Rotating machinery component collaborative diagnosis method considering new failure mode

Abstract

A rotating mechanical component collaborative diagnosis method considering new fault mode is provided.The method constructs a local feature reuse convolution network in an offline stage, constructs a differentiated discriminant index to measure the difference of the diagnosis output of the sample on different local feature reuse convolution networks, and realizes the training of the collaborative diagnosis model; in an online stage, the average integrated diagnosis vector obtained by the trained convolution network, the differentiated discriminant index and the differentiated discriminant threshold are comprehensively used to implement collaborative diagnosis, and reliable fault diagnosis of the rotating mechanical component considering the new fault mode is realized. The local feature reuse convolution network and the triple differentiation enhancement strategy are used to improve the accuracy and differentiation of the collaborative diagnosis base model respectively, so as to drive the collaborative diagnosis of the rotating mechanical component, fully utilize the discrimination ability of the model, and consider the accurate diagnosis of the known fault mode and the new fault mode, overcome the limitation of the diagnosis precision caused by only relying on the data or feature difference, and have important significance for guaranteeing the reliability and safety of the rotating mechanical component.

Description

Technical Field

[0001] This invention relates to a technology in the field of mechanical equipment fault diagnosis, specifically a collaborative diagnosis method for rotating mechanical components that considers emerging fault modes. Background Technology

[0002] Most existing data-driven diagnostic technologies for mechanical components can only identify known fault modes and cannot identify newly emerging fault modes. Some technologies identify new fault modes by evaluating the similarity between data or features, without fully utilizing the model's learning and discriminative capabilities. Due to the non-stationary and noisy characteristics of signals from rotating mechanical components in industrial settings, relying solely on data or feature differences for fault diagnosis has poor reliability. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, such as the lack of targeted design for the fusion learning of fault features and the susceptibility of diagnostic accuracy to test uncertainties. It proposes a collaborative diagnostic method for rotating mechanical components that considers emerging fault modes. By employing a local feature reuse convolutional network and a triple differentiation enhancement strategy, the accuracy and differentiation of the collaborative diagnostic base model are improved, thereby driving collaborative diagnosis of rotating mechanical components. By fully utilizing the model's discriminative capabilities to accurately diagnose both known and emerging fault modes, this method overcomes the limitations in diagnostic accuracy caused by relying solely on data or feature differences. This is of great significance for ensuring the reliability and safety of rotating mechanical components.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a collaborative diagnostic method for rotating mechanical components that considers emerging failure modes. In the offline stage, after constructing a local feature reuse convolutional network, a differential discrimination index is built to measure the differences in diagnostic output of samples on different local feature reuse convolutional networks. This enables the training of a collaborative diagnostic model composed of different local feature reuse convolutional networks and a defined differential discrimination threshold. In the online stage, the average integrated diagnostic vector, differential discrimination index, and differential discrimination threshold obtained from the trained convolutional network are used for collaborative diagnosis, achieving reliable fault diagnosis of rotating mechanical components that considers emerging failure modes.

[0006] The local feature reuse convolutional network includes: several local feature reuse modules, a tiling layer, and a fully connected layer. Each local feature reuse module decomposes the feature map into directly connected feature maps and cross-connected feature maps. By reusing the feature maps locally, it efficiently learns deep fault features and improves the classification ability of the base model. The tiling layer flattens the feature map into one dimension to meet the data requirements of the fully connected layer. The fully connected layer performs classification by performing nonlinear mapping on the flattened feature map.

[0007] The local feature reuse module includes two convolutional layers and one fusion pooling layer, wherein the first convolutional layer performs convolution operations on all input feature maps, specifically: in: This represents the o-th channel of the input feature map. This represents the j-th channel of the feature map of the first convolutional layer. and S1 and S2 represent the weights and biases of the convolution kernel in the first convolutional layer, respectively. S0 and S1 represent the sets of input data and feature map indices of the first convolutional layer, respectively. * represents the convolution operation. This represents the activation function. Before performing the second convolutional layer operation, the feature map of the first convolutional layer is decomposed into a directly connected feature map S′1 and a bridging feature map S″1, where S1 = [S′1, S″1]. The directly connected feature map is then input into the second convolutional layer, specifically: in: Sj represents the j-th channel of the feature map of the second convolutional layer, and S2 represents the set of feature map numbers of the second convolutional layer. and These represent the weights and biases of the convolutional kernels in the second convolutional layer, respectively. The fusion pooling layer merges the crossover feature map of the first convolutional layer with the feature map of the second convolutional layer, specifically: in: The j-th channel of the feature map representing the fusion pooling layer, and ψ represents the pooling operation.

[0008] The training of the collaborative diagnostic model involves collecting vibration signals from rotating mechanical components using sensors, preprocessing them to extract time-domain features, frequency-domain features, and time-frequency-domain features based on wavelet packet decomposition as a fault feature training set. Based on a triple differentiation enhancement strategy—data-layer differentiation enhancement, feature-layer differentiation enhancement, and model-layer differentiation enhancement—a model cluster with a local feature reuse convolutional network as the base model is trained. The differences in diagnostic outputs of the fault feature training set across multiple local feature reuse convolutional networks are then calculated to construct a differentiation discrimination index. A differentiation discrimination threshold is then determined based on the distribution of the differentiation discrimination index of historical monitoring samples.

[0009] The sensor is preferably a piezoelectric accelerometer.

[0010] The aforementioned data layer differentiation enhancement refers to: randomly downsampling historical monitoring samples proportionally to form multiple training data subsets, and then using different data subsets for training different base models. Specifically, it is assumed that the number of training set samples constructed from historical monitoring samples is N. t Set the downsampling ratio to p, where 0.2 ≤ p ≤ 0.8. Use a sampling method without replacement to randomly sample from the training set to construct each training data subset, with a sampling size of pN.t The extraction process of each training data subset is independent of each other, and the number of training data subsets is equal to the number of base models.

[0011] The aforementioned feature layer differentiation enhancement refers to randomly selecting a subset of training features from the extracted features according to a preset number. Different feature subsets are used for training different base models; specifically, it is assumed that there are N extracted fault features. f Dimension, set the feature dimension of the feature subset to N′. f , N′ f <N f Using a no-replacement extraction method from N f Randomly select N′ from dimensional features f The training feature subsets are formed by the dimension, and the extraction process of each training feature subset is independent of each other. The number of training feature subsets is equal to the number of base models.

[0012] The aforementioned model layer differentiation enhancement refers to: pre-setting a range of key structural parameters for the base model, and randomly selecting different structural parameters from this range for different base models. Specifically, this involves first determining the key structural parameters of the base model through cross-validation optimization, and then, for parameter θ... i (i = 1, 2, ..., N) θ Assuming θ is determined through optimization. i =u i When constructing the base model, from [0.5u i 1.5u i ] A uniformly random value is selected as θ i The parameter setting processes of each base model are independent of each other, where: N θ This represents the number of key structural parameters.

[0013] The key structural parameters include, but are not limited to, the number of convolutional kernels in each convolutional layer.

[0014] The difference in diagnostic output of samples on different multi-local feature reuse convolutional networks refers to the difference between the diagnostic output vector of the base model and the average ensemble diagnostic vector. Assume that the diagnostic output vector of sample k on the o-th base model is... C represents the number of known state categories and the average ensemble diagnostic vector. Then the difference in diagnostic output of sample k on the o-th base model is: Where M is the number of base models.

[0015] The aforementioned differential discrimination index refers to: using Euclidean distance to measure the difference in diagnostic outputs of each base model, specifically calculating the mean distance between the diagnostic output vector of each base model and the average ensemble diagnostic vector. Wherein: DI kThis is the differential discrimination index for sample k.

[0016] The aforementioned setting of the differential discrimination threshold refers to: setting the differential discrimination threshold based on the distribution of differential discrimination indicators in historical monitoring samples. Where: T and N t These represent the set of historical monitoring sample numbers and the number of samples, respectively, with α being a preset threshold parameter.

[0017] The online phase includes: preprocessing and extracting features from the collected online monitoring samples of rotating mechanical parts, inputting them into a base model cluster to obtain the diagnostic output vectors of the online monitoring samples on each base model, and calculating the average integrated diagnostic vector and differential discrimination index of the online monitoring samples as the basis for collaborative diagnosis.

[0018] The base model cluster consists of all base models trained in the offline phase using a triple differential enhancement strategy.

[0019] The aforementioned collaborative diagnosis refers to the implementation of collaborative diagnosis by integrating the average ensemble diagnostic vector, differential discrimination index, and differential discrimination threshold to identify the state category of online monitoring samples of rotating mechanical components. When the differential discrimination index is greater than the differential discrimination threshold, the online monitoring sample is determined to belong to a newly emerging fault mode; otherwise, the state category of the sample is identified according to the average ensemble diagnostic vector. Technical effect

[0020] This invention employs a local feature reuse convolutional network as the base model to efficiently learn deep fault features of rotating mechanical components. It utilizes a triple differentiation enhancement strategy to strengthen the differences between base models at the data layer, feature layer, and model layer. Based on a differentiation discriminant index, it measures the differences in diagnostic outputs of samples across multiple local feature reuse convolutional networks. By comprehensively averaging the diagnostic vector, the differentiation discriminant index, and the differentiation threshold, it implements collaborative diagnosis, achieving fault diagnosis of rotating mechanical components that considers emerging fault modes. Compared with existing technologies, this invention improves the classification ability of the base model for monitoring samples of rotating mechanical components through the local feature reuse convolutional network while avoiding a significant increase in model parameters. The triple differentiation enhancement strategy strengthens the differences between different local feature reuse convolutional networks. The combination of these two approaches allows the base model to balance accuracy and differentiation. Collaborative diagnosis based on the differentiation discriminant index can simultaneously achieve reliable identification of both emerging and known fault modes of rotating mechanical components, overcoming the limited diagnostic accuracy caused by relying solely on data or feature differences. This is of great significance for ensuring the reliability and safety of rotating mechanical components. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2A schematic diagram of a convolutional network structure for reusing local features;

[0023] Figure 3 A comparison chart of fault diagnosis accuracy and model parameter quantity under different types of base models in Scenario 2;

[0024] Figure 4 This is a comparison chart of the fault diagnosis effects under different differentiated enhancement strategies in Scenario 1;

[0025] Figure 5 This is a comparison chart of the confusion matrix of diagnostic results under different methods in Scenario 2. Detailed Implementation

[0026] like Figure 1 As shown, this embodiment relates to a collaborative diagnostic method for rotating mechanical components that considers emerging failure modes. Taking rolling bearing fault diagnosis and gearbox fault diagnosis as examples, the method specifically includes:

[0027] Scenario 1: Rolling bearing fault diagnosis. The training set contains four types of samples: healthy bearings, outer ring faults, inner ring faults, and rolling element faults. The test set includes these four types plus a new fault mode (multi-point composite fault). The bearing speed is 13.33 Hz, and the vibration signal sampling frequency is 50 kHz. In both the training and test sets, there are 175 samples in each category, and each sample contains 4000 sampling points.

[0028] Scenario 2: Gearbox Fault Diagnosis. The training set includes five types of samples: healthy gears, combined faults of broken teeth, bearing damage, and shaft bending; combined faults of bearing damage and shaft imbalance; combined faults of broken teeth and bearing damage; and shaft bending. The test set includes these five types plus a newly emerging fault mode (gear spalling). The gearbox shaft speed is 30Hz, and the vibration signal sampling frequency is 66.67kHz. The training set contains 350 samples, and the test set contains 96 samples, with each sample containing 6188 sampling points.

[0029] The collaborative diagnostic method for rotating mechanical components, which considers emerging failure modes, is implemented for both Scenario 1 and Scenario 2, including the following steps:

[0030] Step 1: In the offline phase, the historical monitoring samples in the training set are preprocessed and time-domain, frequency-domain, and time-frequency-domain features are extracted. The time-domain features include 10 items: mean, root mean square value, peak value, amplitude, skewness, kurtosis, waveform factor, peak factor, impulse factor, and margin factor. The frequency-domain features include 3 items: center frequency of the envelope spectrum, root mean square frequency, and frequency standard deviation. The time-frequency-domain features are the sub-band energies calculated after the original signal is decomposed into three layers of wavelet packets, totaling 8 items.

[0031] Step 2: Construct multiple local feature reuse convolutional networks. Each local feature reuse convolutional network contains two local feature reuse modules, one tiling layer, and two fully connected layers. Each local feature reuse module includes two convolutional layers and one fusion pooling layer. Specifically, the first convolutional layer performs convolution operations on all input feature maps. in: This represents the i-th channel of the input feature map. This represents the j-th channel of the feature map of the first convolutional layer. and S1 and S2 represent the weights and biases of the convolution kernel in the first convolutional layer, respectively. S0 and S1 represent the sets of input data and feature map indices of the first convolutional layer, respectively. * represents the convolution operation. The activation function is ReLU. Before the second convolutional layer operation, the feature map of the first convolutional layer is decomposed into a directly connected feature map S′1 and a bridging feature map S″1, where S1 = [S′1, S″1]. During decomposition, the directly connected feature map and the bridging feature map are made to have the same number of channels, i.e., |S′1| = |S″1| = 0.5|S1|. After decomposition, the directly connected feature map is input into the second convolutional layer, specifically: in: Sj represents the j-th channel of the feature map of the second convolutional layer, and S2 represents the set of feature map numbers of the second convolutional layer. and These represent the weights and biases of the convolutional kernels in the second convolutional layer, respectively. The fusion pooling layer merges the crossover feature map of the first convolutional layer with the feature map of the second convolutional layer, specifically: in: The j-th channel of the feature map representing the fusion pooling layer, and ψ represents the pooling operation, specifically one-dimensional average pooling.

[0032] Step 3: Based on the triple differentiation enhancement strategy, train 5 and 10 model clusters with local feature reuse convolutional networks as the base model for scenario 1 and scenario 2, respectively.

[0033] The aforementioned triple differentiation enhancement strategy specifically includes:

[0034] 1) Data Layer Differentiation Enhancement: Random downsampling of historical monitoring samples creates multiple training data subsets. Different data subsets are used for training different base models. Specifically, assuming the training set constructed from historical monitoring samples has N samples... t The downsampling ratio p is set to 0.6, and 0.6N samples are randomly drawn from the training set using a sampling method without replacement. t Each sample is used to construct a subset of training data. The extraction process of each training data subset is independent of each other, and the number of training data subsets is equal to the number of base models.

[0035] 2) Feature layer differentiation enhancement: Randomly select some features from the extracted features to form a training feature subset. For different base models, different feature subsets are used for training. Specifically, a 16-dimensional feature subset is randomly selected from the extracted 21-dimensional features using a no-replacement extraction method. The extraction process of each training feature subset is independent of each other, and the number of training feature subsets is equal to the number of base models.

[0036] 3) Model layer differentiation enhancement: Pre-set the range of key structural parameters of the base model. For different base models, different structural parameters are randomly selected from the preset parameter range. Specifically, the optimal number of convolutional kernels for the convolutional layer is first determined to be 24 through cross-validation optimization. Then, when constructing the base model, the number of convolutional kernels for the four convolutional layers is set by uniformly and randomly selecting values ​​from [12, 36]. The random setting process of parameters for each base model is independent of each other.

[0037] Step 4: Input the historical monitoring samples from the training set into the already trained cluster of base models, calculate the average ensemble diagnostic vector and differential discrimination index of the samples on each base model, and determine the differential discrimination threshold, specifically including:

[0038] 1) Calculate the average ensemble diagnostic vector: Assume that the diagnostic output vector of sample k on the i-th base model is C represents the number of known state categories and the average ensemble diagnostic vector. Where M is the number of base models.

[0039] 2) Calculate the differential discrimination index: Use Euclidean distance to measure the difference in diagnostic output among the base models. Specifically, calculate the mean distance between the diagnostic output vector of each base model and the average ensemble diagnostic vector. Wherein: DI k This is the differential discrimination index for sample k.

[0040] 3) Establish differential discrimination thresholds: Set differential discrimination thresholds based on the distribution of differential discrimination indicators in historical monitoring samples. Where: T and N t These represent the set of historical monitoring sample numbers and the quantity, respectively. α is a preset threshold parameter, set to 3 according to the three sigma criterion.

[0041] Step 5: Enter the online stage. After preprocessing and feature extraction of the online monitoring samples in the test set, the samples are input into the base model cluster trained in the offline stage to obtain the diagnostic output vector of the online monitoring samples on each base model. The average ensemble diagnostic vector and differential discrimination index of the online monitoring samples are calculated as the basis for collaborative diagnosis.

[0042] Step 6: Perform collaborative diagnosis by integrating the average ensemble diagnostic vector, differential discrimination index, and differential discrimination threshold to identify the state category of online monitoring samples of rotating mechanical components. When the differential discrimination index is greater than the differential discrimination threshold, the online monitoring sample is determined to belong to a new fault mode; otherwise, the state category of the sample is identified according to the average ensemble diagnostic vector.

[0043] Table 1 compares the diagnostic performance of our proposed method with that of clustering-based feature similarity recognition methods and traditional convolutional neural network methods, quantitatively evaluated based on diagnostic accuracy and F2 score. The F2 score combines recall and precision, emphasizing the detection of faulty samples; a higher score indicates better diagnostic performance. In both scenarios, the fault diagnosis accuracy and F2 score of our proposed method are superior to the comparative methods.

[0044] Table 1

[0045] like Figure 3 As shown, taking scenario two as an example, a comparison of fault diagnosis accuracy and model parameter count under different types of base models is provided. The diagnostic accuracy of the local feature reuse convolutional network in the embodiment is better than that of traditional convolutional neural networks and densely connected convolutional networks. In addition, compared with densely connected convolutional networks that reuse all features, the local feature reuse convolutional network is more lightweight.

[0046] like Figure 4 As shown, taking Scenario 1 as an example, a comparison of fault diagnosis effects under different differentiation enhancement strategies is provided. Compared with the diagnostic technology without differentiation enhancement, the triple differentiation enhancement strategy of the embodiment improves the F2 score of collaborative fault diagnosis by 19.5%. In addition, compared with using only one differentiation enhancement method, the triple differentiation enhancement strategy further enhances the differentiation within the base model cluster, making collaborative fault diagnosis more reliable.

[0047] like Figure 5 As shown, taking scenario two as an example, a comparison of the confusion matrix of diagnostic results between this method and a clustering-based feature similarity recognition method and a traditional convolutional neural network method is provided. Here, category 1 represents the healthy state, categories 2-5 represent known fault modes, and category 6 represents newly emerging fault modes. The clustering-based feature similarity recognition method has limited diagnostic accuracy for samples of known categories; the traditional convolutional neural network method can classify samples of known categories relatively accurately, but performs poorly in detecting unknown faults. The collaborative diagnostic method for rotating mechanical parts considering newly emerging fault modes in the proposed embodiment provides reliable diagnoses for samples of all categories. In the technology of the proposed embodiment, collaborative diagnosis is implemented through multiple base models that balance accuracy and differentiation to identify newly emerging fault modes, fully utilizing the discriminative power of the diagnostic model.

[0048] Compared with existing technologies, this method has higher diagnostic accuracy for both new and known failure modes of rotating mechanical components. Specifically, it includes the following: the local feature reuse convolutional network has more accurate classification capabilities and a smaller number of model parameters; the triple differentiation enhancement strategy strengthens the differences between base models; and the combination of the two allows the base models to balance accuracy and differentiation, thereby improving the fault diagnosis accuracy of rotating mechanical components that consider new failure modes through collaborative diagnosis.

[0049] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A collaborative diagnostic method for rotating mechanical components considering emerging failure modes, characterized in that, After constructing a local feature reuse convolutional network in the offline stage, a differential discrimination index is built to measure the difference in diagnostic output of samples on different local feature reuse convolutional networks. This enables the training of a collaborative diagnostic model composed of different local feature reuse convolutional networks and a defined differential discrimination threshold. In the online stage, the average integrated diagnostic vector, differential discrimination index, and differential discrimination threshold obtained by the integrated training convolutional network are used for collaborative diagnosis, realizing reliable fault diagnosis of rotating mechanical parts that considers new fault modes. The local feature reuse convolutional network includes: several local feature reuse modules, a tiling layer, and a fully connected layer. Each local feature reuse module decomposes the feature map into directly connected feature maps and cross-connected feature maps, efficiently learning deep fault features and improving the classification ability of the base model through local reuse of the feature maps. The tiling layer flattens the feature map into one dimension to meet the data requirements of the fully connected layer. The fully connected layer performs classification by performing nonlinear mapping on the flattened feature map. The training of the collaborative diagnostic model involves collecting vibration signals from rotating mechanical components using sensors, preprocessing them to extract time-domain features, frequency-domain features, and time-frequency-domain features based on wavelet packet decomposition as a fault feature training set. Based on a triple differentiation enhancement strategy—data-layer differentiation enhancement, feature-layer differentiation enhancement, and model-layer differentiation enhancement—a model cluster with a local feature reuse convolutional network as the base model is trained. The differences in diagnostic outputs of the fault feature training set across multiple local feature reuse convolutional networks are then calculated to construct a differentiation discrimination index. A differentiation discrimination threshold is then determined based on the distribution of the differentiation discrimination index of historical monitoring samples. The local feature reuse module includes two convolutional layers and one fusion pooling layer, wherein the first convolutional layer performs convolution operations on all input feature maps, specifically: ,in: The first feature map representing the input feature map One channel, The first feature map representing the first convolutional layer One channel, and These represent the weights and biases of the convolutional kernel in the first convolutional layer, respectively. and These represent the sets of feature map numbers for the input data and the first convolutional layer, respectively. Represents convolution operation. This represents the activation function; before performing the second convolutional layer operation, the feature map of the first convolutional layer is decomposed into directly connected feature maps. With bridging feature map , Then, the directly connected feature map is input into the second convolutional layer, specifically: ,in: The first feature map representing the second convolutional layer One channel, This represents the set of feature map numbers for the second convolutional layer. and These represent the weights and biases of the convolutional kernels in the second convolutional layer, respectively. The fusion pooling layer merges the crossover feature map of the first convolutional layer with the feature map of the second convolutional layer, specifically: ,in: The first feature map representing the fusion pooling layer One channel, Represents pooling operations; The aforementioned data layer differentiation enhancement refers to: randomly downsampling historical monitoring samples proportionally to form multiple training data subsets, and then using different data subsets for training different base models. Specifically, assuming the number of training set samples constructed from historical monitoring samples is... Set the downsampling ratio to , A sampling method without replacement is used to randomly select samples from the training set to construct each training data subset, with the number of samples being [number missing]. The extraction process of each training data subset is independent of each other, and the number of training data subsets is equal to the number of base models; The aforementioned feature layer differentiation enhancement refers to: randomly selecting a training feature subset from the extracted features according to a preset number; and using different feature subsets for training different base models. Specifically, it is assumed that the extracted fault features total... Dimension, set the feature dimension of the feature subset to be . , Using a no-replacement extraction method from Randomly sampled from 3D features Each training feature subset is formed by multiple dimensions, and the extraction process of each training feature subset is independent of each other. The number of training feature subsets is equal to the number of base models. The aforementioned model layer differentiation enhancement refers to: pre-setting a range of key structural parameters for the base model, and randomly selecting different structural parameters from this range for different base models. Specifically, this involves: first, determining the key structural parameters of the base model through cross-validation optimization, and then, for the parameters... Assuming that the result is determined through optimization When constructing the base model, from Uniformly randomized values ​​are used as The parameter stochastic setting processes of each base model are independent of each other, wherein: The number of key structural parameters. .

2. The collaborative diagnostic method for rotating mechanical components considering emerging failure modes according to claim 1, characterized in that, The difference in diagnostic output of historical monitoring samples on different multi-local feature reuse convolutional networks refers to the difference between the base model diagnostic output vector and the average ensemble diagnostic vector. Assuming the samples... In the The diagnostic output vector on each base model is , Given the number of known state categories, the average ensemble diagnostic vector. Then the sample In the The difference in diagnostic output across the base models is ,in: The number of base models.

3. The collaborative diagnostic method for rotating mechanical components considering emerging failure modes according to claim 1, characterized in that, The aforementioned differential discrimination index refers to: using Euclidean distance to measure the difference in diagnostic outputs of each base model, specifically calculating the mean distance between the diagnostic output vector of each base model and the average ensemble diagnostic vector. ,in: For the sample Differentiation criteria.

4. The collaborative diagnostic method for rotating mechanical components considering emerging failure modes according to claim 3, characterized in that, The aforementioned setting of the differential discrimination threshold refers to: setting the differential discrimination threshold based on the distribution of differential discrimination indicators in historical monitoring samples. ,in: and These are the set of historical monitoring sample numbers and their quantities. This is a preset threshold parameter.

5. The collaborative diagnostic method for rotating mechanical components considering emerging failure modes according to claim 1, characterized in that, The online phase includes: preprocessing and extracting features from the collected online monitoring samples of rotating mechanical parts and inputting them into the base model cluster to obtain the diagnostic output vectors of the online monitoring samples on each base model, and calculating the average integrated diagnostic vector and differential discrimination index of the online monitoring samples as the basis for collaborative diagnosis; The aforementioned base model cluster consists of all base models trained in the offline phase using a triple differential enhancement strategy; The aforementioned collaborative diagnosis refers to: performing collaborative diagnosis by integrating the average integrated diagnostic vector, differential discrimination index, and differential discrimination threshold to identify the state category of online monitoring samples of rotating mechanical components. When the differential discrimination index is greater than the differential discrimination threshold, the online monitoring sample is determined to belong to a new fault mode; otherwise, the state category of the sample is identified according to the average integrated diagnostic vector.

Images