function Specifically:

[0038]

[0039] Where n yThe number of samples indicating label failure; n j Indicates the number of samples of each type of fault.

[0040] According to a second aspect of the present invention, there is provided a rotating mechanical equipment fault diagnosis device, comprising:

[0041] An acquisition module, used for acquiring a vibration signal of a rotating mechanical device to be diagnosed;

[0042] A diagnosis module, used for inputting the vibration signal into a pre-trained fault diagnosis model and outputting a diagnosis result;

[0043] Among them, the fault diagnosis model is obtained by training a cost-sensitive integrated learning network guided by knowledge distillation using training data, the training data includes vibration signals and fault labels corresponding to different faults, the cost-sensitive integrated learning network guided by knowledge distillation includes a multi-scale feature encoding module, a multi-depth knowledge integration module and a classifier module connected in sequence, the multi-scale feature encoding module includes a multi-scale feature extraction module and a high-order interactive fusion module, the multi-depth knowledge integration module includes a multi-depth feature extraction module and a multi-depth knowledge fusion module, the training of the cost-sensitive integrated learning network guided by knowledge distillation is divided into a first training stage and a second training stage, the loss function used in the first training stage is a multi-depth knowledge adaptive migration loss function, the model parameters are frozen after the first training stage is completed, and the second training stage is entered, and the loss function used in the second training stage is a cost-sensitive loss function.

[0044] According to a third aspect of the present invention, there is provided a device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for diagnosing a fault of a rotating mechanical device when executing the computer program.

[0045] Compared with the prior art, the present invention has at least the following beneficial effects:

[0046] The present invention provides a method for fault diagnosis of rotating mechanical equipment. It effectively combines integrated learning and cost-sensitive learning to fully extract multi-scale features, effectively utilize key multi-depth features, and emphasize the classification of the most easily confused categories. Specifically, multi-scale feature extraction and multi-order fusion are first used to fully utilize the limited diagnostic knowledge of minority classes, thereby improving the ability to utilize limited diagnostic information. Then, the diverse and complementary diagnostic knowledge (various unbalanced diagnostic knowledge) from different depths of the network is embedded into the new integrated learning process to better make integrated decisions. Then, the improved knowledge distillation method realizes the mutual migration and sublimation of excellent diagnostic knowledge, fully explores potential diagnostic knowledge, improves the generalization ability of the model, and focuses on the most easily confused fault categories to achieve effective representation of various types of faults. Finally, a cost-sensitive strategy is adopted to further increase the focus on minority classes. Compared with the existing sample imbalance method based on deep networks, the experimental results under different working conditions show that the method proposed by the present invention can achieve the most advanced performance, showing great prospects as an intelligent fault diagnosis method, and is of great significance for solving the problem of unbalanced diagnosis in industrial applications.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the specific implementation modes of the present invention, the drawings required for use in the description of the specific implementation modes will be briefly introduced below. Obviously, the drawings described below are some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 A method for diagnosing a fault of a rotating mechanical device according to an embodiment of the present invention;

[0050] Figure 2 The specific structure of the cost-sensitive ensemble learning network guided by knowledge distillation in the embodiment of the present invention;

[0051] Figure 3 The structure of the MFEM proposed in the embodiment of the present invention;

[0052] Figure 4 The structure of the DKEELM proposed in the embodiment of the present invention;

[0053] Figure 5 The structure of the MKATM proposed in the embodiment of the present invention;

[0054] Figure 6 is the classification result of the wind power gearbox data set according to an embodiment of the present invention;

[0055] Figure 7 This is the confusion matrix analysis of the wind turbine gearbox data set according to an embodiment of the present invention, including: (a) CE loss, (b) CSloss, (c) ECE loss, (d) SMOTE, (e) SMOTETomek, and (f) Proposed CSELF. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Combination Figure 1 As shown, an embodiment of the present invention provides a method for diagnosing a fault of a rotating mechanical device, which effectively solves the problem of data imbalance in the diagnosis of a fault of a rotating mechanical device by constructing a cost-sensitive integrated learning network guided by knowledge distillation. The method for diagnosing a fault of a rotating mechanical device specifically includes the following steps:

[0058] S1. Obtain a vibration signal of a rotating mechanical device to be diagnosed.

[0059] Specifically, vibration signals are obtained from the rotating mechanical equipment to be diagnosed. For example, these vibration signals are collected by sensors and undergo preprocessing steps, such as denoising, filtering, normalization, etc., to improve signal quality. The preprocessed vibration signals will be used as input for subsequent fault diagnosis models.

[0060] S2. Input the vibration signal into a pre-trained fault diagnosis model and output a diagnosis result.

[0061] That is, the trained fault diagnosis model converts the input vibration signal into the health status information of the equipment and outputs the diagnosis result. The diagnosis result includes the predicted probability of each fault, based on which it can be determined whether the equipment has a fault and the type of fault.

[0062] Specifically, the fault diagnosis model is obtained by training a cost-sensitive ensemble learning network guided by knowledge distillation using training data, wherein the training data includes vibration signals corresponding to different faults and fault labels.

[0063] Combination Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, the cost-sensitive ensemble learning network guided by knowledge distillation includes a multi-scale feature encoding module (MFEM), a multi-depth knowledge integration module (DKEELM) and a classifier module connected in sequence. Figure 2 and Figure 3 As shown in FIG. 1 , the multi-scale feature encoding module includes a multi-scale feature extraction module and a high-order interactive fusion module. The multi-scale feature extraction module uses a plurality of parallel connected hole convolution layers to filter the vibration signal through convolution kernels with different time-frequency resolutions to extract the multi-scale features of the vibration signal. The high-order interactive fusion module includes a first normalization layer, a recursive gated convolution module (g n Conv), a second normalization layer and a perceptron, the high-order interactive fusion module is used to encode and fuse the multi-scale features of the vibration signal, and the interaction and fusion between the features are realized through the recursive gated convolution module to obtain the fused features.

[0064] In other words, a lightweight MFEM is used to extract multi-scale features and perform high-order interactive fusion of vibration signals. Multi-scale learning utilizes the multi-resolution characteristics of filters of different scales to effectively mine complementary discriminant features and construct a feature set containing all sensitive frequency bands. Therefore, a multi-scale dilated convolution layer is used as the first layer of the entire network to filter the vibration signal through convolution kernels of different time-frequency resolutions to extract rich multi-scale features. In particular, the parallel connection of multiple dilated convolutions enhances the feature encoding capability while achieving a lightweight model and avoiding the overfitting problem caused by too few fault samples.

[0065] A high-order interactive fusion module is used to deeply encode multi-scale features to further enhance the representation and fusion performance of multi-scale features, thereby obtaining more discriminative features for the subsequent diagnosis process.

[0066] Regarding the interaction and fusion of features through recursive gated convolution modules, more specifically, g n Conv performs high-order and long-term spatial interactions through gated convolution and recursive operations, modeling high-order spatial relationships of features at different levels with low computational cost. n The basic module of Conv is gated convolution. Assume that the input feature is Where W and C represent the length and number of channels of the feature respectively, the process of gated convolution is expressed as:

[0067]

[0068]

[0069]

[0070] Among them, f(·) is the depth convolution layer, φ in ,φ out is a linear projection layer. Specifically,

[0071]

[0072] Where w is the weight of f(·), Ω i is a local window. Gated convolution realizes the first-order interaction between adjacent features through element-wise multiplication.

[0073] On this basis, g n Conv introduces high-order interactions to further improve the model capacity. Specifically, we first use φ in Get a set of projection features p0 and

[0074]

[0075] Then, a gated convolution operation is recursively performed on this basis.

[0076] p k+1 =f k (q k )⊙g k (p k ) / α

[0077] where {g k} Cross-order matching feature dimension, {f k} is a depthwise convolutional layer, k is the interaction order, and α is a scaling factor used to smooth the training process.

[0078] Finally, the high-order feature p obtained recursively in the previous step is n Input φ out Perform linear mapping to obtain g n Conv output. It is worth noting that when performing multi-step gated convolution, the convolution operation only requires a single f to transform the feature Therefore, g n Conv realizes arbitrary-order interactions between coarse to fine features with less computational overhead, greatly enhancing the fusion capability of multi-scale features.

[0079] Combination Figure 2 and Figure 4As shown, the multi-depth knowledge integration module includes a multi-depth feature extraction module and a multi-depth knowledge fusion module. The multi-depth feature extraction module includes a shallow feature extraction module and a deep feature extraction module. The shallow feature extraction module is composed of a plurality of first convolutional layers connected in sequence. The shallow feature extraction module is used to perform shallow feature extraction on the fused features to obtain a plurality of shallow features, specifically:

[0080]

[0081] In the formula, is the shallow feature extracted by the mth first convolutional layer; M is the number of first convolutional layers; Represents the network parameters of the mth first convolutional layer; represents function synthesis, that is, h°g(f I )=h(g(f I )), The fused features.

[0082] That is to say, the fused features are first encoded by multiple first convolutional layers to obtain a series of original shallow features.

[0083] The deep feature extraction module is composed of a plurality of second convolutional layers connected in parallel. The deep feature extraction module is used to perform deep feature extraction on a plurality of shallow features to obtain a plurality of deep knowledge, specifically:

[0084]

[0085] In the formula, The deep knowledge extracted by the r-th second convolutional layer; is the network parameter of the second convolutional layer of the rth layer; It is the shallow feature extracted by the last second convolutional layer; R represents the number of second convolutional layers, M ≥ R.

[0086] That is to say, while the shallow feature extraction module extracts the original shallow features with diverse diagnostic knowledge, it also serves as the front-end shared network of the deep feature extraction module. The deep feature extraction module learning architecture based on lightweight embedding contains multiple second convolutional layers with the same structure, which further learns deep feature encoding and improves the generalization performance of the model.

[0087] The multi-depth knowledge fusion module includes a feature matching module and a knowledge fusion module. The feature matching module is composed of a third convolutional layer corresponding to a plurality of first convolutional layers. The feature matching module is used to perform feature shape matching on a plurality of shallow features to obtain a plurality of shallow knowledge, specifically:

[0088]

[0089] In the formula, It is the shallow knowledge after the mth original shallow feature matching transformation; is the network parameter of the mth third convolutional layer.

[0090] That is, in order to ensure that multiple second convolutional layers have diverse and complementary feature representation capabilities for imbalanced distribution, the diagnostic knowledge of the shallow feature extraction module is embedded into the corresponding third convolutional layer respectively. More specifically, the original shallow features are first convolutionally encoded by the corresponding third convolutional layers to match the shapes of different deep features.

[0091] The knowledge fusion module is used to fuse multiple shallow knowledge and multiple deep knowledge to obtain fused knowledge, specifically:

[0092]

[0093] In the formula, It is the integrated knowledge; It is the shallow knowledge after the rth original shallow feature matching transformation; is element-wise addition.

[0094] That is, shallow knowledge of different depths is aggregated with multiple deep knowledge through the knowledge fusion module. In particular, setting M ≥ R ensures that each third convolutional layer utilizes the shallow knowledge of the only depth in the network, implicitly assigning different imbalanced distribution diagnostic preferences to the third convolutional layer, further enabling the model to have excellent diagnostic performance for all fault categories through comprehensive decision making.

[0095] In summary, the diagnostic knowledge represented by features of different depths in the model has different diagnostic capabilities for different fault categories in the imbalanced distribution. By realizing the fusion and integration of diagnostic knowledge of different depths, the optimal diagnostic knowledge of the entire network can be fundamentally mined, rather than relying solely on deep diagnostic knowledge.

[0096] The classifier module includes a plurality of parallel classifiers, each of which is used to classify the faults of the fused knowledge and output the predicted probability of each fault.

[0097] Combination Figure 2 and Figure 5 As shown, the training of the cost-sensitive ensemble learning network guided by knowledge distillation is divided into two stages: the first training stage and the second training stage. The loss function used in the first training stage is the multi-depth knowledge adaptive migration (MKATM) loss function. It consists of the self-distillation loss of the target class, the dynamic knowledge transfer loss of the non-target class, and the cross entropy loss, specifically:

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] In the formula, represents the self-distillation loss of the target class; KL(·) represents the computational divergence; p j and p k Represent the predicted probabilities of the jth and kth classifiers respectively; Represents the dynamic knowledge transfer loss of non-target classes, and converts the output logits of the classifier Break down into goals and non-target in, is the index set of non-target classes, is the average logits of R classifiers; The most difficult category; For super teachers; and The non-target probabilities of the super teacher and the classifier r = [1, 2, ... R] as students, respectively; represents the cross entropy loss, is the output of the rth classifier regarding the kth type of fault, K is the total number of fault classes; α and β are hyperparameters used to balance the contribution of the three losses during training.

[0107] Specifically, the derivation of the multi-depth knowledge adaptive transfer loss function is described in detail as follows:

[0108] The self-distillation method is introduced to promote mutual learning between any two second convolutional layers, that is, each second convolutional layer acts as a mentor to teach excellent diagnostic knowledge of certain fault categories, and at the same time acts as a student to learn the excellent diagnostic knowledge of other second convolutional layers. Therefore, the mutual knowledge transfer and sublimation between any two second convolutional layers provides an effective way for the model to adaptively aggregate multi-depth knowledge. The above process is expressed as:

[0109]

[0110] in represents the self-distillation loss of the target class, p j and p k denote the predicted probabilities of the j-th and k-th classifiers, respectively.

[0111] However, studies have shown that non-target logits have a significant impact on knowledge distillation, and it is more helpful in dealing with imbalanced distributions because minority classes usually have smaller target logits during training, resulting in the contribution of non-target logits being comparable to that of target logits. In addition, some samples in the minority class will have a fairly high probability of misclassification to the majority class (negative class), especially when their features are confused with each other. The non-target class with the highest probability of misclassification is called the hardest negative class. If all classifiers cannot effectively distinguish between the minority class and the hardest negative class, then there is a high probability that incorrect diagnostic knowledge will be transferred. Therefore, the dynamic knowledge transfer (DKT) method that only considers non-target class predictions is introduced to solve the hardest negative class. Assume that the label corresponding to sample x is y, and the output logits of the classifier are First, decompose the logits into target logits and non-target logits in is the index set of non-target classes. Then knowledge distillation is performed on non-target classes to achieve more detailed knowledge transfer. Specifically, the average logarithm of all classifiers on each non-target class is calculated as:

[0112]

[0113] Where i=1,...,K-1. Then It is the most difficult negative class jointly judged by multiple classifiers. In order to effectively suppress the logits value of the most difficult negative class from being too large, the largest non-target class logits among all classifiers are selected as experienced teachers, and the knowledge of non-target classes is integrated to guide the training process, which can be expressed as:

[0114]

[0115] Due to the diversity of diagnostic knowledge embedded in different classifiers, different classifiers may experience maximum values ​​on different non-target class logits. Therefore, the selected teacher can dynamically suppress the logit of the hardest negative class, thereby reducing the negative impact of the hardest negative class in the knowledge transfer process and alleviating the degree of confusion between the most confusing classes. The hardest negative class is then combined with the maximum non-target logit to create a set of non-target pairs, called super teachers, as shown below:

[0116]

[0117] After selecting the super teacher, the non-target knowledge is refined for each learner. The non-target probability of the super teacher and the individual learner as a student can be expressed as:

[0118]

[0119] Furthermore, the dynamic knowledge transfer (knowledge extraction) based on non-target classes can be expressed as:

[0120]

[0121] For a specific sample, the hardest negative class not only varies among multiple classifiers, but may also keep changing during training. The adopted method can dynamically select the hardest negative class and adaptively construct a super teacher to suppress the logit of the hardest negative class in the non-target class, thereby improving the model's ability to distinguish confusion faults.

[0122] In addition, cross entropy loss is used to ensure classification accuracy, which can be expressed as:

[0123]

[0124] In summary, the overall optimization goal can be expressed as follows:

[0125]

[0126] After the first training phase is completed, the model parameters are frozen and the second training phase is entered. The loss function used in the second training phase is a cost-sensitive loss function. Specifically:

[0127]

[0128] Where n y The number of samples indicating label failure; n j Indicates the number of samples of each type of fault.

[0129] Specifically, the cost-sensitive loss function improves the overall diagnostic ability of the algorithm by increasing the model's attention to the minority class.

[0130] In a specific implementation case, experimental results for various complex data imbalance scenarios such as extreme imbalance, step imbalance, continuous imbalance, inter-class imbalance, and intra-class imbalance all show that a rotating mechanical equipment fault diagnosis method provided by the present invention can achieve optimal performance and provide a promising solution for the actual industrial application of intelligent diagnosis methods.

[0131] Specifically, the wind turbine gearbox data set obtained from the wind turbine transmission system fault simulation experimental platform of Xi'an Jiaotong University was used for verification. The three BK accelerometers installed on the gearbox can collect vibration signals in three directions, of which the radial vertical signal is used for experimental verification. The experiment collected gear data in five healthy states: normal (NS), surface wear (SW), tooth root crack (RC), broken tooth (CT), and missing tooth (MT). During the experiment, the speed of each experiment was 2000r / min.

[0132] Step 1: Data collection and preprocessing. The sampling frequency is set to 20480Hz. For the above five health states, 1400 samples are collected for each health state, and each sample contains 2048 data points. Among them, 1000 samples are used for fault diagnosis model training, and the number of samples for each class in the validation set and the test set is 200. In addition, in order to better simulate the data imbalance scenario of the industrial site, multiple diagnosis tasks are constructed, as shown in Table 1, including both the step imbalance (SI) commonly used in academic research and the more realistic continuous imbalance (CI). More specifically, Task 1 is balanced and contains all health states, and the amount of data for each health state is equal. Tasks 2–6 are unbalanced, with different combinations of imbalance rates and imbalance types to ensure the uniqueness of each task, but all belong to the data form of multiple fault states. From more details, it can be seen that Task 2 belongs to SI with one majority class, Tasks 3–5 are CI with different imbalance rates, and Task 6 corresponds to SI involving multiple majority classes. The training set with unbalanced data volume is used in the two-stage model training process to optimize the feature extractor and classifier respectively, which is conducive to representation learning under data imbalance conditions. The validation set and test set with balanced data volume are used for optimal model selection and model testing respectively.

[0133] Table 1 Data division of diagnostic tasks

[0134]

[0135] Step 2: First training phase. First, MFEM is used to encode the vibration signal and fully exploit the multi-scale information. Then DKEELM is used to achieve the fusion and integration of different depth diagnostic knowledge. Finally, MKATM is used to jointly optimize the network parameters to promote the mutual transfer and sublimation of excellent and complementary diagnostic knowledge, and obtain a feature extractor with excellent performance. For example, Table 2 and Figure 3 The network structure layers of MFEM are introduced in detail.

[0136] Table 2 Network structure of MFEM

[0137]

[0138]

[0139] Step 2.2: A new type of DKEELM is constructed. The diagnostic knowledge represented by the features of different depths in the model has different diagnostic capabilities for different fault categories in the imbalanced distribution. By realizing the fusion and integration of diagnostic knowledge of different depths, the optimal diagnostic knowledge of the entire network is fundamentally mined, rather than relying solely on deep diagnostic knowledge. For example, Table 3 and Figure 4 The network structure layers of DKEELM are introduced in detail.

[0140] Table 3 Network structure of DKEELM

[0141]

[0142] For example, α and β are set to 1 during training. Step 3: Second training phase. With the parameters of the model feature extractor frozen, the classifier is trained using a cost-sensitive loss function to optimize a balanced classifier for each second convolutional layer.

[0143] For example, during the training process, the stochastic gradient descent with restart (SGDR) algorithm is used to optimize the model parameters. The first training stage 1 is iterated 250 times, and the second training stage is iterated 150 times. During the training process of the proposed method, the stochastic gradient descent with restart (SGDR) algorithm is used to optimize the model parameters. The training stage 1 is iterated 250 times, and the training stage 2 is iterated 150 times. Step 4: Testing stage. The diagnostic results are obtained by testing the trained model using the test set.

[0144] In order to demonstrate the superiority of the proposed method, accuracy, F1-Score and G-mean are used as evaluation indicators, and five advanced methods are used for comparison, including a basic model, two algorithm-level methods and two data-level methods.

[0145] 1) The base model only uses CE loss to train the model without any strategy for data imbalance.

[0146] 2) Algorithm-level method 1: Enhance the attention to minority classes by re-weighting samples, relying on the sample size CS loss.

[0147] 3) Algorithm-level method 2: Enhance the attention to minority classes by re-weighting samples, and ECE loss that considers both sample difficulty and sample size.

[0148] 4) Data level method 1: oversampling technique SMOTE.

[0149] 5) Data level method 2: hybrid sampling technology SMOTETomek.

[0150] For the sake of fairness, all the comparison methods use the same network architecture as the method proposed in this invention, but do not include the multi-depth knowledge fusion module proposed in this invention. The Adam algorithm with a learning rate of 0.001 is used to optimize the model parameters of the comparison methods, and 400 iterations of training are performed. The comparison results of the six models are as follows:

[0151] 1) Table 4 and Figure 6 The diagnostic results of different methods for different diagnostic tasks are given. Obviously, the CSELF proposed in the present invention provides the best results on all three evaluation indicators of the six diagnostic tasks, and the smallest standard deviation shows that the proposed method has good stability, which proves the superiority of the method proposed in the present invention. In multiple diagnostic tasks, as the degree of imbalance increases, the performance of the comparison method drops sharply, while the method proposed in the present invention still shows stable and reliable diagnostic performance. This shows that the method proposed in the present invention has strong generalization performance. In addition, the excellence of CSELF is more significant in diagnostic tasks with a large imbalance rate. In particular, in the most unbalanced scenario T6, the F1-Score of the CSELF proposed in the present invention reaches 0.9474, which is 0.27 higher than the best comparison method SMOTE, proving the superiority of the method proposed in the present invention in extremely unbalanced scenarios. In addition, in some diagnostic tasks, since cost-sensitive learning may damage the quality of representation learning in some cases, the diagnostic effect of the algorithm-level method is reduced compared with the basic model, which is precisely the advantage of the two-stage decoupling training method adopted in the present invention. It is worth noting that as the imbalance rate of diagnostic tasks increases, the difference between the accuracy and F1-Score of the comparison methods gradually increases, which indicates that the model overemphasizes the majority class and ignores the minority class. The proposed method integrates and transfers the optimal diagnostic knowledge in the entire network, effectively diagnoses both the majority class and the minority class, and the difference between the accuracy and F1-Score on the six diagnostic tasks remains basically consistent.

[0152] Table 4 Diagnosis results of different methods for different diagnosis tasks

[0153]

[0154]

[0155] 2) In order to analyze the diagnostic effect of each fault category in detail, the best results of 10 tests on the T6 task are selected, and the prediction results of each method are further displayed through confusion matrix analysis. Figure 7It can be seen that all the compared methods have large classification errors in the RC and MT classes, and even completely lose the diagnostic ability. Overall, the comparison methods fluctuate greatly and have great limitations in the diagnosis of minority class faults. In contrast, CSELF has achieved an accuracy of 0.94 on MT faults with very few samples, which has obvious advantages. Further analysis shows that since the RC and MT fault signals are relatively similar, the misclassification is mainly concentrated between RC and MT faults, but compared with the comparison methods, CSELF also obtains satisfactory diagnostic results for these two most easily confused categories. In summary, the method proposed in the present invention integrates and transfers the optimal diagnostic knowledge in the entire model, simultaneously improves the characterization and diagnostic capabilities of various fault types, and greatly improves the diagnostic accuracy of minority classes by increasing the focus on confusing classes.

[0156] In another embodiment of the present invention, a rotating mechanical equipment fault diagnosis device is provided, which is used to implement a rotating mechanical equipment fault diagnosis method in the above embodiment, including:

[0157] The acquisition module is used to acquire the vibration signal of the rotating mechanical equipment to be diagnosed.

[0158] A diagnosis module, used for inputting the vibration signal into a pre-trained fault diagnosis model and outputting a diagnosis result;

[0159] Among them, the fault diagnosis model is obtained by training a cost-sensitive integrated learning network guided by knowledge distillation using training data, the training data includes vibration signals and fault labels corresponding to different faults, the cost-sensitive integrated learning network guided by knowledge distillation includes a multi-scale feature encoding module, a multi-depth knowledge integration module and a classifier module connected in sequence, the multi-scale feature encoding module includes a multi-scale feature extraction module and a high-order interactive fusion module, the multi-depth knowledge integration module includes a multi-depth feature extraction module and a multi-depth knowledge fusion module, the training of the cost-sensitive integrated learning network guided by knowledge distillation is divided into a first training stage and a second training stage, the loss function used in the first training stage is a multi-depth knowledge adaptive migration loss function, the model parameters are frozen after the first training stage is completed, and the second training stage is entered, and the loss function used in the second training stage is a cost-sensitive loss function.

[0160] All relevant contents of each step involved in the aforementioned embodiment of a method for diagnosing a fault of a rotating mechanical device can be referred to the functional description of a functional module corresponding to a device for diagnosing a fault of a rotating mechanical device in an embodiment of the present invention, and will not be repeated here. The division of modules in the embodiment of the present invention is schematic, and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional module in each embodiment of the present invention may be integrated into a processor, or may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module.

[0161] In another embodiment of the present invention, a computer device is provided, the computer device comprising a processor and a memory, the memory being used to store a computer program, the computer program comprising program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in a computer storage medium to implement a corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of a method for diagnosing a fault of a rotating mechanical device.

[0162] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the above-mentioned embodiment of a method for diagnosing a fault of a rotating mechanical device.

[0163] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0164] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0165] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0166] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0167] In the present invention, the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" etc. mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0168] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for diagnosing a fault of a rotating mechanical device, characterized in that: include: Obtaining vibration signals of rotating mechanical equipment to be diagnosed; Inputting the vibration signal into a pre-trained fault diagnosis model and outputting a diagnosis result; Among them, the fault diagnosis model is obtained by training a cost-sensitive integrated learning network guided by knowledge distillation using training data, the training data includes vibration signals and fault labels corresponding to different faults, the cost-sensitive integrated learning network guided by knowledge distillation includes a multi-scale feature encoding module, a multi-depth knowledge integration module and a classifier module connected in sequence, the multi-scale feature encoding module includes a multi-scale feature extraction module and a high-order interactive fusion module, the multi-depth knowledge integration module includes a multi-depth feature extraction module and a multi-depth knowledge fusion module, the training of the cost-sensitive integrated learning network guided by knowledge distillation is divided into a first training stage and a second training stage, the loss function used in the first training stage is a multi-depth knowledge adaptive migration loss function, the model parameters are frozen after the first training stage is completed, and the second training stage is entered, and the loss function used in the second training stage is a cost-sensitive loss function.

2. A method for diagnosing a fault of a rotating mechanical device according to claim 1, characterized in that: The multi-scale feature extraction module uses multiple parallel connected hole convolution layers to filter the vibration signal through convolution kernels with different time-frequency resolutions to extract multi-scale features of the vibration signal; The high-order interactive fusion module includes a first normalization layer, a recursive gated convolution module, a second normalization layer and a perceptron connected in sequence, and residual connections are respectively provided at the recursive gated convolution module and the perceptron. The high-order interactive fusion module is used to encode and fuse the multi-scale features of the vibration signal to obtain fused features; The multi-depth feature extraction module includes a shallow feature extraction module and a deep feature extraction module, wherein the shallow feature extraction module is composed of a plurality of first convolutional layers connected in sequence, and the deep feature extraction module is composed of a plurality of second convolutional layers connected in parallel; the shallow feature extraction module is used to perform shallow feature extraction on the fused features to obtain a plurality of shallow features; the deep feature extraction module is used to perform deep feature extraction on a plurality of shallow features to obtain a plurality of deep knowledge; The multi-depth knowledge fusion module includes a feature matching module and a knowledge fusion module, wherein the feature matching module is composed of a third convolution layer corresponding to a plurality of first convolution layers one by one, and the feature matching module is used to perform feature shape matching on a plurality of shallow features to obtain a plurality of shallow knowledge; the knowledge fusion module is used to fuse a plurality of shallow knowledge with a plurality of deep knowledge to obtain fused knowledge; The classifier module includes a plurality of parallel classifiers, each of which is used to classify the faults of the fused knowledge and output the predicted probability of each fault.

3. A method for diagnosing a fault of a rotating mechanical device according to claim 2, characterized in that: The shallow feature extraction module is used to perform shallow feature extraction on the fused features to obtain multiple shallow features, specifically: In the formula, is the shallow feature extracted by the mth first convolutional layer; M is the number of first convolutional layers; Represents the network parameters of the mth first convolutional layer; represents function composition, that is is the fused feature, W and C represent the length and number of channels of the feature respectively.

4. A method for diagnosing a fault of a rotating mechanical device according to claim 3, characterized in that: The deep feature extraction module is used to perform deep feature extraction on multiple shallow features to obtain multiple deep knowledge, specifically: In the formula, is the deep knowledge extracted by the second convolutional layer of the rth layer; θ D r is the network parameter of the rth second convolutional layer; It is the shallow feature extracted by the last second convolutional layer; R represents the number of second convolutional layers, M ≥ R.

5. A method for diagnosing a fault of a rotating mechanical device according to claim 4, characterized in that: The feature matching module is used to perform feature shape matching on multiple shallow features to obtain multiple shallow knowledge, specifically: In the formula, It is the shallow knowledge after the mth original shallow feature matching transformation; is the network parameter of the mth third convolutional layer.

6. A method for diagnosing a fault of a rotating mechanical device according to claim 5, characterized in that: The knowledge fusion module is used to fuse multiple shallow knowledge and multiple deep knowledge to obtain fused knowledge, specifically: In the formula, It is the integrated knowledge; It is the shallow knowledge after the rth original shallow feature matching transformation; is element-wise addition.

7. A method for diagnosing a fault of a rotating mechanical device according to claim 6, characterized in that: The multi-depth knowledge adaptive transfer loss function It consists of the self-distillation loss of the target class, the dynamic knowledge transfer loss of the non-target class, and the cross entropy loss, specifically: In the formula, represents the self-distillation loss of the target class; KL(·) represents the computational divergence; p j and p k Represent the predicted probabilities of the jth and kth classifiers respectively; Represents the dynamic knowledge transfer loss of non-target classes, and converts the output logits of the classifier Break down into goals and non-target in, is the index set of non-target classes, is the average logits of R classifiers; The most difficult category; For super teachers; and The non-target probabilities of the super teacher and the classifier r = [1, 2, ... R] as students, respectively; represents the cross entropy loss, is the output of the rth classifier regarding the kth type of fault, K is the total number of fault classes; α and β are hyperparameters used to balance the contribution of the three losses during training.

8. A method for diagnosing a fault of a rotating mechanical device according to claim 6, characterized in that: The cost-sensitive loss function Specifically: Where n y The number of samples indicating label failure; n j Indicates the number of samples of each type of fault.

9. A rotating mechanical equipment fault diagnosis device, characterized in that: include: An acquisition module, used for acquiring a vibration signal of a rotating mechanical device to be diagnosed; A diagnosis module, used for inputting the vibration signal into a pre-trained fault diagnosis model and outputting a diagnosis result; Among them, the fault diagnosis model is obtained by training a cost-sensitive integrated learning network guided by knowledge distillation using training data, the training data includes vibration signals and fault labels corresponding to different faults, the cost-sensitive integrated learning network guided by knowledge distillation includes a multi-scale feature encoding module, a multi-depth knowledge integration module and a classifier module connected in sequence, the multi-scale feature encoding module includes a multi-scale feature extraction module and a high-order interactive fusion module, the multi-depth knowledge integration module includes a multi-depth feature extraction module and a multi-depth knowledge fusion module, the training of the cost-sensitive integrated learning network guided by knowledge distillation is divided into a first training stage and a second training stage, the loss function used in the first training stage is a multi-depth knowledge adaptive migration loss function, the model parameters are frozen after the first training stage is completed, and the second training stage is entered, and the loss function used in the second training stage is a cost-sensitive loss function.

10. A device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for diagnosing a fault of a rotating mechanical equipment as described in any one of claims 1 to 8 is implemented.

Images